fix(scouts): fetch single Gmail message instead of bulk in fetch_content

Replace bulk GmailClient.fetch_messages() + linear search with a direct
service.users().messages().get(format="full") call. Adds _extract_plain_text_body
helper for recursive MIME part walking. Update test to patch _get_gmail_service.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.env.example

@@ -2,66 +2,94 @@
 ENV=dev
 
 # ── Database ──────────────────────────────────────────────────────────────────
-DATABASE_URL=postgresql+asyncpg://postgres:postgres@localhost:5432/adiuva
+DATABASE_URL=postgresql+asyncpg://postgres:postgres@localhost:5432/adiuvai
 
-# ── Redis ─────────────────────────────────────────────────────────────────────
+# ── Auth ──────────────────────────────────────────────────────────────────────
-REDIS_URL=redis://localhost:6379/0
+JWT_SECRET=replace-with-a-long-random-secret
-
+JWT_ALGORITHM=HS256
-# ── Auth (JWT RS256) ──────────────────────────────────────────────────────────
-# Generate keypair:
-#   openssl genpkey -algorithm RSA -out private.pem -pkeyopt rsa_keygen_bits:2048
-#   openssl rsa -in private.pem -pubout -out public.pem
-# Paste PEM content with literal \n for newlines.
-#
-# Private key — ONLY used by the Auth Service (JWT signing).
-JWT_PRIVATE_KEY=
-# Public key — used by all services / Traefik ForwardAuth (JWT verification).
-JWT_PUBLIC_KEY=
 JWT_ACCESS_TOKEN_EXPIRE_MINUTES=30
 JWT_REFRESH_TOKEN_EXPIRE_DAYS=30
 
 # ── LLM ───────────────────────────────────────────────────────────────────────
 # LiteLLM model identifiers — change to swap providers without code changes.
 # Examples: gpt-4o, anthropic/claude-sonnet-4-20250514, gemini/gemini-pro, ollama/llama3
+#
+# API keys — only the key(s) matching your chosen provider(s) are required.
+# The correct key is picked automatically from the model prefix (e.g.
+# "anthropic/..." → ANTHROPIC_API_KEY, "gemini/..." → GOOGLE_API_KEY).
 OPENAI_API_KEY=
 ANTHROPIC_API_KEY=
 GOOGLE_API_KEY=
-LLM_MODEL=gpt-4o
+CEREBRAS_API_KEY=
+GROQ_API_KEY=
+DEEPSEEK_API_KEY=
+
+# Default model used by any agent that does not have a specific override below.
+LLM_MODEL=gpt-5-mini
+LLM_EMBED_MODEL=text-embedding-3-small
+
+# GitHub Copilot — leave empty to use the LiteLLM default token directory.
+# In Docker, point this to a named-volume path so tokens survive restarts.
+# GITHUB_COPILOT_TOKEN_DIR=
+
+# ── Per-agent model overrides ─────────────────────────────────────────────────
+# Leave a value empty to fall back to LLM_MODEL.
+# Each agent resolves its API key from the model prefix automatically.
+#
+# Intent classifier — routes user messages to the right domain agent.
+# A small/fast model (e.g. gpt-4o-mini) is usually sufficient here.
+LLM_MODEL_CLASSIFIER=
+
+# Home-agent — handles chat from the home screen (all tools available).
+LLM_MODEL_HOME_AGENT=
+
+# Floating-agent — handles contextual chat triggered from a task/project/note.
+LLM_MODEL_FLOATING_AGENT=
+
+# Unified-processor — processes local directory files (local agent runner).
+LLM_MODEL_UNIFIED_PROCESSOR=
+
+# Cloud-processor — fetches and processes data from cloud connectors.
+LLM_MODEL_CLOUD_PROCESSOR=
+
+# Brief-agent — produces home and project text briefs.
+# A small model (e.g. gpt-4o-mini) is sufficient.
+# LLM_MODEL_BRIEF_AGENT=
+
+# Task-brief-agent — per-task deep research (Stage 1 executive assistant).
+# Needs tool-use + reasoning; a capable model recommended (e.g. gpt-4o, gemini-2.5-flash).
+# LLM_MODEL_TASK_BRIEF_AGENT=
+
+# Setup-agent — guided journey to build an AgentConfig via WebSocket chat.
+LLM_MODEL_SETUP_AGENT=
+
+# Memory-extractor — Mem0-style extract/decide pipeline (Phase 2).
+# Defaults to gpt-4o-mini when empty (fast + cheap, temperature=0).
+LLM_MODEL_MEMORY_EXTRACTOR=
+
+# Memory-miner — proactive pattern mining from episodic history (Phase 5, Power+ only).
+# Defaults to gpt-4o-mini when empty.
+LLM_MODEL_MEMORY_MINER=
+
+# Memory-auditor — weekly contradiction scan + relation label canonicalization (Phase 7).
+# Defaults to LLM_MODEL when empty (a reasoning-capable model is recommended).
+LLM_MODEL_MEMORY_AUDITOR=
+
+# Scheduler — set to false to disable memory cron jobs (automatically false in tests).
+SCHEDULER_ENABLED=true
 
 # ── Stripe (leave empty to stub billing) ──────────────────────────────────────
 STRIPE_SECRET_KEY=
 STRIPE_WEBHOOK_SECRET=
 
-# ── AWS / S3 ──────────────────────────────────────────────────────────────────
-S3_BUCKET=adiuva
-S3_REGION=us-east-1
-S3_ENDPOINT_URL=
-AWS_ACCESS_KEY_ID=
-AWS_SECRET_ACCESS_KEY=
-# For MinIO (homelab): S3_ENDPOINT_URL=http://minio:9000
 
-# ── Vector Store ──────────────────────────────────────────────────────────────
+# ── Langfuse (leave empty to disable observability) ───────────────────────────
-# Pinecone is used when PINECONE_API_KEY is set; otherwise falls back to Qdrant.
+LANGFUSE_SECRET_KEY=
-PINECONE_API_KEY=
+LANGFUSE_PUBLIC_KEY=
-PINECONE_INDEX=adiuva
+# LANGFUSE_BASE_URL=https://cloud.langfuse.com        # EU (default)
-QDRANT_URL=
+# LANGFUSE_BASE_URL=https://us.cloud.langfuse.com     # US
-QDRANT_API_KEY=
+# LANGFUSE_BASE_URL=http://localhost:3000             # Self-hosted
-# For local Qdrant (homelab): QDRANT_URL=http://qdrant:6333
 
 # ── CORS ──────────────────────────────────────────────────────────────────────
 # Comma-separated list parsed by Settings (override default if needed)
 # CORS_ORIGINS=["app://.","http://localhost:3000"]
-
-# ── Langfuse (observability) ─────────────────────────────────────────────────
-LANGFUSE_SECRET_KEY=sk-lf-...
-LANGFUSE_PUBLIC_KEY=pk-lf-...
-LANGFUSE_HOST=https://cloud.langfuse.com  # or self-hosted URL
-
-# ── Cloudflare (Traefik ACME DNS-01 challenge) ───────────────────────────────
-CF_DNS_API_TOKEN=
-ACME_EMAIL=
-
-# ── PostgreSQL (used by docker-compose) ──────────────────────────────────────
-POSTGRES_USER=postgres
-POSTGRES_PASSWORD=postgres
-POSTGRES_DB=adiuva

.gitea/workflows/deploy.yaml

@@ -48,23 +48,23 @@ jobs:
           key: ${{ secrets.SSH_KEY }}
           script: |
             set -e
-            DEPLOY_DIR="/opt/adiuva-api"
+            DEPLOY_DIR="/opt/adiuvai-api"
             REPO_URL="http://10.0.0.119:3000/${{ gitea.repository }}.git"
             TAG="${{ gitea.ref_name }}"
 
             # ── Pull latest code ──
-            cd /tmp && rm -rf adiuva-api-deploy
+            cd /tmp && rm -rf adiuvai-api-deploy
-            git clone --depth 1 --branch "${TAG}" "${REPO_URL}" adiuva-api-deploy
+            git clone --depth 1 --branch "${TAG}" "${REPO_URL}" adiuvai-api-deploy
 
             # ── Sync source (preserve .env) ──
-            cp -rf /tmp/adiuva-api-deploy/app/ \
+            cp -rf /tmp/adiuvai-api-deploy/app/ \
-                   /tmp/adiuva-api-deploy/alembic/ \
+                   /tmp/adiuvai-api-deploy/alembic/ \
-                   /tmp/adiuva-api-deploy/alembic.ini \
+                   /tmp/adiuvai-api-deploy/alembic.ini \
-                   /tmp/adiuva-api-deploy/Dockerfile \
+                   /tmp/adiuvai-api-deploy/Dockerfile \
-                   /tmp/adiuva-api-deploy/docker-compose.yml \
+                   /tmp/adiuvai-api-deploy/docker-compose.yml \
-                   /tmp/adiuva-api-deploy/requirements.txt \
+                   /tmp/adiuvai-api-deploy/requirements.txt \
                    "$DEPLOY_DIR/"
-            rm -rf /tmp/adiuva-api-deploy
+            rm -rf /tmp/adiuvai-api-deploy
 
             # ── Verify .env ──
             if [ ! -f "$DEPLOY_DIR/.env" ]; then

.github/workflows/ci.yml

@@ -58,7 +58,7 @@ jobs:
       - uses: actions/checkout@v4
 
       - name: Build image
-        run: docker build -t adiuva-api:ci .
+        run: docker build -t adiuvai-api:ci .
 
       - name: Verify gunicorn installed
-        run: docker run --rm adiuva-api:ci gunicorn --version
+        run: docker run --rm adiuvai-api:ci gunicorn --version

.gitignore

@@ -13,9 +13,6 @@ env/
 # Environment variables
 .env
 
-# Cryptographic keys
-*.pem
-
 # IDE
 .vscode/
 .idea/
@@ -24,17 +21,18 @@ env/
 .pytest_cache/
 htmlcov/
 .coverage
+tests/fixtures/private*/
 
 # Docker
 *.log
 
 # OS
 .DS_Store
+
+# Smoke scripts (dev-only, not for CI)
+scripts/smoke_*.py
 Thumbs.db
 
 # Claude Code
 .claude/
 logs/
-
-# Eval private test data
-services/batch-agent/eval/fixtures/private_data/

Dockerfile

@@ -3,34 +3,37 @@ FROM python:3.12-slim AS builder
 
 WORKDIR /build
 
-COPY services/chat/requirements.txt ./requirements.txt
+COPY requirements.txt .
 RUN pip install --upgrade pip && \
     pip install --no-cache-dir --prefix=/install -r requirements.txt
 
 # ── runtime ──────────────────────────────────────────────────────────────────
 FROM python:3.12-slim AS runtime
 
+# Non-root user
 RUN addgroup --system appgroup && adduser --system --ingroup appgroup appuser
 
 WORKDIR /app
 
+# Copy installed packages from builder
 COPY --from=builder /install /usr/local
 
-# Shared module
+# Copy application source
-COPY shared/ shared/
+COPY app/ app/
 
-# Service source
+# Copy Alembic migration files
-COPY services/chat/app/ app/
+COPY alembic/ alembic/
+COPY alembic.ini .
 
+# Ensure appuser owns the working directory
 RUN chown -R appuser:appgroup /app
 
 USER appuser
 
 EXPOSE 8000
 
-# Chat service is CPU-bound (LLM calls) — use multiple workers
 CMD ["gunicorn", "app.main:app", \
      "-k", "uvicorn.workers.UvicornWorker", \
      "--bind", "0.0.0.0:8000", \
-     "--workers", "2", \
+     "--workers", "4", \
      "--timeout", "120"]

README.md

@@ -1,793 +1,5 @@
-# Adiuva Cloud API
+## DEV
-
+Run in DEV with command:
-**AI-powered project management backend with E2E encrypted cloud storage, LLM orchestration, and a plugin marketplace.**
-
-Built with FastAPI · Python 3.12 · PostgreSQL · LangChain · Stripe · AWS S3
-
----
-
-## Table of Contents
-
-- [Overview](#overview)
-- [Architecture](#architecture)
-- [Key Features](#key-features)
-- [Tech Stack](#tech-stack)
-- [Getting Started](#getting-started)
-- [Docker Deployment](#docker-deployment)
-- [Environment Variables](#environment-variables)
-- [API Reference](#api-reference)
-- [Data Model](#data-model)
-- [AI Agent System](#ai-agent-system)
-- [Orchestration & Execution Plans](#orchestration--execution-plans)
-- [Middleware](#middleware)
-- [Storage Layer](#storage-layer)
-- [Billing & Tiers](#billing--tiers)
-- [Plugin Marketplace](#plugin-marketplace)
-- [Testing](#testing)
-- [Project Structure](#project-structure)
-- [License](#license)
-
----
-
-## Overview
-
-Adiuva Cloud API is the FastAPI backend that powers the **Adiuva Electron desktop app**. It provides LLM-powered chat orchestration, end-to-end encrypted cloud storage, a vector search engine, an encrypted backup system, a plugin marketplace with revenue sharing, and Stripe-based subscription billing across four tiers.
-
-### Design Principles
-
-1. **Never persist user data in plaintext** — the database stores only auth, billing, storage metadata, and marketplace data. All user content is E2E encrypted by the client before reaching the server.
-2. **Never expose prompts** — system prompts stay server-side; responses are sanitized to strip any leaked prompt fragments.
-3. **Never decrypt user blobs** — the backend performs only checksum verification; no decryption keys ever reach the server.
-4. **Stateless request handling** — all context comes from the client and JWT; no server-side session state.
-5. **Tier gates enforced server-side** — the server always reads the current tier from the database, never trusting client-reported values.
-
----
-
-## Architecture
-
 ```
-┌──────────────┐      ┌────────────────────────────────────────────────────────┐
+uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload --log-config logging.conf
-│  Electron    │      │  FastAPI  (Uvicorn / Gunicorn)                         │
-│  Desktop App │────▶│                                                        │
-│  (Client)    │◀────│  Middleware: RateLimit → Sanitizer → CORS → Router     │
-└──────────────┘      │                                                        │
-                      │  ┌──────────────────┐  ┌────────────────────────────┐  │
-                      │  │  Auth Routes     │  │  Chat Routes               │  │
-                      │  │  Billing Routes  │  │    ↓                       │  │
-                      │  │  Storage Routes  │  │  Orchestrator (GPT-4o-mini)│  │
-                      │  │  Backup Routes   │  │    ↓ classify intent       │  │
-                      │  │  Plugin Routes   │  │  Agent Registry            │  │
-                      │  │  Vector Routes   │  │    ↓                       │  │
-                      │  │  Plans Routes    │  │  TaskAgent  | ProjectAgent │  │
-                      │  └──────────────────┘  │  NoteAgent  | CheckptAgent │  │
-                      │                        │  (GPT-4o + LangChain)      │  │
-                      │                        └────────────────────────────┘  │
-                      └────────────────────────────────────────────────────────┘
-                               │              │              │
-                      ┌────────▼───┐  ┌───────▼───────┐  ┌──▼─────────────┐
-                      │ PostgreSQL │  │  AWS S3       │  │ Pinecone /     │
-                      │ (Auth,     │  │  (E2E blobs,  │  │ Qdrant         │
-                      │  Billing,  │  │   backups)    │  │ (Vectors)      │
-                      │  Metadata) │  └───────────────┘  └────────────────┘
-                      └────────────┘
-                               │
-                      ┌────────▼───┐
-                      │  Stripe    │
-                      │  (Billing, │
-                      │   Connect) │
-                      └────────────┘
 ```
-
----
-
-## Key Features
-
-1. **LLM-powered orchestration** — GPT-4o-mini classifies user intent and routes to the appropriate domain agent.
-2. **4 specialized AI agents** — Tasks (8 tools), Projects (6 tools), Timelines (4 tools), Notes (5 tools), all powered by GPT-4o via LangChain.
-3. **Execution plans & playbooks** — Server-side prompt template registry; clients receive only opaque template IDs, never raw prompts.
-4. **E2E encrypted cloud storage** — The backend never decrypts user data; SHA-256 checksum verification uses constant-time comparison to prevent timing attacks.
-5. **Cloud vector store** — Pinecone or Qdrant with user-isolated namespaces and encrypted blob payloads.
-6. **Encrypted backup system** — Tiered storage limits with `If-Modified-Since` support for efficient syncing.
-7. **Plugin marketplace** — Catalog, admin review/approval workflow, security checklist, and 70/30 revenue sharing via Stripe Connect.
-8. **Stripe billing** — Four-tier subscription model (Free / Pro / Power / Team) with checkout sessions and full webhook lifecycle handling.
-9. **JWT authentication** — Access + refresh tokens with bcrypt password hashing, SHA-256 token hashing, and automatic rotation.
-10. **Prompt IP protection** — Sanitizer middleware strips system prompts, reasoning markers, tool schemas, and agent routing metadata from all chat responses.
-11. **Tier-based rate limiting** — Sliding-window per-user limiter scaling from 20 to 200 requests/min by subscription tier.
-12. **Zero-trust data model** — User content is never stored in plaintext; the database holds only authentication, billing, and metadata records.
-13. **WebSocket streaming** — Real-time chat with 30-second heartbeat keep-alive and chunked text delivery.
-14. **Alembic migrations** — Versioned schema management with seed data for the plugin marketplace.
-15. **Comprehensive test suite** — In-memory SQLite + moto S3 mocks, per-tier test fixtures, and full API coverage without external dependencies.
-
----
-
-## Tech Stack
-
-| Package | Version | Purpose |
-|---|---|---|
-| `fastapi` | ≥ 0.115.0 | Web framework |
-| `uvicorn[standard]` | ≥ 0.34.0 | ASGI development server |
-| `gunicorn` | ≥ 22.0.0 | Production process manager |
-| `langchain` | ≥ 0.3.0 | LLM orchestration framework |
-| `langchain-openai` | ≥ 0.3.0 | OpenAI LLM provider integration |
-| `litellm` | ≥ 1.50.0 | Universal LLM gateway (100+ providers) |
-| `pydantic` | ≥ 2.10.0 | Data validation and serialization |
-| `pydantic-settings` | ≥ 2.7.0 | Environment-based configuration |
-| `python-jose[cryptography]` | ≥ 3.3.0 | JWT encoding and decoding |
-| `stripe` | ≥ 11.0.0 | Billing and payment integration |
-| `boto3` | ≥ 1.35.0 | AWS S3 client |
-| `slowapi` | ≥ 0.1.9 | Rate limiting utilities |
-| `sqlalchemy` | ≥ 2.0.0 | Async ORM and query builder |
-| `asyncpg` | ≥ 0.30.0 | PostgreSQL async driver |
-| `alembic` | ≥ 1.14.0 | Database migration management |
-| `bcrypt` | ≥ 4.2.0 | Password hashing |
-| `python-dotenv` | ≥ 1.0.0 | `.env` file loading |
-| `httpx` | ≥ 0.28.0 | Async HTTP client (used in tests) |
-| `websockets` | ≥ 14.0 | WebSocket protocol support |
-| `psycopg2-binary` | ≥ 2.9.0 | Synchronous PostgreSQL driver (Alembic) |
-| `pinecone` | ≥ 5.0.0 | Pinecone vector store client |
-| `qdrant-client` | ≥ 1.7.0 | Qdrant vector store client |
-| `pytest` | ≥ 8.0.0 | Test framework |
-| `pytest-asyncio` | ≥ 0.24.0 | Async test support |
-| `aiosqlite` | ≥ 0.20.0 | In-memory SQLite for tests |
-| `moto[s3]` | ≥ 5.0.0 | AWS S3 mock for tests |
-| `ruff` | ≥ 0.8.0 | Linter and formatter |
-
----
-
-## Getting Started
-
-### Prerequisites
-
-- Python 3.12+
-- PostgreSQL 16+
-- An OpenAI API key (for LLM features)
-- Stripe API keys (optional — billing stubs gracefully when unconfigured)
-- AWS credentials (optional — needed for S3 storage in production)
-
-### Installation
-
-```bash
-# Clone the repository
-git clone <repo-url> && cd adiuva-api
-
-# Create a virtual environment
-python -m venv .venv && source .venv/bin/activate
-
-# Install dependencies
-pip install -r requirements.txt
-
-# Configure environment
-cp .env.example .env
-# Edit .env with your DATABASE_URL, OPENAI_API_KEY, etc.
-```
-
-### Database Setup
-
-```bash
-# Start PostgreSQL (or use the Docker Compose database)
-docker compose up db -d
-
-# Run migrations
-alembic upgrade head
-```
-
-### Run the Development Server
-
-```bash
-uvicorn app.main:app --reload --host 0.0.0.0 --port 8000
-```
-
-Interactive API docs are available at [http://localhost:8000/docs](http://localhost:8000/docs) in development mode (`ENV=dev`). The `/docs` endpoint is disabled in production.
-
----
-
-## Docker Deployment
-
-### Quick Start
-
-```bash
-docker compose up --build
-```
-
-This starts two services:
-
-- **app** — FastAPI server on port `8000`
-- **db** — PostgreSQL 16 (Alpine) on port `5432` with a persistent volume and health checks
-
-The compose file also includes optional services for fully local deployments:
-
-- **minio** — S3-compatible object storage on ports `9000` (API) and `9001` (console)
-- **qdrant** — Vector search engine on ports `6333` (HTTP) and `6334` (gRPC)
-
-### Dockerfile Details
-
-The Dockerfile uses a multi-stage build:
-
-1. **Builder stage** — Installs Python dependencies into a virtual environment.
-2. **Runtime stage** — Copies only the venv, app source, and Alembic migrations. Runs as a non-root user (`appuser`).
-3. **Production server** — Gunicorn with 4 Uvicorn workers, 120-second timeout, listening on port 8000.
-
-```bash
-# Production command (run by the container)
-gunicorn app.main:app -k uvicorn.workers.UvicornWorker -w 4 --timeout 120 -b 0.0.0.0:8000
-```
-
----
-
-## Homelab / Self-Hosted Deployment
-
-You can run the entire stack locally on a homelab with **no cloud dependencies except the LLM provider**. The compose file includes MinIO (S3 replacement) and Qdrant (vector store) out of the box.
-
-### 1. Start all services
-
-```bash
-docker compose up -d
-```
-
-This starts PostgreSQL, MinIO, and Qdrant alongside the app.
-
-### 2. Create the MinIO bucket
-
-Open the MinIO console at [http://localhost:9001](http://localhost:9001) (login: `minioadmin` / `minioadmin`) and create a bucket named `adiuva`, or use the CLI:
-
-```bash
-docker compose exec minio mc alias set local http://localhost:9000 minioadmin minioadmin
-docker compose exec minio mc mb local/adiuva
-```
-
-### 3. Configure your `.env`
-
-```bash
-# Database (uses the compose PostgreSQL)
-DATABASE_URL=postgresql+asyncpg://postgres:postgres@db:5432/adiuva
-
-# S3 → MinIO
-S3_BUCKET=adiuva
-S3_REGION=us-east-1
-S3_ENDPOINT_URL=http://minio:9000
-AWS_ACCESS_KEY_ID=minioadmin
-AWS_SECRET_ACCESS_KEY=minioadmin
-
-# Vector store → local Qdrant (leave PINECONE_API_KEY empty)
-QDRANT_URL=http://qdrant:6333
-QDRANT_API_KEY=
-PINECONE_API_KEY=
-
-# Billing — leave empty to stub (no Stripe needed)
-STRIPE_SECRET_KEY=
-STRIPE_WEBHOOK_SECRET=
-
-# LLM — the only external service
-OPENAI_API_KEY=sk-...
-LLM_MODEL=gpt-4o
-LLM_ROUTER_MODEL=gpt-4o-mini
-
-# Auth
-JWT_SECRET=your-secret-here
-ENV=dev
-```
-
-### 4. Run migrations
-
-```bash
-docker compose exec app alembic upgrade head
-```
-
-### What runs where
-
-| Service | Runs on | Port | Notes |
-|---|---|---|---|
-| FastAPI app | Docker | 8000 | API server |
-| PostgreSQL | Docker | 5432 | Auth, billing, metadata |
-| MinIO | Docker | 9000 / 9001 | S3-compatible blob & backup storage |
-| Qdrant | Docker | 6333 / 6334 | Vector search (replaces Pinecone) |
-| Stripe | — | — | Stubbed when keys are empty |
-| OpenAI / LLM | Cloud | — | Only external dependency |
-
-> **Want fully offline AI too?** Set `LLM_MODEL=ollama/llama3` and `LLM_ROUTER_MODEL=ollama/llama3`, then add an Ollama container or point at a local Ollama instance. See the [LLM provider switching](#switching-llm-providers) section.
-
----
-
-## Environment Variables
-
-All variables are loaded from a `.env` file via Pydantic Settings. Source: `app/config/settings.py`
-
-| Variable | Type | Default | Description |
-|---|---|---|---|
-| `DATABASE_URL` | `str` | `postgresql+asyncpg://postgres:postgres@localhost:5432/adiuva` | Async SQLAlchemy connection string |
-| `JWT_SECRET` | `str` | `change-me-in-production` | HMAC secret for JWT signing |
-| `JWT_ALGORITHM` | `str` | `HS256` | JWT signing algorithm |
-| `JWT_ACCESS_TOKEN_EXPIRE_MINUTES` | `int` | `30` | Access token time-to-live |
-| `JWT_REFRESH_TOKEN_EXPIRE_DAYS` | `int` | `30` | Refresh token time-to-live |
-| `STRIPE_SECRET_KEY` | `str` | `""` | Stripe API key (empty = stub mode) |
-| `STRIPE_WEBHOOK_SECRET` | `str` | `""` | Stripe webhook signature secret |
-| `S3_BUCKET` | `str` | `""` | S3 bucket for encrypted blobs and backups |
-| `S3_REGION` | `str` | `us-east-1` | AWS region |
-| `S3_ENDPOINT_URL` | `str` | `""` | Custom S3 endpoint (e.g. `http://minio:9000` for MinIO). Leave empty for AWS. |
-| `AWS_ACCESS_KEY_ID` | `str` | `""` | AWS credentials |
-| `AWS_SECRET_ACCESS_KEY` | `str` | `""` | AWS credentials |
-| `PINECONE_API_KEY` | `str` | `""` | Pinecone API key (if set, Pinecone is used for vectors) |
-| `PINECONE_INDEX` | `str` | `adiuva` | Pinecone index name |
-| `QDRANT_URL` | `str` | `""` | Qdrant URL (used when Pinecone is not configured) |
-| `QDRANT_API_KEY` | `str` | `""` | Qdrant API key |
-| `OPENAI_API_KEY` | `str` | `""` | OpenAI key for LLM agent calls |
-| `LLM_MODEL` | `str` | `gpt-4o` | LiteLLM model identifier for agents (e.g. `anthropic/claude-3.5-sonnet`, `gemini/gemini-pro`, `ollama/llama3`) |
-| `LLM_ROUTER_MODEL` | `str` | `gpt-4o-mini` | Lighter model used for intent classification / routing |
-| `CORS_ORIGINS` | `list[str]` | `["app://.", "http://localhost:3000", "http://localhost:5173"]` | Allowed CORS origins |
-| `ENV` | `Literal` | `dev` | `dev` or `prod` — controls `/docs` visibility and SQL echo |
-
----
-
-## API Reference
-
-All routes are prefixed with `/api/v1`. **27 endpoints** total (25 REST + 1 WebSocket + 1 health check).
-
-### Health
-
-| Method | Path | Auth | Description |
-|---|---|---|---|
-| `GET` | `/api/v1/health` | No | Returns `{"status": "ok", "version": "0.1.0"}` |
-
-### Auth
-
-| Method | Path | Auth | Description |
-|---|---|---|---|
-| `POST` | `/api/v1/auth/register` | No | Create account with bcrypt-hashed password, returns `AuthTokens` |
-| `POST` | `/api/v1/auth/login` | No | Validate credentials, returns `AuthTokens` |
-| `POST` | `/api/v1/auth/refresh` | No | Rotate refresh token, returns new `AuthTokens` |
-| `GET` | `/api/v1/auth/me` | JWT | Returns `UserProfile` for the authenticated user |
-
-### Chat
-
-| Method | Path | Auth | Description |
-|---|---|---|---|
-| `POST` | `/api/v1/chat` | JWT | Route message through the orchestrator; returns `ChatResponse` or `ExecutionPlan` depending on execution mode |
-| `WS` | `/api/v1/chat/stream` | JWT (query param `?token=`) | Streaming chat — first frame is a `ChatRequest`, server yields text chunks, final frame is `{"done": true, "response": "...", "actions": [...]}`. 30-second heartbeat ping. |
-
-### Plans
-
-| Method | Path | Auth | Description |
-|---|---|---|---|
-| `GET` | `/api/v1/plans/playbook` | JWT | List all cached execution plan playbooks |
-| `GET` | `/api/v1/plans/playbook/{plan_id}` | JWT | Retrieve a specific playbook by ID |
-
-### Storage (Cloud Records)
-
-| Method | Path | Auth | Description |
-|---|---|---|---|
-| `POST` | `/api/v1/storage/records` | JWT | Upload an E2E encrypted record (verifies checksum, enforces storage quota) |
-| `GET` | `/api/v1/storage/records` | JWT | List record metadata with pagination (`?table`, `?page`, `?limit`); no blob bytes returned |
-| `GET` | `/api/v1/storage/records/{id}` | JWT | Download encrypted blob with `X-Checksum` response header |
-| `PUT` | `/api/v1/storage/records/{id}` | JWT | Replace an existing blob (verifies checksum, enforces quota) |
-| `DELETE` | `/api/v1/storage/records/{id}` | JWT | Delete a record and its S3 blob |
-
-### Vectors (Cloud Vector Store)
-
-| Method | Path | Auth | Description |
-|---|---|---|---|
-| `POST` | `/api/v1/storage/vectors/upsert` | JWT | Verify checksums and upsert encrypted vectors |
-| `POST` | `/api/v1/storage/vectors/search` | JWT | Search user-scoped vector namespace |
-| `DELETE` | `/api/v1/storage/vectors` | JWT | Delete vectors by ID list |
-
-### Backup
-
-| Method | Path | Auth | Description |
-|---|---|---|---|
-| `PUT` | `/api/v1/backup` | JWT | Upload encrypted backup blob with custom headers (`X-Backup-Version`, `X-Backup-Timestamp`, `X-Backup-Checksum`). Tier quota enforced. |
-| `GET` | `/api/v1/backup` | JWT | Download latest backup blob. Supports `If-Modified-Since`. |
-| `GET` | `/api/v1/backup/history` | JWT | List backup metadata (no blob content) |
-| `DELETE` | `/api/v1/backup/{backup_id}` | JWT | Delete a specific backup |
-
-### Plugins (Marketplace)
-
-| Method | Path | Auth | Description |
-|---|---|---|---|
-| `GET` | `/api/v1/plugins` | JWT (Power+) | Browse the marketplace (`?category`, `?q`, `?page`, `?sort=rating\|installs\|newest`) |
-| `GET` | `/api/v1/plugins/{id}` | JWT (Power+) | Plugin detail with install count and ratings |
-| `POST` | `/api/v1/plugins/{id}/install` | JWT (Power+) | Install plugin; triggers Stripe Connect revenue split for paid plugins |
-| `DELETE` | `/api/v1/plugins/{id}/install` | JWT | Uninstall plugin |
-
-### Billing
-
-| Method | Path | Auth | Description |
-|---|---|---|---|
-| `POST` | `/api/v1/billing/checkout` | JWT | Create a Stripe checkout session, returns `{"checkout_url": "..."}` |
-| `POST` | `/api/v1/billing/webhook` | Stripe signature | Handle Stripe events: `checkout.session.completed`, `customer.subscription.updated`, `customer.subscription.deleted`, `invoice.payment_failed` |
-| `GET` | `/api/v1/billing/subscription` | JWT | Get current subscription information |
-| `DELETE` | `/api/v1/billing/subscription` | JWT | Cancel subscription and revert to free tier |
-
----
-
-## Data Model
-
-9 tables managed by Alembic migrations. Source: `app/models.py`
-
-### Tables
-
-| Table | Primary Key | Key Columns | Purpose |
-|---|---|---|---|
-| `users` | `id` (UUID) | `email` (unique), `password_hash`, `tier`, `stripe_customer_id`, timestamps | User accounts |
-| `refresh_tokens` | `id` (UUID) | `user_id` (FK), `token_hash` (SHA-256, unique), `expires_at` | Hashed refresh tokens for rotation |
-| `subscriptions` | `id` (UUID) | `user_id` (FK, unique), `stripe_subscription_id`, `tier`, `status`, `current_period_end` | Stripe subscription records |
-| `storage_records` | `id` (UUID) | `user_id` (FK), `table_name`, `s3_key`, `checksum`, `size_bytes`, timestamps | S3 blob metadata (no plaintext content) |
-| `backup_metadata` | `id` (UUID) | `user_id` (FK), `s3_key`, `version`, `timestamp`, `checksum`, `size_bytes` | Backup manifests |
-| `plugins` | `id` (String) | `name`, `description`, `version`, `author_id` (FK), `category`, `price_cents`, `permissions` (JSON), `status`, `s3_package_key`, `install_count`, `avg_rating` | Marketplace plugin catalog |
-| `plugin_installations` | `id` (UUID) | `plugin_id` (FK), `user_id` (FK), unique constraint on (`plugin_id`, `user_id`) | Per-user install tracking |
-| `plugin_reviews` | `id` (UUID) | `plugin_id` (FK), `reviewer_id` (FK), `decision`, `notes`, `reviewed_at` | Admin review decisions |
-| `revenue_events` | `id` (UUID) | `plugin_id` (FK), `user_id` (FK), `amount_cents`, `developer_share_cents`, `stripe_transfer_id` | 70/30 revenue split ledger |
-
-### Enum Types
-
-| Enum | Values |
-|---|---|
-| `billing_tier` | `free`, `pro`, `power`, `team` |
-| `plugin_status` | `pending_review`, `approved`, `rejected` |
-| `review_decision` | `approved`, `rejected` |
-
-### Migrations
-
-| Version | Description |
-|---|---|
-| `001_initial_schema` | Creates all 9 tables with indexes and foreign key constraints |
-| `002_seed_plugins` | Seeds 3 approved plugins: GitHub Sync (free), Slack Notifier (€4.99), Time Tracker (€9.99) |
-
----
-
-## AI Agent System
-
-The agent system uses a registry pattern with LangChain tool-calling agents powered by GPT-4o. Source: `app/agents/`, `app/core/agent_registry.py`
-
-### Architecture
-
-- **`BaseAgent`** — Abstract base with `user_id`, `shared_memory`, and `vector_store_context`.
-- **`ChatAgent(BaseAgent)`** — Abstract `handle(query, context)` and `get_tools()` methods, plus a shared `_tool_loop(llm, messages, tools, max_iter=5)` for iterative tool calling.
-- **`AgentRegistry`** — Singleton registry with `@register` decorator, `get(name)`, `list_agents()`, and `call_agent(name, query, context)`.
-
-### Registered Agents
-
-| Agent | Registry Name | Tools | Description |
-|---|---|---|---|
-| **TaskAgent** | `task_agent` | 8 | Full task and comment CRUD. Status: `todo` / `in_progress` / `done`. Priority: `high` / `medium` / `low`. Tools: `list_tasks`, `create_task`, `update_task`, `delete_task`, `list_tasks_due_today`, `list_task_comments`, `add_task_comment`, `delete_task_comment` |
-| **ProjectAgent** | `project_agent` | 6 | Project lifecycle management. Status: `active` / `archived`. Prefers archiving over deletion. Tools: `list_projects`, `list_all_projects`, `get_project`, `create_project`, `update_project`, `delete_project` |
-| **TimelineAgent** | `timeline_agent` | 4 | Project milestones. Requires `project_id` for creation. Supports AI-suggestion and approval workflows. Tools: `list_timelines`, `create_timeline`, `update_timeline`, `delete_timeline` |
-| **NoteAgent** | `note_agent` | 5 | Markdown note management. Optionally linked to projects. Tools: `list_notes`, `get_note`, `create_note`, `update_note`, `delete_note` |
-
-All agents use the model configured by `LLM_MODEL` (default: GPT-4o) with `temperature=0` via LiteLLM. Tools return JSON action descriptors that the Electron client interprets and applies locally.
-
-### Switching LLM Providers
-
-The backend uses **LiteLLM** as a universal LLM gateway. All agents and the orchestrator instantiate models through a centralized factory in `app/core/llm.py`. To switch providers, change environment variables — no code changes required:
-
-```bash
-# OpenAI (default)
-LLM_MODEL=gpt-4o
-LLM_ROUTER_MODEL=gpt-4o-mini
-
-# Anthropic
-LLM_MODEL=anthropic/claude-3.5-sonnet
-LLM_ROUTER_MODEL=anthropic/claude-3-haiku
-
-# Google Gemini
-LLM_MODEL=gemini/gemini-pro
-LLM_ROUTER_MODEL=gemini/gemini-flash
-
-# Local Ollama
-LLM_MODEL=ollama/llama3
-LLM_ROUTER_MODEL=ollama/llama3
-
-# AWS Bedrock
-LLM_MODEL=bedrock/anthropic.claude-v2
-LLM_ROUTER_MODEL=bedrock/anthropic.claude-instant-v1
-```
-
-See the [LiteLLM provider docs](https://docs.litellm.ai/docs/providers) for the full list of 100+ supported providers and model naming conventions.
-
----
-
-## Orchestration & Execution Plans
-
-Source: `app/core/orchestrator.py`, `app/core/execution_plan.py`
-
-### Orchestrator
-
-1. **`classify_intent(message, context, registry)`** — Uses the router model (`LLM_ROUTER_MODEL`, default: GPT-4o-mini) to determine which agent should handle a message. Falls back to `task_agent` when classification is ambiguous.
-2. **`route_single(agent_name, message, context)`** — Routes to a single agent and returns a `ChatResponse`.
-3. **`route_pipeline(agent_names, message, context)`** — Executes agents sequentially; each receives `previous_results` from earlier agents. A final LLM synthesis step merges all results.
-4. **`orchestrate(request)`** — Main entry point. In `direct` mode, returns a `ChatResponse`. In `plan` mode, returns an `ExecutionPlan`.
-5. **`orchestrate_stream(request)`** — Streaming variant that yields 50-character text chunks with a final JSON frame.
-
-### Execution Plans
-
-- **`PromptTemplateRegistry`** — Maps template IDs to server-side prompt text. Clients only ever see opaque IDs, never raw prompts.
-- **`ExecutionPlanBuilder`** — Fluent builder API: `add_step()`, `add_llm_step(template_id, vars)`, `add_data_step(action, data_from_step)`. Validates step references on `build()`.
-- **`PlanCache`** — LRU cache (maxsize 1000) for storing plans as reusable playbooks.
-
-### Built-in Templates (6)
-
-`tpl_task_agent_default`, `tpl_timeline_agent_default`, `tpl_project_agent_default`, `tpl_note_agent_default`, `tpl_task_extract_from_project`, `tpl_note_weekly_summary`
-
-### Built-in Playbooks (2)
-
-| Playbook | Description |
-|---|---|
-| `create_tasks_from_project` | LLM extracts actionable tasks from project context, then creates task records |
-| `generate_weekly_note` | LLM generates a weekly summary, then creates a note record |
-
----
-
-## Middleware
-
-Middleware executes in this order on each request: **TierRateLimit → Sanitizer → CORS → Router**
-
-### JWT Authentication
-
-Source: `app/api/middleware/auth.py`
-
-- FastAPI dependency `get_current_user` validates the `Bearer` JWT and extracts `user_id` and `email`.
-- **Live tier lookup** — The current tier is fetched from the `subscriptions` table on every request (not cached in the JWT), so upgrades and downgrades take immediate effect.
-- Falls back to `free` when no subscription row exists.
-- Raises `401 Unauthorized` on invalid or expired tokens.
-- **Exempt paths:** `/api/v1/auth/register`, `/api/v1/auth/login`, `/api/v1/billing/webhook`
-
-### Tier-Based Rate Limiter
-
-Source: `app/api/middleware/rate_limit.py`
-
-- `TierRateLimitMiddleware` — Sliding-window in-process rate limiter (no Redis dependency).
-- Per-user 60-second window sized by subscription tier:
-
-| Tier | Requests / Minute |
-|---|---|
-| Free | 20 |
-| Pro | 60 |
-| Power | 120 |
-| Team | 200 |
-
-- Returns `429 Too Many Requests` with a `Retry-After` header when the limit is exceeded.
-- **Exempt paths:** register, login, webhook, health
-
-### Response Sanitizer
-
-Source: `app/api/middleware/sanitizer.py`
-
-- Runs only on `/api/v1/chat` endpoints.
-- Scans JSON response bodies and replaces leaked prompt IP fragments with `[REDACTED]`.
-- Detects: system prompt openers, agent routing metadata, LangChain tool schemas, internal reasoning markers (`<thinking>`, `[INST]`), and known prompt fingerprints.
-- Logs sanitization events as `WARNING`.
-- Binary responses (storage, backup) are never touched.
-
----
-
-## Storage Layer
-
-### Blob Store
-
-Source: `app/storage/blob_store.py`
-
-- S3-backed storage for E2E encrypted blobs.
-- Object keys follow the pattern: `{user_id}/{table}/{record_id}`
-- Server-side SSE-S3 encryption at rest (additional layer on top of client-side E2E encryption).
-- Methods: `upload()`, `download()`, `delete()` (idempotent), `list_keys()`
-- The backend **never inspects or decrypts blob content**.
-
-### Vector Store
-
-Source: `app/storage/vector_store.py`
-
-- Runtime-configurable: **Pinecone** (when `PINECONE_API_KEY` is set) or **Qdrant** (fallback).
-- User isolation: Pinecone uses `namespace=user_id`; Qdrant filters by `user_id` payload field.
-- 32-dimensional SHA-256-derived float vectors (deterministic, not semantically meaningful on encrypted data — a documented trade-off for privacy).
-- Encrypted blobs are stored as base64 in metadata/payload for verbatim retrieval.
-- Methods: `upsert()`, `search()`, `delete()`
-
-### Encryption Utilities
-
-Source: `app/storage/encryption.py`
-
-- `verify_checksum(blob, checksum)` — SHA-256 hash comparison using `hmac.compare_digest` (constant-time to prevent timing attacks).
-- `reject_if_tampered(blob, checksum)` — Raises HTTP 400 on checksum mismatch.
-- **No decryption key ever reaches the backend.**
-
----
-
-## Billing & Tiers
-
-Source: `app/billing/stripe_service.py`, `app/billing/tier_manager.py`
-
-### Feature Matrix
-
-| Feature | Free | Pro | Power | Team |
-|---|---|---|---|---|
-| AI Agents | 3 | Unlimited | Unlimited | Unlimited |
-| Batch Active | 2 | 10 | Unlimited | Unlimited |
-| Cloud Storage | 0 GB | 5 GB | 25 GB | Unlimited |
-| Backup Storage | 0 GB | 5 GB | 25 GB | Unlimited |
-| LLM Providers | 1 | Unlimited | Unlimited | Unlimited |
-| Batch Builder | — | — | ✓ | ✓ |
-| Plugin Marketplace | — | — | ✓ | ✓ |
-| SSO | — | — | — | ✓ |
-| Rate Limit | 20 req/min | 60 req/min | 120 req/min | 200 req/min |
-
-### Stripe Integration
-
-- **Checkout** — `create_checkout_session(user_id, tier)` creates a Stripe Checkout session. Returns a stub URL when Stripe is not configured.
-- **Webhooks** — Handles `checkout.session.completed`, `customer.subscription.updated`, `customer.subscription.deleted`, and `invoice.payment_failed`.
-- **Subscription management** — `get_subscription()` returns the current subscription record; `cancel_subscription()` cancels via the Stripe API and reverts the user to the free tier.
-- **Price IDs:** `price_pro_monthly`, `price_power_monthly`, `price_team_monthly`
-
-### Tier Manager
-
-- `get_tier(user_id)` — Returns the user's current billing tier.
-- `check_feature(tier, feature)` — Boolean feature gate check.
-- `require_feature(tier, feature)` — Raises HTTP 403 if the feature is not available.
-- `enforce_quota(user_id, tier)` / `enforce_backup_quota(user_id, tier)` — Raises HTTP 402 if storage limits are exceeded.
-
----
-
-## Plugin Marketplace
-
-Source: `app/marketplace/`
-
-### Plugin Registry
-
-- PostgreSQL-backed catalog of submitted and approved plugins.
-- `list_plugins(db, category, query, page, sort)` — Paginated listing (page size: 20) with optional filtering by category, text search, and sorting by `rating`, `installs`, or `newest`.
-- `get_plugin(db, plugin_id)` — Full manifest with install count and ratings.
-- `submit_plugin(db, manifest, s3_key)` — Submits a plugin with `pending_review` status.
-- `approve_plugin()` / `reject_plugin(reason)` — Admin workflow for plugin approval.
-- `record_install()` / `record_uninstall()` — Tracks per-user installations and updates install counts.
-
-### Review Queue
-
-- Automated security checklist before human review:
-  - Plugin ID must match `^[a-z0-9-]+$`
-  - Permissions must be from the allowed set only
-  - No binary blobs in the manifest
-- **Allowed permissions:** `read:tasks`, `write:tasks`, `read:projects`, `write:projects`, `read:notes`, `write:notes`, `read:timelines`, `write:timelines`, `read:calendar`, `write:calendar`
-- `get_pending(db)` — Lists plugins awaiting review.
-- `submit_review(db, plugin_id, reviewer_id, decision, notes)` — Records the review decision.
-
-### Revenue Sharing
-
-- **70% developer / 30% platform** split on all paid plugin sales.
-- `record_install(db, plugin_id, user_id, amount_cents)` — Records the revenue event and triggers a Stripe Connect transfer for the developer share.
-- `get_earnings(db, developer_id, period)` — Aggregated earnings report for plugin developers.
-- Gracefully stubs transfers when Stripe is not configured.
-
-### Seed Plugins
-
-| Plugin | Category | Price |
-|---|---|---|
-| GitHub Sync | Productivity | Free |
-| Slack Notifier | Communication | €4.99 |
-| Time Tracker | Productivity | €9.99 |
-
----
-
-## Testing
-
-### Running Tests
-
-```bash
-# Run all tests
-pytest
-
-# Run a specific test file
-pytest tests/test_auth.py
-
-# Run with verbose output
-pytest -v
-```
-
-### Test Infrastructure
-
-- **Database:** Async SQLite in-memory via `aiosqlite` + `StaticPool` — fast, no PostgreSQL needed.
-- **S3 mock:** `moto[s3]` with a fixture that patches `BlobStore` settings.
-- **Auth helpers:** `make_jwt(tier)` and `auth_header(tier)` generate per-tier test tokens.
-- **Seed data:** Auto-creates one `User` + `Subscription` per tier (free/pro/power/team) before each test.
-- **Plugin seeds:** Fixture adds 3 approved plugins for marketplace tests.
-- **FK enforcement:** SQLite `PRAGMA foreign_keys=ON`.
-- **No external dependencies** — all tests run fully offline.
-
-### Test Coverage
-
-| File | Coverage |
-|---|---|
-| `test_auth.py` | Register, login, token access, refresh, expiration |
-| `test_orchestrator.py` | Intent classification, single agent routing, pipeline, plan mode |
-| `test_agents.py` | Each agent with mocked LLM: registration, tools, handle method |
-| `test_storage.py` | Create, list, download, update, delete records; checksum rejection; quota enforcement |
-| `test_backup.py` | Upload, download, history, delete; tier-based storage limits |
-| `test_plugins.py` | List, install, uninstall, revenue events, tier gate enforcement |
-| `test_agent_registry.py` | Registry singleton, registration, lookup, listing |
-| `test_execution_plan.py` | Plan builder, template registry, plan cache |
-| `test_middleware.py` | Rate limiting by tier, sanitizer prompt leak detection |
-
----
-
-## Project Structure
-
-```
-adiuva-api/
-├── alembic.ini                  # Alembic configuration
-├── BACKEND_PLAN.md              # Architecture & design decisions
-├── docker-compose.yml           # Docker Compose (app + PostgreSQL)
-├── Dockerfile                   # Multi-stage production build
-├── requirements.txt             # Python dependencies
-│
-├── alembic/                     # Database migrations
-│   ├── env.py                   # Alembic environment config
-│   ├── script.py.mako           # Migration template
-│   └── versions/
-│       ├── 001_initial_schema.py    # Tables, indexes, FKs
-│       └── 002_seed_plugins.py      # Seed marketplace plugins
-│
-├── app/                         # Application source
-│   ├── main.py                  # FastAPI app factory, middleware, routes
-│   ├── db.py                    # Async SQLAlchemy engine & session
-│   ├── models.py                # SQLAlchemy ORM models (9 tables)
-│   ├── schemas.py               # Pydantic request/response schemas
-│   │
-│   ├── config/
-│   │   └── settings.py          # Pydantic Settings (env vars)
-│   │
-│   ├── agents/                  # LLM-powered domain agents
-│   │   ├── task_agent.py        # Task & comment CRUD (8 tools)
-│   │   ├── project_agent.py     # Project lifecycle (6 tools)
-│   │   ├── timeline_agent.py  # Milestones (4 tools)
-│   │   └── note_agent.py        # Markdown notes (5 tools)
-│   │
-│   ├── core/                    # Orchestration engine
-│   │   ├── agent_registry.py    # BaseAgent, ChatAgent, AgentRegistry
-│   │   ├── llm.py               # LiteLLM factory (get_llm)
-│   │   ├── orchestrator.py      # Intent classification & routing
-│   │   └── execution_plan.py    # Plan builder, templates, cache
-│   │
-│   ├── api/                     # HTTP layer
-│   │   ├── deps.py              # Shared FastAPI dependencies
-│   │   ├── middleware/
-│   │   │   ├── auth.py          # JWT validation, live tier lookup
-│   │   │   ├── rate_limit.py    # Sliding-window tier rate limiter
-│   │   │   └── sanitizer.py     # Prompt IP leak protection
-│   │   └── routes/
-│   │       ├── auth.py          # Register, login, refresh, me
-│   │       ├── chat.py          # Chat + WebSocket streaming
-│   │       ├── plans.py         # Execution plan playbooks
-│   │       ├── storage.py       # E2E encrypted record CRUD
-│   │       ├── vectors.py       # Vector upsert, search, delete
-│   │       ├── backup.py        # Encrypted backup management
-│   │       ├── plugins.py       # Marketplace browse & install
-│   │       └── billing.py       # Stripe checkout & webhooks
-│   │
-│   ├── storage/                 # Storage backends
-│   │   ├── blob_store.py        # S3 blob storage
-│   │   ├── vector_store.py      # Pinecone / Qdrant vector store
-│   │   └── encryption.py        # Checksum verification utilities
-│   │
-│   ├── billing/                 # Subscription management
-│   │   ├── stripe_service.py    # Stripe API integration
-│   │   └── tier_manager.py      # Feature matrix & quota enforcement
-│   │
-│   └── marketplace/             # Plugin ecosystem
-│       ├── plugin_registry.py   # Catalog CRUD & search
-│       ├── plugin_review.py     # Security checklist & review queue
-│       └── revenue_share.py     # 70/30 split & Stripe Connect
-│
-└── tests/                       # Test suite
-    ├── conftest.py              # Fixtures: DB, S3, auth, seeds
-    ├── test_auth.py
-    ├── test_orchestrator.py
-    ├── test_agents.py
-    ├── test_storage.py
-    ├── test_backup.py
-    ├── test_plugins.py
-    ├── test_agent_registry.py
-    ├── test_execution_plan.py
-    └── test_middleware.py
-```
-
----
-
-## License
-
-*To be determined.*

alembic/env.py

@@ -16,7 +16,7 @@ import re
 from logging.config import fileConfig
 
 from alembic import context
-from sqlalchemy import engine_from_config, pool
+from sqlalchemy import pool
 from sqlalchemy.ext.asyncio import create_async_engine
 
 # Alembic Config object (gives access to alembic.ini values).

alembic/versions/001_initial_schema.py

@@ -1,5 +1,4 @@
-"""Initial schema: users, refresh_tokens, subscriptions, storage_records,
+"""Initial schema: users, refresh_tokens, subscriptions.
-backup_metadata, plugins, plugin_installations, plugin_reviews, revenue_events.
 
 Revision ID: 001
 Revises:
@@ -28,18 +27,6 @@ def upgrade() -> None:
         EXCEPTION WHEN duplicate_object THEN NULL;
         END $$;
     """)
-    op.execute("""
-        DO $$ BEGIN
-            CREATE TYPE plugin_status AS ENUM ('pending_review', 'approved', 'rejected');
-        EXCEPTION WHEN duplicate_object THEN NULL;
-        END $$;
-    """)
-    op.execute("""
-        DO $$ BEGIN
-            CREATE TYPE review_decision AS ENUM ('approved', 'rejected');
-        EXCEPTION WHEN duplicate_object THEN NULL;
-        END $$;
-    """)
 
     # ── users ─────────────────────────────────────────────────────────────
     op.create_table(
@@ -88,122 +75,10 @@ def upgrade() -> None:
     op.create_index("ix_subscriptions_user_id", "subscriptions", ["user_id"])
     op.create_index("ix_subscriptions_stripe_id", "subscriptions", ["stripe_subscription_id"])
 
-    # ── storage_records ───────────────────────────────────────────────────
-    op.create_table(
-        "storage_records",
-        sa.Column("id", postgresql.UUID(as_uuid=False), nullable=False),
-        sa.Column("user_id", postgresql.UUID(as_uuid=False), nullable=False),
-        sa.Column("table_name", sa.String(100), nullable=False),
-        sa.Column("s3_key", sa.String(500), nullable=False),
-        sa.Column("checksum", sa.String(64), nullable=False),
-        sa.Column("size_bytes", sa.Integer, nullable=False),
-        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
-        sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
-        sa.PrimaryKeyConstraint("id"),
-        sa.ForeignKeyConstraint(["user_id"], ["users.id"], ondelete="CASCADE"),
-    )
-    op.create_index("ix_storage_records_user_id", "storage_records", ["user_id"])
-
-    # ── backup_metadata ───────────────────────────────────────────────────
-    op.create_table(
-        "backup_metadata",
-        sa.Column("id", postgresql.UUID(as_uuid=False), nullable=False),
-        sa.Column("user_id", postgresql.UUID(as_uuid=False), nullable=False),
-        sa.Column("s3_key", sa.String(500), nullable=False),
-        sa.Column("version", sa.Integer, nullable=False),
-        sa.Column("timestamp", sa.BigInteger, nullable=False),
-        sa.Column("checksum", sa.String(64), nullable=False),
-        sa.Column("size_bytes", sa.Integer, nullable=False),
-        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
-        sa.PrimaryKeyConstraint("id"),
-        sa.ForeignKeyConstraint(["user_id"], ["users.id"], ondelete="CASCADE"),
-    )
-    op.create_index("ix_backup_metadata_user_id", "backup_metadata", ["user_id"])
-
-    # ── plugins ───────────────────────────────────────────────────────────
-    op.create_table(
-        "plugins",
-        sa.Column("id", sa.String(255), nullable=False),
-        sa.Column("name", sa.String(255), nullable=False),
-        sa.Column("description", sa.Text, nullable=False, server_default=""),
-        sa.Column("version", sa.String(50), nullable=False, server_default="1.0.0"),
-        sa.Column("author_id", postgresql.UUID(as_uuid=False), nullable=True),
-        sa.Column("author_name", sa.String(255), nullable=False, server_default=""),
-        sa.Column("category", sa.String(100), nullable=False, server_default=""),
-        sa.Column("price_cents", sa.Integer, nullable=False, server_default="0"),
-        sa.Column("permissions", sa.Text, nullable=False, server_default="[]"),
-        sa.Column("status", postgresql.ENUM("pending_review", "approved", "rejected", name="plugin_status", create_type=False), nullable=False, server_default="pending_review"),
-        sa.Column("s3_package_key", sa.String(500), nullable=True),
-        sa.Column("install_count", sa.Integer, nullable=False, server_default="0"),
-        sa.Column("avg_rating", sa.Float, nullable=False, server_default="0.0"),
-        sa.Column("rejection_reason", sa.Text, nullable=True),
-        sa.Column("submitted_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
-        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
-        sa.PrimaryKeyConstraint("id"),
-        sa.ForeignKeyConstraint(["author_id"], ["users.id"], ondelete="SET NULL"),
-    )
-
-    # ── plugin_installations ──────────────────────────────────────────────
-    op.create_table(
-        "plugin_installations",
-        sa.Column("id", postgresql.UUID(as_uuid=False), nullable=False),
-        sa.Column("plugin_id", sa.String(255), nullable=False),
-        sa.Column("user_id", postgresql.UUID(as_uuid=False), nullable=False),
-        sa.Column("installed_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
-        sa.PrimaryKeyConstraint("id"),
-        sa.ForeignKeyConstraint(["plugin_id"], ["plugins.id"], ondelete="CASCADE"),
-        sa.ForeignKeyConstraint(["user_id"], ["users.id"], ondelete="CASCADE"),
-        sa.UniqueConstraint("plugin_id", "user_id", name="uq_plugin_user"),
-    )
-    op.create_index("ix_plugin_installations_plugin_id", "plugin_installations", ["plugin_id"])
-    op.create_index("ix_plugin_installations_user_id", "plugin_installations", ["user_id"])
-
-    # ── plugin_reviews ────────────────────────────────────────────────────
-    op.create_table(
-        "plugin_reviews",
-        sa.Column("id", postgresql.UUID(as_uuid=False), nullable=False),
-        sa.Column("plugin_id", sa.String(255), nullable=False),
-        sa.Column("reviewer_id", postgresql.UUID(as_uuid=False), nullable=True),
-        sa.Column("decision", postgresql.ENUM("approved", "rejected", name="review_decision", create_type=False), nullable=False),
-        sa.Column("notes", sa.Text, nullable=True),
-        sa.Column("reviewed_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
-        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
-        sa.PrimaryKeyConstraint("id"),
-        sa.ForeignKeyConstraint(["plugin_id"], ["plugins.id"], ondelete="CASCADE"),
-        sa.ForeignKeyConstraint(["reviewer_id"], ["users.id"], ondelete="SET NULL"),
-    )
-    op.create_index("ix_plugin_reviews_plugin_id", "plugin_reviews", ["plugin_id"])
-
-    # ── revenue_events ────────────────────────────────────────────────────
-    op.create_table(
-        "revenue_events",
-        sa.Column("id", postgresql.UUID(as_uuid=False), nullable=False),
-        sa.Column("plugin_id", sa.String(255), nullable=False),
-        sa.Column("user_id", postgresql.UUID(as_uuid=False), nullable=False),
-        sa.Column("amount_cents", sa.Integer, nullable=False, server_default="0"),
-        sa.Column("developer_share_cents", sa.Integer, nullable=False, server_default="0"),
-        sa.Column("stripe_transfer_id", sa.String(255), nullable=True),
-        sa.Column("paid_at", sa.DateTime(timezone=True), nullable=True),
-        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
-        sa.PrimaryKeyConstraint("id"),
-        sa.ForeignKeyConstraint(["plugin_id"], ["plugins.id"], ondelete="CASCADE"),
-        sa.ForeignKeyConstraint(["user_id"], ["users.id"], ondelete="CASCADE"),
-    )
-    op.create_index("ix_revenue_events_plugin_id", "revenue_events", ["plugin_id"])
-    op.create_index("ix_revenue_events_user_id", "revenue_events", ["user_id"])
-
 
 def downgrade() -> None:
-    op.drop_table("revenue_events")
-    op.drop_table("plugin_reviews")
-    op.drop_table("plugin_installations")
-    op.drop_table("plugins")
-    op.drop_table("backup_metadata")
-    op.drop_table("storage_records")
     op.drop_table("subscriptions")
     op.drop_table("refresh_tokens")
     op.drop_table("users")
 
-    op.execute("DROP TYPE IF EXISTS review_decision")
-    op.execute("DROP TYPE IF EXISTS plugin_status")
     op.execute("DROP TYPE IF EXISTS billing_tier")

alembic/versions/002_seed_plugins.py

@@ -1,92 +0,0 @@
-"""Seed approved plugins: GitHub Sync, Slack Notifier, Time Tracker.
-
-Revision ID: 002
-Revises: 001
-Create Date: 2026-03-03
-"""
-
-from __future__ import annotations
-
-import json
-from datetime import datetime, timezone
-from typing import Sequence, Union
-
-import sqlalchemy as sa
-from alembic import op
-
-revision: str = "002"
-down_revision: Union[str, None] = "001"
-branch_labels: Union[str, Sequence[str], None] = None
-depends_on: Union[str, Sequence[str], None] = None
-
-_SEED_PLUGINS = [
-    {
-        "id": "plugin-github-sync",
-        "name": "GitHub Sync",
-        "description": "Sync tasks with GitHub Issues and pull requests.",
-        "version": "1.0.0",
-        "author_name": "Adiuva",
-        "category": "productivity",
-        "price_cents": 0,
-        "permissions": json.dumps(["read:tasks", "write:tasks"]),
-        "status": "approved",
-        "s3_package_key": "plugins/plugin-github-sync/1.0.0/package.zip",
-        "install_count": 0,
-        "avg_rating": 0.0,
-    },
-    {
-        "id": "plugin-slack-notify",
-        "name": "Slack Notifier",
-        "description": "Post task and timeline updates to Slack channels.",
-        "version": "1.2.0",
-        "author_name": "Adiuva",
-        "category": "communication",
-        "price_cents": 499,
-        "permissions": json.dumps(["read:tasks", "read:timelines"]),
-        "status": "approved",
-        "s3_package_key": "plugins/plugin-slack-notify/1.2.0/package.zip",
-        "install_count": 0,
-        "avg_rating": 0.0,
-    },
-    {
-        "id": "plugin-time-tracker",
-        "name": "Time Tracker",
-        "description": "Track time spent on tasks with automatic reporting.",
-        "version": "0.9.1",
-        "author_name": "Third Party",
-        "category": "productivity",
-        "price_cents": 999,
-        "permissions": json.dumps(["read:tasks", "write:tasks"]),
-        "status": "approved",
-        "s3_package_key": "plugins/plugin-time-tracker/0.9.1/package.zip",
-        "install_count": 0,
-        "avg_rating": 0.0,
-    },
-]
-
-
-def upgrade() -> None:
-    plugins = sa.table(
-        "plugins",
-        sa.column("id", sa.String),
-        sa.column("name", sa.String),
-        sa.column("description", sa.Text),
-        sa.column("version", sa.String),
-        sa.column("author_name", sa.String),
-        sa.column("category", sa.String),
-        sa.column("price_cents", sa.Integer),
-        sa.column("permissions", sa.Text),
-        sa.column("status", sa.Enum("pending_review", "approved", "rejected", name="plugin_status")),
-        sa.column("s3_package_key", sa.String),
-        sa.column("install_count", sa.Integer),
-        sa.column("avg_rating", sa.Float),
-    )
-    op.bulk_insert(plugins, _SEED_PLUGINS)
-
-
-def downgrade() -> None:
-    op.execute(
-        "DELETE FROM plugins WHERE id IN ("
-        "'plugin-github-sync', 'plugin-slack-notify', 'plugin-time-tracker'"
-        ")"
-    )

alembic/versions/003_agent_tables.py

@@ -14,7 +14,7 @@ from alembic import op
 from sqlalchemy.dialects import postgresql
 
 revision: str = "003"
-down_revision: Union[str, None] = "002"
+down_revision: Union[str, None] = "001"
 branch_labels: Union[str, Sequence[str], None] = None
 depends_on: Union[str, Sequence[str], None] = None
 

alembic/versions/005_associative_pgvector.py

@@ -0,0 +1,54 @@
+"""Phase 1 — confirm pgvector activation on memory_associative.
+
+Migration 004 created the embedding column as vector(1536) and added the
+IVFFlat index.  This migration is the Phase-1 checkpoint:
+  1. Ensures the pgvector extension is enabled (idempotent).
+  2. Ensures the canonical Phase-1 IVFFlat index exists under the name
+     memory_associative_embedding_idx (creates it only if absent).
+
+Revision ID: 005
+Revises: 9a1f2d0b6c7e
+Create Date: 2026-04-15
+"""
+
+from __future__ import annotations
+
+from typing import Sequence, Union
+
+from alembic import op
+
+revision: str = "005"
+down_revision: Union[str, None] = "e04100e88ace"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Ensure pgvector extension is enabled (also done in 004, idempotent).
+    op.execute("CREATE EXTENSION IF NOT EXISTS vector;")
+
+    # Ensure the canonical Phase-1 IVFFlat index exists.
+    # 004 may have created ix_memory_associative_embedding; this adds the
+    # Phase-1 name memory_associative_embedding_idx if it is missing.
+    op.execute(
+        """
+        DO $$
+        BEGIN
+            IF NOT EXISTS (
+                SELECT 1
+                FROM   pg_indexes
+                WHERE  tablename  = 'memory_associative'
+                  AND  indexname  = 'memory_associative_embedding_idx'
+            ) THEN
+                CREATE INDEX memory_associative_embedding_idx
+                ON memory_associative
+                USING ivfflat (embedding vector_cosine_ops)
+                WITH  (lists = 100);
+            END IF;
+        END $$;
+        """
+    )
+
+
+def downgrade() -> None:
+    op.execute("DROP INDEX IF EXISTS memory_associative_embedding_idx;")

alembic/versions/006_memory_relations.py

@@ -0,0 +1,74 @@
+"""Add memory_relations table (Phase 3 — relational tier).
+
+Revision ID: 006
+Revises: 1f5975a4f3f4
+Create Date: 2026-04-16
+"""
+
+from __future__ import annotations
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+revision: str = "006"
+down_revision: Union[str, None] = "1f5975a4f3f4"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "memory_relations",
+        sa.Column("id", postgresql.UUID(as_uuid=False), primary_key=True),
+        sa.Column(
+            "user_id",
+            postgresql.UUID(as_uuid=False),
+            sa.ForeignKey("users.id", ondelete="CASCADE"),
+            nullable=False,
+        ),
+        sa.Column("subject_label", sa.String(128), nullable=False),
+        sa.Column("subject_type", sa.String(32), nullable=False),
+        sa.Column("predicate", sa.String(64), nullable=False),
+        sa.Column("object_label", sa.String(128), nullable=False),
+        sa.Column("object_type", sa.String(32), nullable=False),
+        sa.Column("confidence", sa.Float, nullable=False, server_default="0.7"),
+        sa.Column(
+            "source_episode_id",
+            postgresql.UUID(as_uuid=False),
+            sa.ForeignKey("memory_episodic.id", ondelete="SET NULL"),
+            nullable=True,
+        ),
+        sa.Column("notes_encrypted", sa.LargeBinary, nullable=True),
+        sa.Column(
+            "created_at",
+            sa.DateTime(timezone=True),
+            nullable=False,
+            server_default=sa.func.now(),
+        ),
+        sa.Column(
+            "updated_at",
+            sa.DateTime(timezone=True),
+            nullable=False,
+            server_default=sa.func.now(),
+        ),
+        sa.Column("last_confirmed_at", sa.DateTime(timezone=True), nullable=True),
+    )
+    op.create_index(
+        "memory_relations_user_subject_idx",
+        "memory_relations",
+        ["user_id", "subject_label"],
+    )
+    op.create_index(
+        "memory_relations_user_predicate_idx",
+        "memory_relations",
+        ["user_id", "predicate"],
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("memory_relations_user_predicate_idx", "memory_relations")
+    op.drop_index("memory_relations_user_subject_idx", "memory_relations")
+    op.drop_table("memory_relations")

alembic/versions/007_rename_agents_to_scouts.py

@@ -0,0 +1,41 @@
+"""Rename agents to scouts.
+
+Revision ID: 007
+Revises: d6e3f4a5b6c7
+Create Date: 2026-05-15
+
+Renames the entire agents subsystem identifiers to scouts.
+Pre-1.0 — no data preservation concerns beyond ALTER TABLE rename.
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+
+revision: str = "007"
+down_revision: Union[str, None] = "d6e3f4a5b6c7"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Tables
+    op.rename_table("local_agent_configs", "local_scout_configs")
+    op.rename_table("cloud_agent_configs", "cloud_scout_configs")
+    op.rename_table("agent_run_logs", "scout_run_logs")
+
+    # Columns
+    op.alter_column("local_scout_configs", "agent_config", new_column_name="scout_config")
+    op.alter_column("scout_run_logs", "agent_id", new_column_name="scout_id")
+    op.alter_column("scout_run_logs", "agent_type", new_column_name="scout_type")
+
+
+def downgrade() -> None:
+    op.alter_column("scout_run_logs", "scout_type", new_column_name="agent_type")
+    op.alter_column("scout_run_logs", "scout_id", new_column_name="agent_id")
+    op.alter_column("local_scout_configs", "scout_config", new_column_name="agent_config")
+
+    op.rename_table("scout_run_logs", "agent_run_logs")
+    op.rename_table("cloud_scout_configs", "cloud_agent_configs")
+    op.rename_table("local_scout_configs", "local_agent_configs")

alembic/versions/008_scout_triage_queue.py

@@ -0,0 +1,59 @@
+"""Scout triage queue + cloud_scout_configs alterations.
+
+Revision ID: 008
+Revises: 007
+Create Date: 2026-05-16
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+
+revision: str = "008"
+down_revision: Union[str, None] = "007"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "scout_triage_queue",
+        sa.Column("id", sa.Uuid(as_uuid=False), primary_key=True),
+        sa.Column("user_id", sa.Uuid(as_uuid=False), sa.ForeignKey("users.id", ondelete="CASCADE"), nullable=False, index=True),
+        sa.Column("scout_id", sa.Uuid(as_uuid=False), sa.ForeignKey("cloud_scout_configs.id", ondelete="CASCADE"), nullable=False),
+        sa.Column("source_type", sa.String(50), nullable=False),
+        sa.Column("source_msg_ref", sa.String(255), nullable=False),
+        sa.Column("triage_verdict", sa.String(20), nullable=False),
+        sa.Column("triage_reason", sa.Text, nullable=True),
+        sa.Column("status", sa.String(20), nullable=False, server_default="queued"),
+        sa.Column("triaged_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.func.now()),
+        sa.Column("delivered_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column("acked_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column("expires_at", sa.DateTime(timezone=True), nullable=False),
+        sa.UniqueConstraint("scout_id", "source_msg_ref", name="uq_scout_triage_queue_scout_msg"),
+    )
+    op.create_index("ix_scout_triage_queue_user_status", "scout_triage_queue", ["user_id", "status"])
+    op.create_index(
+        "ix_scout_triage_queue_expires_active",
+        "scout_triage_queue",
+        ["expires_at"],
+        postgresql_where=sa.text("status != 'acked'"),
+    )
+
+    op.add_column("cloud_scout_configs", sa.Column("auto_trash_spam", sa.Boolean(), nullable=False, server_default=sa.text("false")))
+    op.add_column("cloud_scout_configs", sa.Column("gmail_history_id", sa.String(64), nullable=True))
+    op.add_column("cloud_scout_configs", sa.Column("gmail_watch_expires_at", sa.DateTime(timezone=True), nullable=True))
+    op.add_column("cloud_scout_configs", sa.Column("device_inactivity_pause_days", sa.Integer(), nullable=False, server_default="14"))
+
+
+def downgrade() -> None:
+    op.drop_column("cloud_scout_configs", "device_inactivity_pause_days")
+    op.drop_column("cloud_scout_configs", "gmail_watch_expires_at")
+    op.drop_column("cloud_scout_configs", "gmail_history_id")
+    op.drop_column("cloud_scout_configs", "auto_trash_spam")
+
+    op.drop_index("ix_scout_triage_queue_expires_active", table_name="scout_triage_queue")
+    op.drop_index("ix_scout_triage_queue_user_status", table_name="scout_triage_queue")
+    op.drop_table("scout_triage_queue")

alembic/versions/1f5975a4f3f4_add_extraction_queue.py

@@ -0,0 +1,38 @@
+"""add extraction_queue
+
+Revision ID: 1f5975a4f3f4
+Revises: 005
+Create Date: 2026-04-16 17:26:25.790870
+
+"""
+from __future__ import annotations
+
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision: str = '1f5975a4f3f4'
+down_revision: Union[str, None] = '005'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        'extraction_queue',
+        sa.Column('id', sa.Uuid(as_uuid=False), nullable=False),
+        sa.Column('user_id', sa.Uuid(as_uuid=False), nullable=False),
+        sa.Column('episode_id', sa.Uuid(as_uuid=False), nullable=True),
+        sa.Column('created_at', sa.DateTime(timezone=True), server_default=sa.text('now()'), nullable=False),
+        sa.ForeignKeyConstraint(['user_id'], ['users.id'], ondelete='CASCADE'),
+        sa.PrimaryKeyConstraint('id'),
+    )
+    op.create_index(op.f('ix_extraction_queue_user_id'), 'extraction_queue', ['user_id'], unique=False)
+
+
+def downgrade() -> None:
+    op.drop_index(op.f('ix_extraction_queue_user_id'), table_name='extraction_queue')
+    op.drop_table('extraction_queue')

alembic/versions/a3b9c0d1e2f3_add_agent_config_to_local_agents.py

@@ -0,0 +1,107 @@
+"""Restore agent config tables and add agent_config column.
+
+9a1f2d0b6c7e dropped local_agent_configs and cloud_agent_configs, but both
+ORM models are still active. This migration recreates them with agent_config
+added to local_agent_configs.
+
+Revision ID: a3b9c0d1e2f3
+Revises: 9a1f2d0b6c7e
+Create Date: 2026-04-07 00:00:00.000000
+
+"""
+from __future__ import annotations
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+
+# revision identifiers, used by Alembic.
+revision: str = "a3b9c0d1e2f3"
+down_revision: Union[str, None] = "9a1f2d0b6c7e"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Recreate enum types (idempotent — they may already exist from migration 003)
+    op.execute("""
+        DO $$ BEGIN
+            CREATE TYPE agent_type AS ENUM ('local', 'cloud');
+        EXCEPTION WHEN duplicate_object THEN NULL;
+        END $$;
+    """)
+    op.execute("""
+        DO $$ BEGIN
+            CREATE TYPE agent_run_status AS ENUM ('running', 'success', 'error', 'partial');
+        EXCEPTION WHEN duplicate_object THEN NULL;
+        END $$;
+    """)
+    op.execute("""
+        DO $$ BEGIN
+            CREATE TYPE cloud_provider AS ENUM ('gmail', 'teams', 'outlook');
+        EXCEPTION WHEN duplicate_object THEN NULL;
+        END $$;
+    """)
+
+    bind = op.get_bind()
+    inspector = sa.inspect(bind)
+    existing = set(inspector.get_table_names())
+
+    # ── local_agent_configs (with agent_config column) ────────────────────
+    if "local_agent_configs" not in existing:
+        op.create_table(
+            "local_agent_configs",
+            sa.Column("id", postgresql.UUID(as_uuid=False), nullable=False),
+            sa.Column("user_id", postgresql.UUID(as_uuid=False), nullable=False),
+            sa.Column("device_id", sa.String(255), nullable=False),
+            sa.Column("name", sa.String(255), nullable=False),
+            sa.Column("directory_paths", sa.JSON, nullable=False, server_default="[]"),
+            sa.Column("data_types", sa.JSON, nullable=False, server_default="[]"),
+            sa.Column("prompt_template", sa.Text, nullable=False, server_default=""),
+            sa.Column("agent_config", sa.JSON, nullable=True),
+            sa.Column("file_extensions", sa.JSON, nullable=False, server_default="[]"),
+            sa.Column("schedule_cron", sa.String(100), nullable=False, server_default="0 */6 * * *"),
+            sa.Column("enabled", sa.Boolean, nullable=False, server_default=sa.true()),
+            sa.Column("last_run_at", sa.DateTime(timezone=True), nullable=True),
+            sa.Column("created_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
+            sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
+            sa.PrimaryKeyConstraint("id"),
+            sa.ForeignKeyConstraint(["user_id"], ["users.id"], ondelete="CASCADE"),
+        )
+        op.create_index("ix_local_agent_configs_user_id", "local_agent_configs", ["user_id"])
+
+    # ── cloud_agent_configs ───────────────────────────────────────────────
+    if "cloud_agent_configs" not in existing:
+        op.create_table(
+            "cloud_agent_configs",
+            sa.Column("id", postgresql.UUID(as_uuid=False), nullable=False),
+            sa.Column("user_id", postgresql.UUID(as_uuid=False), nullable=False),
+            sa.Column(
+                "provider",
+                postgresql.ENUM("gmail", "teams", "outlook", name="cloud_provider", create_type=False),
+                nullable=False,
+            ),
+            sa.Column("name", sa.String(255), nullable=False),
+            sa.Column("data_types", sa.JSON, nullable=False, server_default="[]"),
+            sa.Column("prompt_template", sa.Text, nullable=False, server_default=""),
+            sa.Column("oauth_token_encrypted", sa.Text, nullable=True),
+            sa.Column("filter_config", sa.JSON, nullable=True),
+            sa.Column("schedule_cron", sa.String(100), nullable=False, server_default="0 */6 * * *"),
+            sa.Column("enabled", sa.Boolean, nullable=False, server_default=sa.true()),
+            sa.Column("last_run_at", sa.DateTime(timezone=True), nullable=True),
+            sa.Column("created_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
+            sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False, server_default=sa.text("now()")),
+            sa.PrimaryKeyConstraint("id"),
+            sa.ForeignKeyConstraint(["user_id"], ["users.id"], ondelete="CASCADE"),
+        )
+        op.create_index("ix_cloud_agent_configs_user_id", "cloud_agent_configs", ["user_id"])
+
+
+def downgrade() -> None:
+    op.drop_index("ix_cloud_agent_configs_user_id", table_name="cloud_agent_configs")
+    op.drop_table("cloud_agent_configs")
+    op.drop_index("ix_local_agent_configs_user_id", table_name="local_agent_configs")
+    op.drop_table("local_agent_configs")

alembic/versions/b4c0d1e2f3a4_add_oauth_and_avatar.py

@@ -0,0 +1,56 @@
+"""Add oauth_accounts table, nullable password_hash, avatar_url to users.
+
+Revision ID: b4c0d1e2f3a4
+Revises: a3b9c0d1e2f3
+Create Date: 2026-04-10 00:00:00.000000
+
+"""
+from __future__ import annotations
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+
+# revision identifiers, used by Alembic.
+revision: str = "b4c0d1e2f3a4"
+down_revision: Union[str, None] = "a3b9c0d1e2f3"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ── users: make password_hash nullable (social users have no password) ──
+    op.alter_column("users", "password_hash", existing_type=sa.String(255), nullable=True)
+
+    # ── users: add avatar_url ─────────────────────────────────────────────
+    op.add_column("users", sa.Column("avatar_url", sa.String(2048), nullable=True))
+
+    # ── oauth_accounts ────────────────────────────────────────────────────
+    op.create_table(
+        "oauth_accounts",
+        sa.Column("id", postgresql.UUID(as_uuid=False), nullable=False),
+        sa.Column("user_id", postgresql.UUID(as_uuid=False), nullable=False),
+        sa.Column("provider", sa.String(50), nullable=False),
+        sa.Column("provider_user_id", sa.String(255), nullable=False),
+        sa.Column("provider_email", sa.String(255), nullable=True),
+        sa.Column(
+            "created_at",
+            sa.DateTime(timezone=True),
+            nullable=False,
+            server_default=sa.text("now()"),
+        ),
+        sa.PrimaryKeyConstraint("id"),
+        sa.ForeignKeyConstraint(["user_id"], ["users.id"], ondelete="CASCADE"),
+        sa.UniqueConstraint("provider", "provider_user_id", name="uq_oauth_provider_user"),
+    )
+    op.create_index("ix_oauth_accounts_user_id", "oauth_accounts", ["user_id"])
+
+
+def downgrade() -> None:
+    op.drop_index("ix_oauth_accounts_user_id", table_name="oauth_accounts")
+    op.drop_table("oauth_accounts")
+    op.drop_column("users", "avatar_url")
+    op.alter_column("users", "password_hash", existing_type=sa.String(255), nullable=False)

alembic/versions/c5d1e2f3a4b5_add_onboarding_completed_at.py

@@ -0,0 +1,31 @@
+"""Add onboarding_completed_at column to users table.
+
+Revision ID: c5d1e2f3a4b5
+Revises: b4c0d1e2f3a4
+Create Date: 2026-04-11 00:00:00.000000
+
+"""
+from __future__ import annotations
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+
+# revision identifiers, used by Alembic.
+revision: str = "c5d1e2f3a4b5"
+down_revision: Union[str, None] = "b4c0d1e2f3a4"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "users",
+        sa.Column("onboarding_completed_at", sa.DateTime(timezone=True), nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("users", "onboarding_completed_at")

alembic/versions/d6e3f4a5b6c7_folder_index_tables.py

@@ -0,0 +1,46 @@
+"""Add token tracking columns for folder integration.
+
+Revision ID: d6e3f4a5b6c7
+Revises: 006
+Create Date: 2026-05-11 00:00:00.000000
+
+"""
+from __future__ import annotations
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects.postgresql import UUID
+
+# revision identifiers, used by Alembic.
+revision: str = "d6e3f4a5b6c7"
+down_revision: Union[str, None] = "006"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "agent_run_logs",
+        sa.Column("tokens_used", sa.Integer(), nullable=False, server_default="0"),
+    )
+    op.create_table(
+        "monthly_token_usage",
+        sa.Column("user_id", UUID(as_uuid=False), sa.ForeignKey("users.id", ondelete="CASCADE"), nullable=False),
+        sa.Column("year_month", sa.String(7), nullable=False),
+        sa.Column("feature", sa.String(64), nullable=False),
+        sa.Column("tokens_used", sa.Integer(), nullable=False, server_default="0"),
+        sa.PrimaryKeyConstraint("user_id", "year_month", "feature"),
+    )
+    op.create_index(
+        "ix_monthly_token_usage_user_month",
+        "monthly_token_usage",
+        ["user_id", "year_month"],
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("ix_monthly_token_usage_user_month", table_name="monthly_token_usage")
+    op.drop_table("monthly_token_usage")
+    op.drop_column("agent_run_logs", "tokens_used")

alembic/versions/e04100e88ace_avatar_url_varchar_to_text.py

@@ -0,0 +1,34 @@
+"""avatar_url_varchar_to_text
+
+Revision ID: e04100e88ace
+Revises: c5d1e2f3a4b5
+Create Date: 2026-04-13 09:13:06.733674
+
+"""
+from __future__ import annotations
+
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision: str = 'e04100e88ace'
+down_revision: Union[str, None] = 'c5d1e2f3a4b5'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.alter_column('users', 'avatar_url',
+               existing_type=sa.VARCHAR(length=2048),
+               type_=sa.Text(),
+               existing_nullable=True)
+
+
+def downgrade() -> None:
+    op.alter_column('users', 'avatar_url',
+               existing_type=sa.Text(),
+               type_=sa.VARCHAR(length=2048),
+               existing_nullable=True)

app/agents/__init__.py

@@ -0,0 +1,5 @@
+"""Expose tool modules used by deep orchestrator-worker graphs."""
+
+from app.agents import filesystem_agent, timeline_agent, note_agent, project_agent, task_agent
+
+__all__ = ["filesystem_agent", "timeline_agent", "note_agent", "project_agent", "task_agent"]

app/agents/client_agent.py

@@ -0,0 +1,52 @@
+"""Client agent — read-only tools for the clients table."""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from langchain_core.tools import tool
+
+from app.core.ws_context import execute_on_client
+
+
+@tool
+async def list_clients(search: str = "", limit: int = 20) -> str:
+    """List clients, optionally filtered by a name/email substring search.
+
+    search: optional substring to match against client name or email.
+    limit: max rows to return (default 20).
+    """
+    filters: dict[str, Any] = {"limit": limit}
+    if search:
+        filters["search"] = search
+
+    result = await execute_on_client(action="select", table="clients", filters=filters)
+    rows = result.get("rows", [])
+    if not rows:
+        return "No clients found."
+    lines = [
+        f"- {r.get('name', '?')} (id: {r.get('id')}, email: {r.get('email', '')}, "
+        f"company: {r.get('company', '')})"
+        for r in rows
+    ]
+    return f"Found {len(rows)} client(s):\n" + "\n".join(lines)
+
+
+@tool
+async def get_client(id: str) -> str:
+    """Get full details for one client by UUID.
+
+    id: the client's UUID.
+    """
+    if not id:
+        return "Client id is required."
+
+    result = await execute_on_client(action="get", table="clients", data={"id": id})
+    row = result.get("row") or result.get("rows", [None])[0] if result else None
+    if not row:
+        return f"Client '{id}' not found."
+    return f"Client details:\n{json.dumps(row, ensure_ascii=False, indent=2)}"
+
+
+CLIENT_TOOLS: list[Any] = [list_clients, get_client]

app/agents/filesystem_agent.py

@@ -0,0 +1,194 @@
+"""Filesystem agent — tools for reading local directories and files on Electron.
+
+These tools delegate to the Electron client via ``execute_on_client()`` using
+the same WS tool-call round-trip pattern as CRUD tools.  The Electron app
+handles actual disk I/O and responds with ``tool_result`` frames.
+"""
+
+from __future__ import annotations
+
+import os
+import re
+from pathlib import Path
+from typing import Any
+
+from langchain_core.tools import tool
+
+from app.core.ws_context import execute_on_client
+
+# Max characters returned by read_file_content in journey (exploration) tools.
+# The journey only needs to understand file structure, not full content.
+_JOURNEY_READ_MAX_CHARS: int = 4000
+
+
+def _resolve_path(path: str, base: str) -> str:
+    """Resolve *path* against *base* when *path* is relative.
+
+    The LLM often passes ``"."`` meaning "the configured directory".
+    Without this, Electron resolves ``"."`` relative to its own CWD instead
+    of the user's chosen directory.
+    """
+    if os.path.isabs(path):
+        return path
+    return str(Path(base) / path)
+
+
+@tool
+async def list_directory(path: str) -> str:
+    """List files and folders in a local directory on the user's device.
+
+    Returns a formatted listing of entries with name, type (file/directory),
+    and full path.
+    """
+    result = await execute_on_client(
+        action="list_directory",
+        data={"path": path},
+    )
+    entries: list[dict[str, Any]] = result.get("entries", [])
+    if not entries:
+        return f"Directory '{path}' is empty or does not exist."
+    lines: list[str] = []
+    for entry in entries:
+        entry_type = entry.get("type", "unknown")
+        entry_name = entry.get("name", "")
+        entry_path = entry.get("path", "")
+        lines.append(f"- [{entry_type}] {entry_name}  ({entry_path})")
+    return f"Directory listing for '{path}' ({len(entries)} entries):\n" + "\n".join(lines)
+
+
+@tool
+async def read_file_content(path: str) -> str:
+    """Read the text content of a local file on the user's device.
+
+    Returns the file content as a string.  Large files may be truncated
+    by the Electron client.
+    """
+    result = await execute_on_client(
+        action="read_file_content",
+        data={"path": path},
+    )
+    content: str = result.get("content", "")
+    if not content:
+        return f"File '{path}' is empty or could not be read."
+    return content
+
+
+@tool
+async def get_file_metadata(path: str) -> str:
+    """Get metadata for a local file: size, creation date, modification date, extension.
+
+    Returns a formatted summary of the file's metadata.
+    """
+    result = await execute_on_client(
+        action="get_file_metadata",
+        data={"path": path},
+    )
+    size = result.get("size", "unknown")
+    created = result.get("createdAt", "unknown")
+    modified = result.get("modifiedAt", "unknown")
+    extension = result.get("extension", "unknown")
+    name = result.get("name", path)
+    return (
+        f"File: {name}\n"
+        f"  Extension: {extension}\n"
+        f"  Size: {size} bytes\n"
+        f"  Created: {created}\n"
+        f"  Modified: {modified}"
+    )
+
+
+FILESYSTEM_TOOLS: list[Any] = [
+    list_directory,
+    read_file_content,
+    get_file_metadata,
+]
+
+
+def make_directory_tools(base_directory: str) -> list[Any]:
+    """Return filesystem tools that resolve relative paths against *base_directory*.
+
+    Use this instead of ``FILESYSTEM_TOOLS`` whenever you know the user's target
+    directory upfront (e.g., journey setup sessions).  Relative paths like ``"."``
+    from the LLM are resolved to the correct absolute path before being sent to
+    the Electron client, preventing it from falling back to its own CWD.
+    """
+
+    def _compact_for_journey(raw: str) -> str:
+        """Strip HTML noise and truncate for journey exploration.
+
+        The journey LLM only needs to understand file structure (headers,
+        first paragraphs).  Full CSS/style blocks are pure noise that eat
+        up context window budget.
+        """
+        text = re.sub(r"<style[^>]*>.*?</style>", "", raw, flags=re.DOTALL | re.IGNORECASE)
+        text = re.sub(r"<script[^>]*>.*?</script>", "", text, flags=re.DOTALL | re.IGNORECASE)
+        text = re.sub(r"<!--.*?-->", "", text, flags=re.DOTALL)
+        if len(text) > _JOURNEY_READ_MAX_CHARS:
+            text = text[:_JOURNEY_READ_MAX_CHARS] + "\n[…truncated for exploration]"
+        return text
+
+    @tool
+    async def list_directory(path: str) -> str:  # noqa: F811
+        """List files and folders in a local directory on the user's device.
+
+        Returns a formatted listing of entries with name, type (file/directory),
+        and full path.
+        """
+        resolved = _resolve_path(path, base_directory)
+        result = await execute_on_client(
+            action="list_directory",
+            data={"path": resolved},
+        )
+        entries: list[dict[str, Any]] = result.get("entries", [])
+        if not entries:
+            return f"Directory '{resolved}' is empty or does not exist."
+        lines: list[str] = []
+        for entry in entries:
+            entry_type = entry.get("type", "unknown")
+            entry_name = entry.get("name", "")
+            entry_path = entry.get("path", "")
+            lines.append(f"- [{entry_type}] {entry_name}  ({entry_path})")
+        return f"Directory listing for '{resolved}' ({len(entries)} entries):\n" + "\n".join(lines)
+
+    @tool
+    async def read_file_content(path: str) -> str:  # noqa: F811
+        """Read the text content of a local file on the user's device.
+
+        Returns the file content as a string.  Large files may be truncated
+        by the Electron client.
+        """
+        resolved = _resolve_path(path, base_directory)
+        result = await execute_on_client(
+            action="read_file_content",
+            data={"path": resolved},
+        )
+        content: str = result.get("content", "")
+        if not content:
+            return f"File '{resolved}' is empty or could not be read."
+        return _compact_for_journey(content)
+
+    @tool
+    async def get_file_metadata(path: str) -> str:  # noqa: F811
+        """Get metadata for a local file: size, creation date, modification date, extension.
+
+        Returns a formatted summary of the file's metadata.
+        """
+        resolved = _resolve_path(path, base_directory)
+        result = await execute_on_client(
+            action="get_file_metadata",
+            data={"path": resolved},
+        )
+        size = result.get("size", "unknown")
+        created = result.get("createdAt", "unknown")
+        modified = result.get("modifiedAt", "unknown")
+        extension = result.get("extension", "unknown")
+        name = result.get("name", resolved)
+        return (
+            f"File: {name}\n"
+            f"  Extension: {extension}\n"
+            f"  Size: {size} bytes\n"
+            f"  Created: {created}\n"
+            f"  Modified: {modified}"
+        )
+
+    return [list_directory, read_file_content, get_file_metadata]

app/agents/folder_agent.py

@@ -0,0 +1,168 @@
+"""Scoped file-read and search tools for the project folder feature."""
+from __future__ import annotations
+
+from langchain_core.tools import tool
+
+from app.core.folder_indexer import _extract_docx_text, _extract_pdf_text
+from app.core.ws_context import execute_on_client
+
+# Cap returned slice size to keep tool output under control.
+_MAX_RETURN_CHARS = 50_000
+_MAX_SEARCH_MATCHES = 20
+
+
+def _is_unsafe_path(rel: str) -> bool:
+    if not rel:
+        return True
+    norm = rel.replace("\\", "/")
+    if norm.startswith("/"):
+        return True
+    # Windows drive letter
+    if len(rel) >= 2 and rel[1] == ":":
+        return True
+    parts = norm.split("/")
+    return ".." in parts
+
+
+async def _fetch_file(project_id: str, relative_path: str, offset: int, length: int) -> dict:
+    """Return the raw Electron tool_result dict for a file read."""
+    return await execute_on_client(
+        action="read_project_folder_file",
+        data={
+            "projectId": project_id,
+            "relativePath": relative_path,
+            "offset": offset,
+            "length": length,
+        },
+    )
+
+
+def _decode(result: dict) -> tuple[str, str, int]:
+    """Decode a tool_result into (text, kind, total_size). For pdf/docx,
+    extracts text from base64. For images, returns a placeholder string.
+    For text, content is already a sliced utf-8 string.
+    """
+    kind = result.get("kind", "text")
+    content = result.get("content", "") or ""
+    total = int(result.get("totalSize", 0) or 0)
+    if kind == "image":
+        return ("[Image file — cannot be navigated as text. See manifest summary.]", kind, total)
+    if kind == "pdf":
+        return (_extract_pdf_text(content), kind, total)
+    if kind == "docx":
+        return (_extract_docx_text(content), kind, total)
+    return (content, kind, total)
+
+
+@tool
+async def read_project_folder_file(
+    project_id: str,
+    relative_path: str,
+    offset: int = 0,
+    length: int = _MAX_RETURN_CHARS,
+) -> str:
+    """Read a slice of a file inside the project's linked folder.
+
+    Args:
+        project_id: project ID.
+        relative_path: path relative to the linked folder root.
+        offset: char offset to start reading from (0 = beginning).
+        length: max chars to return. Default 50000. Use smaller values to save tokens.
+
+    Returns text content slice with a header showing position. Header tells you
+    when more content is available; call again with the suggested next offset.
+
+    For PDF / DOCX files the backend extracts text first, then applies offset/length
+    on the extracted text. For images returns a placeholder; navigate with the
+    manifest summary instead.
+    """
+    if _is_unsafe_path(relative_path):
+        return "Access denied"
+
+    result = await _fetch_file(project_id, relative_path, offset, length)
+    text, kind, total_size = _decode(result)
+
+    if not text and kind in ("missing", "error"):
+        return f"File not found or unreadable: {relative_path}"
+
+    if kind in ("pdf", "docx"):
+        # Backend extracted full text — apply offset/length on chars.
+        sliced = text[offset:offset + length]
+        slice_end = min(offset + length, len(text))
+        header = (
+            f"[file={relative_path} kind={kind} offset={offset} end={slice_end} "
+            f"totalChars={len(text)}]"
+        )
+        if slice_end < len(text):
+            header += f"\n[More content available — call again with offset={slice_end}.]"
+        return header + "\n" + sliced
+
+    if kind == "text":
+        slice_end = offset + len(text)
+        header = (
+            f"[file={relative_path} kind=text offset={offset} end={slice_end} "
+            f"totalBytes={total_size}]"
+        )
+        if slice_end < total_size:
+            header += f"\n[More content available — call again with offset={slice_end}.]"
+        return header + "\n" + text
+
+    # image or unknown
+    return text
+
+
+@tool
+async def search_project_folder_file(
+    project_id: str,
+    relative_path: str,
+    query: str,
+    context_lines: int = 3,
+) -> str:
+    """Search a project folder file for a query string (case-insensitive substring).
+
+    Args:
+        project_id: project ID.
+        relative_path: path relative to the linked folder root.
+        query: text to search for.
+        context_lines: number of lines of context around each match (default 3).
+
+    Returns matching line ranges with surrounding context and 1-based line numbers.
+    Capped at 20 matches; if more exist the header shows the total.
+
+    Works on text, code, markdown, PDF (extracted), and DOCX (extracted).
+    Images and binary files are not searchable.
+    """
+    if _is_unsafe_path(relative_path):
+        return "Access denied"
+    if not query:
+        return "Empty query."
+
+    # For text we still need full file; pass length=very large.
+    result = await _fetch_file(project_id, relative_path, offset=0, length=10_000_000)
+    text, kind, _ = _decode(result)
+
+    if not text and kind in ("missing", "error"):
+        return f"File not found or unreadable: {relative_path}"
+    if kind == "image":
+        return "Cannot search inside images."
+
+    lines = text.splitlines()
+    q = query.lower()
+    matches = [i for i, line in enumerate(lines) if q in line.lower()]
+    if not matches:
+        return f"No matches for '{query}' in {relative_path}."
+
+    shown = matches[:_MAX_SEARCH_MATCHES]
+    snippets: list[str] = []
+    for i in shown:
+        start = max(0, i - context_lines)
+        end = min(len(lines), i + context_lines + 1)
+        block = "\n".join(f"{n + 1:5d}: {lines[n]}" for n in range(start, end))
+        snippets.append(block)
+
+    header = f"[file={relative_path} matches={len(matches)} showing={len(shown)} query='{query}']"
+    body = "\n---\n".join(snippets)
+    return header + "\n" + body
+
+
+FOLDER_TOOLS = [read_project_folder_file, search_project_folder_file]

app/agents/note_agent.py

@@ -0,0 +1,206 @@
+"""Note agent — Markdown note management (list, get, create, update, propose edit)."""
+
+from __future__ import annotations
+
+import asyncio
+import re
+from typing import Any
+
+from langchain_core.tools import tool
+
+from app.core.note_summarizer import generate_note_summary
+from app.core.ws_context import execute_on_client
+
+_UUID_RE = re.compile(
+    r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-5][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$"
+)
+
+
+def _is_uuid(value: str) -> bool:
+    return bool(_UUID_RE.match(value))
+
+
+def _fmt_summary(row: dict) -> str:
+    summary = (row.get("aiSummary") or row.get("ai_summary") or "").strip()
+    if summary:
+        return f" — {summary}"
+    snippet = (row.get("content") or "")[:120].replace("\n", " ").strip()
+    return f" — {snippet}" if snippet else ""
+
+
+@tool
+async def list_notes(project_id: str = "") -> str:
+    """List notes with AI summaries, optionally scoped to a project by project_id.
+
+    Returns id, title, and ai_summary for each note so you can decide which
+    note to read in full with get_note before creating or updating.
+    """
+    normalized_project_id = project_id if (project_id and _is_uuid(project_id)) else ""
+    result = await execute_on_client(
+        action="select",
+        table="notes",
+        filters={"projectId": normalized_project_id or None},
+    )
+    rows = result.get("rows", [])
+    if not rows:
+        return "No notes found."
+    lines = [f"  - [{r['id']}] {r['title']}{_fmt_summary(r)}" for r in rows]
+    return f"Found {len(rows)} note(s):\n" + "\n".join(lines)
+
+
+@tool
+async def get_note(note_id: str) -> str:
+    """Fetch a single note by its UUID to read its full Markdown content."""
+    result = await execute_on_client(action="get", table="notes", data={"id": note_id})
+    row = result.get("row")
+    if not row:
+        return f"Note {note_id} not found."
+    return f"Note '{row['title']}' (id: {row['id']}):\n\n{row['content']}"
+
+
+@tool
+async def create_note(
+    title: str,
+    content: str,
+    project_id: str = "",
+) -> str:
+    """Create a new note.
+    title: note heading (required)
+    content: Markdown body text (required)
+    project_id: optional UUID linking this note to a project
+    """
+    result = await execute_on_client(
+        action="insert",
+        table="notes",
+        data={
+            "title": title,
+            "content": content,
+            "projectId": project_id or None,
+        },
+    )
+    row = result["row"]
+    note_id: str = row["id"]
+    # Generate summary asynchronously — fire-and-forget.
+    asyncio.create_task(_refresh_summary(note_id, title, content))
+    return f"Note created: '{row['title']}' (id: {note_id})."
+
+
+@tool
+async def update_note(
+    note_id: str,
+    title: str = "",
+    content: str = "",
+) -> str:
+    """Update an existing note directly (no approval required).
+    Use propose_note_edit instead when human review is needed.
+    note_id: UUID of the note (required)
+    If you need to preserve existing content, call get_note first.
+    """
+    updates: dict[str, Any] = {}
+    if title:
+        updates["title"] = title
+    if content:
+        updates["content"] = content
+    result = await execute_on_client(
+        action="update",
+        table="notes",
+        data={"id": note_id, "updates": updates},
+    )
+    row = result["row"]
+    if content:
+        new_title = title or row.get("title", "")
+        asyncio.create_task(_refresh_summary(note_id, new_title, content))
+    return f"Note updated: '{row['title']}' (id: {row['id']})."
+
+
+@tool
+async def propose_note_edit(
+    note_id: str,
+    edit_type: str,
+    proposed_content: str,
+    reasoning: str = "",
+    anchor_before: str = "",
+    anchor_text: str = "",
+    agent_id: str = "",
+    run_id: str = "",
+) -> str:
+    """Propose an AI edit to an existing note, pending human approval.
+
+    Use this instead of update_note when review_required is true.
+    The user will see the proposal highlighted before it is merged.
+
+    note_id: UUID of the target note (required)
+    edit_type: 'append' | 'insert' | 'replace'
+      - append: adds proposed_content at the end of the note
+      - insert: inserts proposed_content immediately after anchor_before text
+      - replace: replaces the first occurrence of anchor_text with proposed_content
+    proposed_content: the new Markdown text to add or substitute (required)
+    reasoning: brief explanation shown to the user (recommended)
+    anchor_before: for 'insert' — the text snippet that precedes the insertion point
+    anchor_text: for 'replace' — the exact text to be replaced
+    agent_id: agent identifier (for traceability)
+    run_id: run identifier (for traceability)
+    """
+    if edit_type not in ("append", "insert", "replace"):
+        return f"Invalid edit_type '{edit_type}'. Use 'append', 'insert', or 'replace'."
+
+    result = await execute_on_client(
+        action="propose_note_edit",
+        data={
+            "noteId": note_id,
+            "type": edit_type,
+            "proposedContent": proposed_content,
+            "reasoning": reasoning or None,
+            "anchorBefore": anchor_before or None,
+            "anchorText": anchor_text or None,
+            "agentId": agent_id or None,
+            "runId": run_id or None,
+        },
+    )
+    edit_id = result.get("id", "?")
+    return (
+        f"Edit proposal created (id: {edit_id}) for note {note_id}. "
+        f"Status: pending user approval."
+    )
+
+
+@tool
+async def delete_note(note_id: str) -> str:
+    """Delete a note permanently by its UUID."""
+    await execute_on_client(action="delete", table="notes", data={"id": note_id})
+    return f"Note {note_id} deleted."
+
+
+async def _refresh_summary(note_id: str, title: str, content: str) -> None:
+    """Generate and persist the AI summary for a note.  Fire-and-forget."""
+    try:
+        summary = await generate_note_summary(title, content)
+        if summary:
+            await execute_on_client(
+                action="update",
+                table="notes",
+                data={
+                    "id": note_id,
+                    "updates": {
+                        "aiSummary": summary,
+                        "aiSummaryUpdatedAt": int(__import__("time").time() * 1000),
+                    },
+                },
+            )
+    except Exception:
+        pass  # fire-and-forget; errors logged by generate_note_summary
+
+
+NOTE_TOOLS: list[Any] = [
+    list_notes,
+    get_note,
+    create_note,
+    update_note,
+    propose_note_edit,
+    delete_note,
+]
+
+NOTE_READ_TOOLS: list[Any] = [
+    list_notes,
+    get_note,
+]

app/agents/project_agent.py

@@ -1,7 +1,4 @@
-"""Project agent — full lifecycle management (list, get, create, update, archive, delete).
+"""Project agent — full lifecycle management (list, get, create, update, archive, delete)."""
-
-Shared tool definitions used by both Chat and Batch Agent services.
-"""
 
 from __future__ import annotations
 
@@ -9,23 +6,7 @@ from typing import Any
 
 from langchain_core.tools import tool
 
-from shared.ws_context import execute_on_client
+from app.core.ws_context import execute_on_client
-
-PROJECT_SYSTEM_PROMPT = (
-    "You are a project management assistant. You help users create, find,\n"
-    "update, and archive projects in their workspace.\n\n"
-    "Rules:\n"
-    "  - status must be one of: active, archived\n"
-    "  - client_id is optional; link to a client only when explicitly mentioned\n"
-    "  - ai_summary is populated only when the user asks for a project summary;\n"
-    "    derive it from context data — do not fabricate content\n"
-    "  - Use list_projects for scoped queries; list_all_projects only when the\n"
-    "    user wants a complete cross-client view including archived projects\n"
-    "  - get_project requires a project UUID; resolve the ID first by calling\n"
-    "    list_projects if you only have a project name\n"
-    "  - Prefer archiving (update_project status=archived) over deletion;\n"
-    "    only call delete_project when the user explicitly confirms deletion."
-)
 
 
 @tool
@@ -144,3 +125,9 @@ PROJECT_TOOLS: list[Any] = [
     update_project,
     delete_project,
 ]
+
+PROJECT_READ_TOOLS: list[Any] = [
+    list_projects,
+    list_all_projects,
+    get_project,
+]

app/agents/relations_agent.py

@@ -0,0 +1,63 @@
+"""Relations agent — read-only tool wrapping MemoryMiddleware.query_relations."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from langchain_core.tools import tool
+
+from app.core.memory_middleware import MemoryMiddleware
+from app.db import async_session
+
+# Injected at tool-factory time by _brief_research_tools(); not a module-level global.
+# Each tool closure captures the user_id bound at factory time.
+
+
+def make_query_relations_tool(user_id: str, trace_id: str | None = None) -> Any:
+    """Return a query_relations tool bound to *user_id*."""
+
+    @tool
+    async def query_relations(
+        subject_label: str = "",
+        predicate: str = "",
+        object_label: str = "",
+        limit: int = 10,
+    ) -> str:
+        """Query the relational memory graph for entity relationships.
+
+        Returns rows where subject ↔ predicate ↔ object match the given filters.
+        All parameters are optional — omit to retrieve all relations up to limit.
+
+        subject_label: entity label on the left side (e.g. a client name, "Acme Corp").
+        predicate: relationship type (e.g. "mentioned_in", "works_at", "related_to").
+        object_label: entity label on the right side (e.g. a project name, "Website Redesign").
+        limit: max rows to return (default 10).
+        """
+        import logging
+        logger = logging.getLogger(__name__)
+        logger.info(
+            "relations_agent: query_relations trace=%s user=%s subject=%r predicate=%r object=%r",
+            trace_id or "-", user_id, subject_label, predicate, object_label,
+        )
+
+        async with async_session() as db:
+            memory = MemoryMiddleware(db)
+            rows = await memory.query_relations(
+                user_id=user_id,
+                subject=subject_label or None,
+                predicate=predicate or None,
+                object_=object_label or None,
+                limit=limit,
+            )
+
+        if not rows:
+            return "No relational memory entries found for the given filters."
+
+        lines = [
+            f"- {r.subject_label} —[{r.predicate}]→ {r.object_label}"
+            + (f" (confidence: {r.confidence:.2f})" if r.confidence is not None else "")
+            for r in rows
+        ]
+        return f"Found {len(rows)} relation(s):\n" + "\n".join(lines)
+
+    return query_relations

app/agents/task_agent.py

@@ -0,0 +1,358 @@
+"""Task agent — full CRUD for tasks and task comments."""
+
+from __future__ import annotations
+
+from datetime import datetime, timezone
+import re
+from typing import Any
+
+from langchain_core.tools import tool
+
+from app.core.ws_context import execute_on_client
+
+_UUID_RE = re.compile(
+    r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-5][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$"
+)
+
+
+def _is_uuid(value: str) -> bool:
+    return bool(_UUID_RE.match(value))
+
+
+# ── Task tools ────────────────────────────────────────────────────────
+
+
+@tool
+async def list_tasks(
+    project_id: str = "",
+    status: str = "",
+    priority: str = "",
+    assignee: str = "",
+    search: str = "",
+    order_by: str = "",
+    order_dir: str = "",
+    due_date_from: int = -1,
+    due_date_to: int = -1,
+    created_at_from: int = -1,
+    created_at_to: int = -1,
+    completed_at_from: int = -1,
+    completed_at_to: int = -1,
+    is_ai_suggested: int = -1,
+    limit: int = 50,
+    offset: int = 0,
+) -> str:
+    """List tasks with optional filters. Returns up to `limit` results (default 50).
+
+    project_id: UUID of the project to scope results to.
+    status: filter by status — todo | in_progress | done.
+    priority: filter by priority — high | medium | low.
+    assignee: substring to match against assignee names. OMIT unless the user explicitly
+              names a person or refers to themselves ("my tasks", "assigned to me", "mine").
+              Do NOT default to the current user.
+    search: substring search across title and description.
+    order_by: sort field — dueDate | priority | createdAt | completedAt.
+    order_dir: asc (default) | desc.
+    due_date_from / due_date_to: ms epoch range for dueDate. Use -1 to omit.
+    created_at_from / created_at_to: ms epoch range for createdAt. Use -1 to omit.
+    completed_at_from / completed_at_to: ms epoch range for completedAt. Use -1 to omit.
+    is_ai_suggested: 0 or 1 to filter by AI-suggested flag; -1 = any.
+    limit: max rows to return (default 50). Use with offset to paginate.
+    offset: skip first N rows (default 0).
+
+    Tip — combine *_from and *_to for a closed range; pass only one for open-ended.
+    Tip — prefer count_tasks for "how many" questions to avoid listing rows.
+    Tip — for natural-language windows ("today", "tomorrow", "this week", "last month", etc.)
+    take due_date_from / due_date_to verbatim from the DATE CONTEXT block in the system prompt;
+    do not compute boundaries from the current UTC instant.
+    """
+    normalized_project_id = project_id if (project_id and _is_uuid(project_id)) else ""
+    filters: dict[str, Any] = {
+        "projectId": normalized_project_id or None,
+        "status": status or None,
+        "priority": priority or None,
+        "search": search or None,
+        "orderBy": order_by or None,
+        "orderDir": order_dir or None,
+        "limit": limit,
+        "offset": offset,
+    }
+    if assignee:
+        filters["assignee"] = assignee
+    if due_date_from != -1:
+        filters["dueDateFrom"] = due_date_from
+    if due_date_to != -1:
+        filters["dueDateTo"] = due_date_to
+    if created_at_from != -1:
+        filters["createdAtFrom"] = created_at_from
+    if created_at_to != -1:
+        filters["createdAtTo"] = created_at_to
+    if completed_at_from != -1:
+        filters["completedAtFrom"] = completed_at_from
+    if completed_at_to != -1:
+        filters["completedAtTo"] = completed_at_to
+    if is_ai_suggested != -1:
+        filters["isAiSuggested"] = is_ai_suggested
+
+    result = await execute_on_client(action="select", table="tasks", filters=filters)
+    rows = result.get("rows", [])
+    if not rows:
+        return "No tasks found matching the given filters."
+    lines = [
+        f"- {r['title']} (status: {r['status']}, priority: {r['priority']}, "
+        f"dueDate: {r.get('dueDate')}, completedAt: {r.get('completedAt')}, "
+        f"projectId: {r.get('projectId')}, id: {r['id']})"
+        for r in rows
+    ]
+    return f"Found {len(rows)} task(s):\n" + "\n".join(lines)
+
+
+@tool
+async def count_tasks(
+    project_id: str = "",
+    status: str = "",
+    priority: str = "",
+    assignee: str = "",
+    search: str = "",
+    due_date_from: int = -1,
+    due_date_to: int = -1,
+    created_at_from: int = -1,
+    created_at_to: int = -1,
+    completed_at_from: int = -1,
+    completed_at_to: int = -1,
+    is_ai_suggested: int = -1,
+) -> str:
+    """Count tasks matching the given filters without returning rows.
+
+    Use this instead of list_tasks for "how many" questions — it is much cheaper.
+    Same filter parameters as list_tasks (no limit/offset/order_by needed).
+    assignee: OMIT unless the user explicitly names a person or refers to themselves
+              ("my tasks"). Do NOT default to the current user.
+    due_date_from / due_date_to: ms epoch range for dueDate. Use -1 to omit.
+    created_at_from / created_at_to: ms epoch range for createdAt. Use -1 to omit.
+    completed_at_from / completed_at_to: ms epoch range for completedAt. Use -1 to omit.
+    Tip — for natural-language windows take due_date_from / due_date_to from the DATE CONTEXT block;
+    do not compute boundaries from the current UTC instant.
+    """
+    normalized_project_id = project_id if (project_id and _is_uuid(project_id)) else ""
+    filters: dict[str, Any] = {
+        "projectId": normalized_project_id or None,
+        "status": status or None,
+        "priority": priority or None,
+        "search": search or None,
+    }
+    if assignee:
+        filters["assignee"] = assignee
+    if due_date_from != -1:
+        filters["dueDateFrom"] = due_date_from
+    if due_date_to != -1:
+        filters["dueDateTo"] = due_date_to
+    if created_at_from != -1:
+        filters["createdAtFrom"] = created_at_from
+    if created_at_to != -1:
+        filters["createdAtTo"] = created_at_to
+    if completed_at_from != -1:
+        filters["completedAtFrom"] = completed_at_from
+    if completed_at_to != -1:
+        filters["completedAtTo"] = completed_at_to
+    if is_ai_suggested != -1:
+        filters["isAiSuggested"] = is_ai_suggested
+
+    result = await execute_on_client(action="count", table="tasks", filters=filters)
+    return f"Task count: {result.get('count', 0)}"
+
+
+@tool
+async def create_task(
+    title: str,
+    description: str = "",
+    status: str = "todo",
+    priority: str = "medium",
+    assignees: str = "[]",
+    due_date: int = 0,
+    project_id: str = "",
+    is_ai_suggested: int = 0,
+) -> str:
+    """Create a new task.
+    title: task title (required)
+    description: optional details
+    status: todo | in_progress | done  (default: todo)
+    priority: high | medium | low  (default: medium)
+    assignees: JSON-encoded array of assignee names, e.g. '["Alice"]'
+    due_date: Unix timestamp in milliseconds; 0 means no due date
+    project_id: optional UUID of the parent project
+    is_ai_suggested: 1 if proactively suggested, 0 if user-requested
+
+    completedAt is set automatically when status is 'done'.
+    """
+    result = await execute_on_client(
+        action="insert",
+        table="tasks",
+        data={
+            "title": title,
+            "description": description or None,
+            "status": status,
+            "priority": priority,
+            "assignee": assignees,
+            "dueDate": due_date or None,
+            "projectId": project_id or None,
+            "isAiSuggested": is_ai_suggested,
+        },
+    )
+    row = result["row"]
+    return (
+        f"Task created: '{row['title']}' "
+        f"(id: {row['id']}, status: {row['status']}, priority: {row['priority']}, projectId: {row.get('projectId')})"
+    )
+
+
+@tool
+async def update_task(
+    task_id: str,
+    title: str = "",
+    description: str = "",
+    status: str = "",
+    priority: str = "",
+    assignees: str = "",
+    due_date: int = -1,
+    project_id: str = "",
+) -> str:
+    """Update fields on an existing task. Only pass fields you want to change.
+    task_id: the task's UUID (required)
+    due_date: -1 means unchanged; 0 clears the due date; any positive value sets it
+
+    completedAt is managed automatically:
+      - setting status to 'done' records the current timestamp
+      - changing status away from 'done' clears completedAt
+    """
+    updates: dict[str, Any] = {}
+    if title:
+        updates["title"] = title
+    if description:
+        updates["description"] = description
+    if status:
+        updates["status"] = status
+    if priority:
+        updates["priority"] = priority
+    if assignees:
+        updates["assignee"] = assignees
+    if due_date != -1:
+        updates["dueDate"] = due_date or None
+    if project_id:
+        updates["projectId"] = project_id
+    result = await execute_on_client(
+        action="update",
+        table="tasks",
+        data={"id": task_id, "updates": updates},
+    )
+    row = result["row"]
+    return f"Task updated: '{row['title']}' (id: {row['id']}, status: {row['status']}, projectId: {row.get('projectId')})"
+
+
+@tool
+async def delete_task(task_id: str) -> str:
+    """Delete a task permanently by its UUID."""
+    await execute_on_client(action="delete", table="tasks", data={"id": task_id})
+    return f"Task {task_id} deleted."
+
+
+@tool
+async def list_tasks_due_today(user_timezone: str = "UTC", include_done: bool = False) -> str:
+    """List all tasks whose due date falls on today's date.
+
+    user_timezone: IANA timezone name (e.g. 'Europe/Rome', 'America/New_York').
+    Always pass the user's timezone so 'today' is computed in their local time.
+    include_done: set True to also include already-completed tasks due today (default False).
+    """
+    try:
+        from zoneinfo import ZoneInfo
+        tz = ZoneInfo(user_timezone or "UTC")
+    except Exception:
+        tz = timezone.utc
+    now_local = datetime.now(tz=tz)
+    start_dt = datetime(now_local.year, now_local.month, now_local.day, tzinfo=tz)
+    start_ms = int(start_dt.timestamp() * 1000)
+    end_ms = start_ms + 86_400_000 - 1
+    filters: dict[str, Any] = {"dueDateFrom": start_ms, "dueDateTo": end_ms}
+    if not include_done:
+        filters["status"] = "todo"
+    result = await execute_on_client(
+        action="select",
+        table="tasks",
+        filters=filters,
+    )
+    rows = result.get("rows", [])
+    if not rows:
+        return "No tasks are due today."
+    lines = [
+        f"- {r['title']} (priority: {r['priority']}, status: {r['status']}, "
+        f"projectId: {r.get('projectId')}, id: {r['id']})"
+        for r in rows
+    ]
+    return f"Tasks due today ({len(rows)}):\n" + "\n".join(lines)
+
+
+# ── Task comment tools ────────────────────────────────────────────────
+
+
+@tool
+async def list_task_comments(task_id: str) -> str:
+    """List all comments on a task by its UUID."""
+    result = await execute_on_client(
+        action="select",
+        table="taskComments",
+        filters={"taskId": task_id},
+    )
+    rows = result.get("rows", [])
+    if not rows:
+        return f"No comments found for task {task_id}."
+    lines = [f"- [{r['author']}]: {r['content']} (id: {r['id']})" for r in rows]
+    return f"Found {len(rows)} comment(s):\n" + "\n".join(lines)
+
+
+@tool
+async def add_task_comment(task_id: str, author: str, content: str) -> str:
+    """Add a comment to a task.
+    task_id: UUID of the task to comment on
+    author: name or ID of the comment author
+    content: comment text
+    """
+    result = await execute_on_client(
+        action="insert",
+        table="taskComments",
+        data={"taskId": task_id, "author": author, "content": content},
+    )
+    row = result.get("row", {})
+    row_author = row.get("author", author)
+    row_task_id = row.get("taskId") or row.get("task_id") or task_id
+    row_comment_id = row.get("id", "unknown")
+    return f"Comment added by {row_author} on task {row_task_id} (comment id: {row_comment_id})."
+
+
+@tool
+async def delete_task_comment(comment_id: str) -> str:
+    """Delete a task comment by its UUID."""
+    await execute_on_client(action="delete", table="taskComments", data={"id": comment_id})
+    return f"Comment {comment_id} deleted."
+
+
+# ── Agent ─────────────────────────────────────────────────────────────
+
+
+TASK_TOOLS: list[Any] = [
+    list_tasks,
+    count_tasks,
+    create_task,
+    update_task,
+    delete_task,
+    list_tasks_due_today,
+    list_task_comments,
+    add_task_comment,
+    delete_task_comment,
+]
+
+TASK_READ_TOOLS: list[Any] = [
+    list_tasks,
+    count_tasks,
+    list_tasks_due_today,
+    list_task_comments,
+]

app/agents/timeline_agent.py

@@ -0,0 +1,270 @@
+"""Timeline agent — project milestone management (list, create, update, delete)."""
+
+from __future__ import annotations
+
+import re
+from datetime import datetime, timezone
+from typing import Any
+
+from langchain_core.tools import tool
+
+from app.core.ws_context import execute_on_client
+
+_UUID_RE = re.compile(
+    r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-5][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$"
+)
+
+
+def _is_uuid(value: str) -> bool:
+    return bool(_UUID_RE.match(value))
+
+
+@tool
+async def list_timelines(
+    project_id: str = "",
+    type: str = "",
+    is_completed: int = -1,
+    is_ai_suggested: int = -1,
+    order_by: str = "",
+    order_dir: str = "",
+    date_from: int = -1,
+    date_to: int = -1,
+    created_at_from: int = -1,
+    created_at_to: int = -1,
+    completed_at_from: int = -1,
+    completed_at_to: int = -1,
+    limit: int = 50,
+    offset: int = 0,
+) -> str:
+    """List timeline events (milestones, checkpoints, activities) with optional filters.
+
+    project_id: UUID to scope results to a specific project.
+    type: filter by event type — milestone | checkpoint | activity.
+    is_completed: 0 = incomplete only, 1 = completed only, -1 = any (default).
+    is_ai_suggested: 0 or 1 to filter by AI-suggested flag; -1 = any.
+    order_by: sort field — date (default) | createdAt | completedAt.
+    order_dir: asc (default) | desc.
+    date_from / date_to: ms epoch range for the event date. Use -1 to omit.
+    created_at_from / created_at_to: ms epoch range for createdAt. Use -1 to omit.
+    completed_at_from / completed_at_to: ms epoch range for completedAt. Use -1 to omit.
+    limit: max rows to return (default 50). Use with offset to paginate.
+    offset: skip first N rows (default 0).
+
+    Tip — combine *_from and *_to for a closed range; pass only one for open-ended.
+    Tip — prefer count_timelines for "how many" questions to avoid listing rows.
+    Tip — for natural-language windows ("today", "this week", "last month", etc.)
+    take date_from / date_to verbatim from the DATE CONTEXT block in the system prompt;
+    do not compute boundaries from the current UTC instant.
+    """
+    normalized_project_id = project_id if (project_id and _is_uuid(project_id)) else ""
+    filters: dict[str, Any] = {
+        "projectId": normalized_project_id or None,
+        "orderBy": order_by or None,
+        "orderDir": order_dir or None,
+        "limit": limit,
+        "offset": offset,
+    }
+    if type:
+        filters["type"] = type
+    if is_completed != -1:
+        filters["isCompleted"] = is_completed
+    if is_ai_suggested != -1:
+        filters["isAiSuggested"] = is_ai_suggested
+    if date_from != -1:
+        filters["dateFrom"] = date_from
+    if date_to != -1:
+        filters["dateTo"] = date_to
+    if created_at_from != -1:
+        filters["createdAtFrom"] = created_at_from
+    if created_at_to != -1:
+        filters["createdAtTo"] = created_at_to
+    if completed_at_from != -1:
+        filters["completedAtFrom"] = completed_at_from
+    if completed_at_to != -1:
+        filters["completedAtTo"] = completed_at_to
+
+    result = await execute_on_client(action="select", table="timelines", filters=filters)
+    rows = result.get("rows", [])
+    if not rows:
+        return "No timeline events found."
+    lines = [
+        f"- {r['title']} (date: {r['date']}, type: {r.get('type')}, "
+        f"completed: {bool(r.get('isCompleted'))}, completedAt: {r.get('completedAt')}, "
+        f"projectId: {r.get('projectId')}, id: {r['id']})"
+        for r in rows
+    ]
+    return f"Found {len(rows)} timeline event(s):\n" + "\n".join(lines)
+
+
+@tool
+async def count_timelines(
+    project_id: str = "",
+    type: str = "",
+    is_completed: int = -1,
+    is_ai_suggested: int = -1,
+    date_from: int = -1,
+    date_to: int = -1,
+    created_at_from: int = -1,
+    created_at_to: int = -1,
+    completed_at_from: int = -1,
+    completed_at_to: int = -1,
+) -> str:
+    """Count timeline events matching the given filters without returning rows.
+
+    Use this instead of list_timelines for "how many" questions — it is much cheaper.
+    Same filter parameters as list_timelines (no limit/offset/order_by needed).
+
+    date_from / date_to: ms epoch range for the event date. Use -1 to omit.
+    completed_at_from / completed_at_to: ms epoch range for completedAt. Use -1 to omit.
+    Tip — for natural-language windows take date_from / date_to from the DATE CONTEXT block;
+    do not compute boundaries from the current UTC instant.
+    """
+    normalized_project_id = project_id if (project_id and _is_uuid(project_id)) else ""
+    filters: dict[str, Any] = {"projectId": normalized_project_id or None}
+    if type:
+        filters["type"] = type
+    if is_completed != -1:
+        filters["isCompleted"] = is_completed
+    if is_ai_suggested != -1:
+        filters["isAiSuggested"] = is_ai_suggested
+    if date_from != -1:
+        filters["dateFrom"] = date_from
+    if date_to != -1:
+        filters["dateTo"] = date_to
+    if created_at_from != -1:
+        filters["createdAtFrom"] = created_at_from
+    if created_at_to != -1:
+        filters["createdAtTo"] = created_at_to
+    if completed_at_from != -1:
+        filters["completedAtFrom"] = completed_at_from
+    if completed_at_to != -1:
+        filters["completedAtTo"] = completed_at_to
+
+    result = await execute_on_client(action="count", table="timelines", filters=filters)
+    return f"Timeline event count: {result.get('count', 0)}"
+
+
+@tool
+async def create_timeline(
+    project_id: str,
+    title: str,
+    date: int,
+    type: str = "milestone",
+    is_completed: int = 0,
+    is_ai_suggested: int = 0,
+) -> str:
+    """Create a project timeline event.
+    project_id: REQUIRED UUID of the parent project
+    title: descriptive name for the event
+    date: Unix timestamp in milliseconds for the event date
+    type: milestone (default) | checkpoint | activity
+    is_completed: 1 if already completed, 0 if not (default 0)
+    is_ai_suggested: 1 if proactively suggested, 0 if user-requested
+
+    completedAt is set automatically when is_completed is 1.
+    """
+    result = await execute_on_client(
+        action="insert",
+        table="timelines",
+        data={
+            "projectId": project_id,
+            "title": title,
+            "date": date,
+            "type": type,
+            "isCompleted": is_completed,
+            "isAiSuggested": is_ai_suggested,
+        },
+    )
+    row = result["row"]
+    return f"Timeline event created: '{row['title']}' (id: {row['id']}, date: {row['date']}, type: {row.get('type')})"
+
+
+@tool
+async def update_timeline(
+    timeline_id: str,
+    title: str = "",
+    date: int = -1,
+    is_completed: int = -1,
+) -> str:
+    """Update a timeline event. Only pass fields that should change.
+    timeline_id: UUID of the event (required)
+    date: -1 means unchanged; any other value sets the new date (ms timestamp)
+    is_completed: 0 = mark incomplete, 1 = mark complete, -1 = unchanged
+
+    completedAt is managed automatically:
+      - setting is_completed to 1 records the current timestamp
+      - setting is_completed to 0 clears completedAt
+    """
+    updates: dict[str, Any] = {}
+    if title:
+        updates["title"] = title
+    if date != -1:
+        updates["date"] = date
+    if is_completed != -1:
+        updates["isCompleted"] = is_completed
+    result = await execute_on_client(
+        action="update",
+        table="timelines",
+        data={"id": timeline_id, "updates": updates},
+    )
+    row = result["row"]
+    return f"Timeline event updated: '{row['title']}' (id: {row['id']})"
+
+
+@tool
+async def delete_timeline(timeline_id: str) -> str:
+    """Delete a timeline event permanently by its UUID."""
+    await execute_on_client(action="delete", table="timelines", data={"id": timeline_id})
+    return f"Timeline event {timeline_id} deleted."
+
+
+@tool
+async def list_timelines_today(user_timezone: str = "UTC", include_completed: bool = True) -> str:
+    """List all timeline events whose date falls on today.
+
+    user_timezone: IANA timezone name (e.g. 'Europe/Rome', 'America/New_York').
+    Always pass the user's timezone so 'today' is computed in their local time.
+    include_completed: set False to exclude already-completed events (default True).
+    """
+    try:
+        from zoneinfo import ZoneInfo
+        tz = ZoneInfo(user_timezone or "UTC")
+    except Exception:
+        tz = timezone.utc
+    now_local = datetime.now(tz=tz)
+    start_dt = datetime(now_local.year, now_local.month, now_local.day, tzinfo=tz)
+    start_ms = int(start_dt.timestamp() * 1000)
+    end_ms = start_ms + 86_400_000 - 1
+    filters: dict[str, Any] = {"dateFrom": start_ms, "dateTo": end_ms}
+    if not include_completed:
+        filters["isCompleted"] = 0
+    result = await execute_on_client(
+        action="select",
+        table="timelines",
+        filters=filters,
+    )
+    rows = result.get("rows", [])
+    if not rows:
+        return "No timeline events today."
+    lines = [
+        f"- {r['title']} (date: {r['date']}, type: {r.get('type')}, "
+        f"completed: {bool(r.get('isCompleted'))}, projectId: {r.get('projectId')}, id: {r['id']})"
+        for r in rows
+    ]
+    return f"Timeline events today ({len(rows)}):\n" + "\n".join(lines)
+
+
+TIMELINE_TOOLS: list[Any] = [
+    list_timelines,
+    count_timelines,
+    list_timelines_today,
+    create_timeline,
+    update_timeline,
+    delete_timeline,
+]
+
+TIMELINE_READ_TOOLS: list[Any] = [
+    list_timelines,
+    count_timelines,
+    list_timelines_today,
+]

app/api/deps.py

@@ -0,0 +1,14 @@
+"""Shared FastAPI dependencies.
+
+``get_current_user`` and ``oauth2_scheme`` live in ``app.api.middleware.auth``
+(the canonical location per Step 9).  This module re-exports them so that all
+existing route imports (``from app.api.deps import get_current_user``) continue
+to work without modification.
+
+Step 12 will update ``get_current_user`` to fetch the live tier from PostgreSQL
+instead of reading it from the JWT payload.
+"""
+
+from app.api.middleware.auth import get_current_user, oauth2_scheme  # noqa: F401
+
+__all__ = ["get_current_user", "oauth2_scheme"]

app/api/middleware/__init__.py

@@ -0,0 +1,19 @@
+"""API middleware package.
+
+Exports the three middleware components introduced in Step 9:
+  - Auth:        ``get_current_user`` FastAPI dependency + ``oauth2_scheme``
+  - Rate limit:  ``TierRateLimitMiddleware`` + ``limiter`` (slowapi Limiter)
+  - Sanitizer:   ``SanitizerMiddleware``
+"""
+
+from app.api.middleware.auth import get_current_user, oauth2_scheme
+from app.api.middleware.rate_limit import TierRateLimitMiddleware, limiter
+from app.api.middleware.sanitizer import SanitizerMiddleware
+
+__all__ = [
+    "get_current_user",
+    "oauth2_scheme",
+    "TierRateLimitMiddleware",
+    "limiter",
+    "SanitizerMiddleware",
+]

app/api/middleware/auth.py

@@ -0,0 +1,103 @@
+"""Auth middleware — JWT validation dependency.
+
+``get_current_user`` is the FastAPI dependency used by all protected routes.
+It decodes the Bearer JWT (identity + expiry), then fetches the current tier
+from the ``subscriptions`` table so that tier changes take effect immediately
+without requiring token re-issue.
+
+Exempt routes (no JWT required):
+  - POST /api/v1/auth/register
+  - POST /api/v1/auth/login
+  - POST /api/v1/billing/webhook
+"""
+
+from __future__ import annotations
+
+from fastapi import Depends, HTTPException, status
+from fastapi.security import OAuth2PasswordBearer
+from jose import JWTError, jwt
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.config.settings import settings
+from app.db import get_session
+from app.schemas import UserProfile
+
+oauth2_scheme = OAuth2PasswordBearer(tokenUrl="/api/v1/auth/login")
+
+
+async def get_current_user(
+    token: str = Depends(oauth2_scheme),
+    db: AsyncSession = Depends(get_session),
+) -> UserProfile:
+    """Validate a Bearer JWT and return the authenticated user.
+
+    The JWT is used for identity and expiry only.  The tier is fetched live
+    from the ``subscriptions`` table so that upgrades/downgrades take effect
+    immediately.  Falls back to ``'free'`` when no subscription row exists.
+
+    Raises HTTP 401 on any invalid or expired token.
+    """
+    credentials_exc = HTTPException(
+        status_code=status.HTTP_401_UNAUTHORIZED,
+        detail="Could not validate credentials",
+        headers={"WWW-Authenticate": "Bearer"},
+    )
+    try:
+        payload = jwt.decode(
+            token, settings.JWT_SECRET, algorithms=[settings.JWT_ALGORITHM]
+        )
+        user_id: str | None = payload.get("sub")
+        email: str | None = payload.get("email")
+        if not user_id or not email:
+            raise credentials_exc
+    except JWTError:
+        raise credentials_exc
+
+    # Live tier lookup — subscription row is the authoritative source.
+    # In dev, fall back to 'power' (unlimited) so quota limits don't
+    # block local development when no Stripe subscription exists.
+    from app.models import Subscription, User  # noqa: PLC0415
+
+    result = await db.execute(
+        select(Subscription.tier).where(Subscription.user_id == user_id)
+    )
+    default_tier = "power" if settings.ENV == "dev" else "free"
+    tier: str = result.scalar_one_or_none() or default_tier
+
+    # Fetch name/surname/avatar_url/onboarding_completed_at/password_hash from user row.
+    user_result = await db.execute(
+        select(
+            User.name, User.surname, User.avatar_url, User.onboarding_completed_at,
+            User.password_hash,
+        ).where(User.id == user_id)
+    )
+    user_row = user_result.one_or_none()
+
+    # Convert onboarding_completed_at to epoch ms (int) or None.
+    onboarding_ms: int | None = None
+    if user_row and user_row.onboarding_completed_at is not None:
+        onboarding_ms = int(user_row.onboarding_completed_at.timestamp() * 1000)
+
+    # Load decrypted core memory.
+    from app.core.memory_middleware import MemoryMiddleware  # noqa: PLC0415
+
+    memory_dict: dict[str, str] = {}
+    try:
+        mw = MemoryMiddleware(db)
+        blocks = await mw.list_core_blocks(user_id)
+        memory_dict = {b["label"]: b["value"] for b in blocks}
+    except Exception:
+        pass  # Non-critical — return empty memory on failure
+
+    return UserProfile(
+        id=user_id,
+        email=email,
+        name=user_row.name if user_row else None,
+        surname=user_row.surname if user_row else None,
+        avatar_url=user_row.avatar_url if user_row else None,
+        has_password=bool(user_row.password_hash) if user_row else False,
+        tier=tier,
+        onboarding_completed_at=onboarding_ms,
+        memory=memory_dict,
+    )  # type: ignore[arg-type]

app/api/middleware/rate_limit.py

@@ -0,0 +1,129 @@
+"""Tier-aware rate limiting middleware.
+
+Uses a per-user sliding-window counter (in-process, no Redis required).
+The ``slowapi`` Limiter is also exported for optional route-level decoration.
+
+Limits (requests per minute):
+  - free:  20
+  - pro:   60
+  - power: 120
+  - team:  200
+
+Exempt paths bypass the limiter entirely:
+  - POST /api/v1/auth/register
+  - POST /api/v1/auth/login
+  - POST /api/v1/billing/webhook
+  - GET  /api/v1/health
+"""
+
+from __future__ import annotations
+
+import json
+import time
+from collections import defaultdict
+
+from fastapi import Request, Response
+from jose import JWTError, jwt
+from slowapi import Limiter
+from slowapi.util import get_remote_address
+from starlette.middleware.base import BaseHTTPMiddleware
+from starlette.types import ASGIApp
+
+from app.config.settings import settings
+
+_TIER_LIMITS: dict[str, int] = {
+    "free": 20,
+    "pro": 60,
+    "power": 120,
+    "team": 200,
+}
+
+_EXEMPT_PATHS: frozenset[str] = frozenset(
+    {
+        "/api/v1/auth/register",
+        "/api/v1/auth/login",
+        "/api/v1/billing/webhook",
+        "/api/v1/health",
+    }
+)
+
+
+def _get_user_id_from_jwt(request: Request) -> str:
+    """Key function for the slowapi Limiter: returns JWT sub or remote IP."""
+    auth = request.headers.get("Authorization", "")
+    token = auth.removeprefix("Bearer ").strip()
+    if not token:
+        return get_remote_address(request)
+    try:
+        payload = jwt.decode(
+            token, settings.JWT_SECRET, algorithms=[settings.JWT_ALGORITHM]
+        )
+        return payload.get("sub") or get_remote_address(request)
+    except JWTError:
+        return get_remote_address(request)
+
+
+# Exported Limiter instance — available for optional route-level decoration.
+limiter = Limiter(key_func=_get_user_id_from_jwt)
+
+
+class TierRateLimitMiddleware(BaseHTTPMiddleware):
+    """Sliding-window rate limiter applied globally across all non-exempt routes.
+
+    Each authenticated user gets their own 60-second window sized by tier.
+    Unauthenticated requests pass through (the auth dependency will reject them
+    with 401 before the route handler runs).
+    """
+
+    def __init__(self, app: ASGIApp) -> None:
+        super().__init__(app)
+        # user_id → list of request timestamps (float, seconds since epoch)
+        self._window: dict[str, list[float]] = defaultdict(list)
+
+    async def dispatch(self, request: Request, call_next) -> Response:  # type: ignore[override]
+        if request.url.path in _EXEMPT_PATHS:
+            return await call_next(request)
+
+        # Extract JWT claims — if no valid token, pass through for auth dep to handle.
+        auth = request.headers.get("Authorization", "")
+        token = auth.removeprefix("Bearer ").strip()
+        if not token:
+            return await call_next(request)
+
+        try:
+            payload = jwt.decode(
+                token, settings.JWT_SECRET, algorithms=[settings.JWT_ALGORITHM]
+            )
+            user_id: str = payload.get("sub") or get_remote_address(request)
+            tier: str = payload.get("tier", "free")
+        except JWTError:
+            return await call_next(request)
+
+        limit = _TIER_LIMITS.get(tier, _TIER_LIMITS["free"])
+        now = time.monotonic()
+        window_start = now - 60.0
+
+        # Slide the window: discard timestamps older than 60 seconds.
+        timestamps = [t for t in self._window[user_id] if t > window_start]
+
+        if len(timestamps) >= limit:
+            retry_after = max(1, int(60 - (now - min(timestamps))))
+            return Response(
+                content=json.dumps(
+                    {
+                        "detail": (
+                            f"Rate limit exceeded ({limit} req/min for {tier} tier). "
+                            f"Retry in {retry_after}s."
+                        )
+                    }
+                ),
+                status_code=429,
+                headers={
+                    "Retry-After": str(retry_after),
+                    "Content-Type": "application/json",
+                },
+            )
+
+        timestamps.append(now)
+        self._window[user_id] = timestamps
+        return await call_next(request)

app/api/middleware/sanitizer.py

@@ -0,0 +1,138 @@
+"""Response sanitizer middleware.
+
+Scans JSON responses from the /api/v1/chat endpoint and strips any fragments
+that could reveal server-side prompt IP:
+  - System prompt openers ("You are a/an/the …")
+  - Agent routing metadata ("Available agents:", "intent classifier", …)
+  - LangChain tool schema fragments (``"type": "function"``)
+  - Internal reasoning markers (<thinking>, <reasoning>, [INST], …)
+  - Exact-match known prompt fingerprints
+
+The middleware only activates for paths under /api/v1/chat.
+
+Any sanitisation event is logged as a WARNING with the request path and the
+names of the fields that were modified.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+
+from fastapi import Request, Response
+from starlette.middleware.base import BaseHTTPMiddleware
+from starlette.types import ASGIApp
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Detection patterns — order matters: fingerprints checked first (exact),
+# then compiled regexes.
+# ---------------------------------------------------------------------------
+
+_FINGERPRINTS: tuple[str, ...] = (
+    "You are an intent classifier",
+    "Respond with just the agent name",
+    "Summarize these agent results",
+    "Available agents:",
+    "route to:",
+)
+
+_PATTERNS: tuple[re.Pattern[str], ...] = (
+    re.compile(r"You are (a|an|the)\b.{0,200}", re.IGNORECASE | re.DOTALL),
+    re.compile(r"Available agents\s*:", re.IGNORECASE),
+    re.compile(r"\bintent classifier\b", re.IGNORECASE),
+    re.compile(r'"type"\s*:\s*"function"'),           # LangChain tool schema
+    re.compile(r"<(thinking|reasoning|system|prompt)>", re.IGNORECASE),
+    re.compile(r"\[INST\]|\[/INST\]"),                # Llama instruct markers
+    re.compile(r"route\s+to\s*:", re.IGNORECASE),
+    re.compile(r"prompt_template\s*:\s*['\"].{10,}", re.IGNORECASE),
+)
+
+
+def _sanitize_text(text: str) -> tuple[str, bool]:
+    """Scan *text* for prompt fragments and replace matches with ``[REDACTED]``.
+
+    Returns ``(cleaned_text, was_changed)``.
+    """
+    # Fingerprint check — if any exact phrase is present, redact the whole string.
+    for fp in _FINGERPRINTS:
+        if fp in text:
+            return "[REDACTED]", True
+
+    changed = False
+    for pattern in _PATTERNS:
+        new_text, n = pattern.subn("[REDACTED]", text)
+        if n:
+            text = new_text
+            changed = True
+
+    return text, changed
+
+
+class SanitizerMiddleware(BaseHTTPMiddleware):
+    """Strip prompt IP from /api/v1/chat JSON responses."""
+
+    def __init__(self, app: ASGIApp) -> None:
+        super().__init__(app)
+
+    async def dispatch(self, request: Request, call_next) -> Response:  # type: ignore[override]
+        response: Response = await call_next(request)
+
+        # Only process chat endpoint responses.
+        if not request.url.path.startswith("/api/v1/chat"):
+            return response
+
+        # Read body — collect streaming chunks.
+        body_bytes = b""
+        async for chunk in response.body_iterator:
+            body_bytes += chunk if isinstance(chunk, bytes) else chunk.encode()
+
+        # Skip non-JSON bodies (shouldn't happen on /chat, but be safe).
+        try:
+            body = json.loads(body_bytes.decode("utf-8"))
+        except (json.JSONDecodeError, UnicodeDecodeError):
+            return Response(
+                content=body_bytes,
+                status_code=response.status_code,
+                headers=dict(response.headers),
+                media_type=response.media_type,
+            )
+
+        if not isinstance(body, dict):
+            return Response(
+                content=body_bytes,
+                status_code=response.status_code,
+                headers=dict(response.headers),
+                media_type=response.media_type,
+            )
+
+        # Walk top-level string fields and sanitise.
+        sanitised_fields: list[str] = []
+        for key, value in body.items():
+            if isinstance(value, str):
+                cleaned, changed = _sanitize_text(value)
+                if changed:
+                    body[key] = cleaned
+                    sanitised_fields.append(key)
+
+        if sanitised_fields:
+            logger.warning(
+                "Sanitizer redacted prompt fragments",
+                extra={
+                    "path": request.url.path,
+                    "fields": sanitised_fields,
+                },
+            )
+
+        new_body = json.dumps(body).encode("utf-8")
+        headers = dict(response.headers)
+        headers["content-length"] = str(len(new_body))
+
+        return Response(
+            content=new_body,
+            status_code=response.status_code,
+            headers=headers,
+            media_type="application/json",
+        )

app/api/routes/auth.py

@@ -0,0 +1,795 @@
+"""Auth routes: register, login, refresh, me, OAuth social login, onboarding.
+
+Users and refresh tokens are persisted in PostgreSQL (users + refresh_tokens
+tables).  Passwords are hashed with bcrypt; refresh tokens are stored as
+SHA-256 hashes so plaintext never reaches the DB.
+
+OAuth (Google):
+  GET  /auth/oauth/{provider}/authorize  — returns consent-screen URL + state
+  POST /auth/oauth/{provider}/callback   — exchanges code, issues JWT tokens
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import time
+import urllib.parse
+import uuid
+from datetime import datetime, timedelta, timezone
+from typing import Literal
+
+import bcrypt
+from cryptography.fernet import Fernet
+from fastapi import APIRouter, Depends, HTTPException, status
+from fastapi.responses import RedirectResponse
+from jose import jwt
+from pydantic import BaseModel, Field
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.api.deps import get_current_user
+from app.auth.oauth_providers import GoogleOAuthProvider, generate_pkce_pair
+from app.config.settings import settings
+from app.core.llm import get_llm
+from app.core.memory_middleware import MemoryMiddleware
+from app.db import get_session
+from app.models import OAuthAccount, RefreshToken, User
+from app.schemas import AuthTokens, UserProfile
+
+router = APIRouter(prefix="/auth", tags=["auth"])
+
+
+# ── OAuth provider registry ───────────────────────────────────────────
+
+def _get_google_provider() -> GoogleOAuthProvider:
+    if not settings.GOOGLE_AUTH_CLIENT_ID or not settings.GOOGLE_AUTH_CLIENT_SECRET:
+        raise HTTPException(
+            status.HTTP_503_SERVICE_UNAVAILABLE,
+            "Google login is not configured on this server",
+        )
+    return GoogleOAuthProvider(
+        client_id=settings.GOOGLE_AUTH_CLIENT_ID,
+        client_secret=settings.GOOGLE_AUTH_CLIENT_SECRET,
+        redirect_uri=settings.OAUTH_REDIRECT_URI,
+    )
+
+
+_PROVIDERS = {"google": _get_google_provider}
+
+# In-memory state store: state → (code_verifier, expires_at_epoch_s)
+# Production note: replace with Redis for multi-process deployments.
+_pending_states: dict[str, tuple[str, float]] = {}
+_STATE_TTL_SECONDS = 600  # 10 minutes
+
+
+# ── Internal helpers ─────────────────────────────────────────────────
+
+
+def _hash_password(password: str) -> str:
+    return bcrypt.hashpw(password.encode(), bcrypt.gensalt()).decode()
+
+
+def _verify_password(password: str, hashed: str) -> bool:
+    return bcrypt.checkpw(password.encode(), hashed.encode())
+
+
+def _hash_token(plain_token: str) -> str:
+    """SHA-256 of the plain refresh token string."""
+    return hashlib.sha256(plain_token.encode()).hexdigest()
+
+
+def _make_access_token(user_id: str, email: str, tier: str) -> tuple[str, int]:
+    """Return (signed JWT, expires_at_ms)."""
+    now = int(time.time())
+    exp = now + settings.JWT_ACCESS_TOKEN_EXPIRE_MINUTES * 60
+    payload = {
+        "sub": user_id,
+        "email": email,
+        "tier": tier,
+        "exp": exp,
+        "iat": now,
+    }
+    token = jwt.encode(payload, settings.JWT_SECRET, algorithm=settings.JWT_ALGORITHM)
+    return token, exp * 1000  # ms for client
+
+
+# ── Request bodies ────────────────────────────────────────────────────
+
+
+class _RegisterRequest(BaseModel):
+    email: str
+    password: str
+    name: str | None = None
+    surname: str | None = None
+
+
+class _LoginRequest(BaseModel):
+    email: str
+    password: str
+
+
+class _RefreshRequest(BaseModel):
+    refresh_token: str
+
+
+# ── Routes ────────────────────────────────────────────────────────────
+
+
+@router.post("/register", response_model=AuthTokens, status_code=status.HTTP_201_CREATED)
+async def register(
+    body: _RegisterRequest,
+    db: AsyncSession = Depends(get_session),
+) -> AuthTokens:
+    """Create a new account and return JWT tokens."""
+    existing = await db.execute(select(User).where(User.email == body.email))
+    if existing.scalar_one_or_none() is not None:
+        raise HTTPException(status.HTTP_409_CONFLICT, "Email already registered")
+
+    user = User(
+        id=str(uuid.uuid4()),
+        email=body.email,
+        name=body.name,
+        surname=body.surname,
+        password_hash=_hash_password(body.password),
+        tier="free",
+        encryption_key=Fernet.generate_key().decode(),
+    )
+    db.add(user)
+    await db.flush()  # get user.id without committing
+
+    plain_token = str(uuid.uuid4())
+    expires_at = datetime.now(timezone.utc) + timedelta(
+        days=settings.JWT_REFRESH_TOKEN_EXPIRE_DAYS
+    )
+    rt = RefreshToken(
+        user_id=user.id,
+        token_hash=_hash_token(plain_token),
+        expires_at=expires_at,
+    )
+    db.add(rt)
+    await db.commit()
+
+    access_token, expires_at_ms = _make_access_token(user.id, user.email, user.tier)
+    return AuthTokens(
+        access_token=access_token,
+        refresh_token=plain_token,
+        expires_at=expires_at_ms,
+    )
+
+
+@router.post("/login", response_model=AuthTokens)
+async def login(
+    body: _LoginRequest,
+    db: AsyncSession = Depends(get_session),
+) -> AuthTokens:
+    """Validate credentials and return JWT tokens."""
+    result = await db.execute(select(User).where(User.email == body.email))
+    user = result.scalar_one_or_none()
+    if user is None or not _verify_password(body.password, user.password_hash):
+        raise HTTPException(status.HTTP_401_UNAUTHORIZED, "Invalid credentials")
+
+    plain_token = str(uuid.uuid4())
+    expires_at = datetime.now(timezone.utc) + timedelta(
+        days=settings.JWT_REFRESH_TOKEN_EXPIRE_DAYS
+    )
+    rt = RefreshToken(
+        user_id=user.id,
+        token_hash=_hash_token(plain_token),
+        expires_at=expires_at,
+    )
+    db.add(rt)
+    await db.commit()
+
+    access_token, expires_at_ms = _make_access_token(user.id, user.email, user.tier)
+    return AuthTokens(
+        access_token=access_token,
+        refresh_token=plain_token,
+        expires_at=expires_at_ms,
+    )
+
+
+@router.post("/refresh", response_model=AuthTokens)
+async def refresh(
+    body: _RefreshRequest,
+    db: AsyncSession = Depends(get_session),
+) -> AuthTokens:
+    """Rotate a refresh token and return a new token pair."""
+    token_hash = _hash_token(body.refresh_token)
+    result = await db.execute(
+        select(RefreshToken).where(RefreshToken.token_hash == token_hash)
+    )
+    rt = result.scalar_one_or_none()
+
+    now = datetime.now(timezone.utc)
+    if rt is None or rt.expires_at.replace(tzinfo=timezone.utc) < now:
+        raise HTTPException(status.HTTP_401_UNAUTHORIZED, "Invalid or expired refresh token")
+
+    # Rotate: delete old token, issue new one.
+    await db.delete(rt)
+
+    user_result = await db.execute(select(User).where(User.id == rt.user_id))
+    user = user_result.scalar_one_or_none()
+    if user is None:
+        raise HTTPException(status.HTTP_401_UNAUTHORIZED, "User not found")
+
+    plain_token = str(uuid.uuid4())
+    new_expires = now + timedelta(days=settings.JWT_REFRESH_TOKEN_EXPIRE_DAYS)
+    new_rt = RefreshToken(
+        user_id=user.id,
+        token_hash=_hash_token(plain_token),
+        expires_at=new_expires,
+    )
+    db.add(new_rt)
+    await db.commit()
+
+    access_token, expires_at_ms = _make_access_token(user.id, user.email, user.tier)
+    return AuthTokens(
+        access_token=access_token,
+        refresh_token=plain_token,
+        expires_at=expires_at_ms,
+    )
+
+
+class _UpdateProfileRequest(BaseModel):
+    name: str | None = None
+    surname: str | None = None
+
+
+@router.get("/me", response_model=UserProfile)
+async def me(current_user: UserProfile = Depends(get_current_user)) -> UserProfile:
+    """Return the profile for the authenticated user."""
+    return current_user
+
+
+@router.put("/me", response_model=UserProfile)
+async def update_profile(
+    body: _UpdateProfileRequest,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> UserProfile:
+    """Update the authenticated user's name and surname."""
+    result = await db.execute(select(User).where(User.id == current_user.id))
+    user = result.scalar_one()
+
+    if body.name is not None:
+        user.name = body.name
+    if body.surname is not None:
+        user.surname = body.surname
+
+    await db.commit()
+    await db.refresh(user)
+
+    return UserProfile(
+        id=user.id,
+        email=user.email,
+        name=user.name,
+        surname=user.surname,
+        avatar_url=user.avatar_url,
+        tier=current_user.tier,
+    )
+
+
+# ── OAuth helpers ─────────────────────────────────────────────────────
+
+
+async def _issue_refresh_token(user: User, db: AsyncSession) -> tuple[str, AuthTokens]:
+    """Create a refresh token row and return (plain_token, AuthTokens)."""
+    plain_token = str(uuid.uuid4())
+    expires_at = datetime.now(timezone.utc) + timedelta(
+        days=settings.JWT_REFRESH_TOKEN_EXPIRE_DAYS
+    )
+    rt = RefreshToken(
+        user_id=user.id,
+        token_hash=_hash_token(plain_token),
+        expires_at=expires_at,
+    )
+    db.add(rt)
+    access_token, expires_at_ms = _make_access_token(user.id, user.email, user.tier)
+    return plain_token, AuthTokens(
+        access_token=access_token,
+        refresh_token=plain_token,
+        expires_at=expires_at_ms,
+    )
+
+
+# ── OAuth request/response schemas ───────────────────────────────────
+
+
+class _OAuthAuthorizeResponse(BaseModel):
+    url: str
+    state: str
+
+
+class _OAuthCallbackRequest(BaseModel):
+    code: str
+    state: str
+
+
+# ── OAuth routes ──────────────────────────────────────────────────────
+
+
+@router.get(
+    "/oauth/{provider}/web-callback",
+    summary="Web-facing OAuth redirect — bounces to the adiuvai:// deep link",
+    include_in_schema=False,
+)
+async def oauth_web_callback(
+    provider: Literal["google"],
+    code: str,
+    state: str,
+) -> RedirectResponse:
+    """Google redirects here after user consent.
+
+    This endpoint immediately redirects to the Electron deep-link URI so the
+    desktop app receives the authorization code.  It is intentionally simple —
+    no state validation here (the Electron app + backend callback do that).
+
+    Registered in Google Cloud Console as:
+      http://localhost:8000/api/v1/auth/oauth/google/web-callback  (dev)
+      https://api.adiuvai.com/api/v1/auth/oauth/google/web-callback  (prod)
+    """
+    params = urllib.parse.urlencode({"code": code, "state": state, "provider": provider})
+    deep_link = f"adiuvai://oauth/callback?{params}"
+    return RedirectResponse(url=deep_link, status_code=302)
+
+
+@router.get(
+    "/oauth/{provider}/authorize",
+    response_model=_OAuthAuthorizeResponse,
+    summary="Start OAuth flow — returns the provider consent-screen URL",
+)
+async def oauth_authorize(
+    provider: Literal["google"],
+) -> _OAuthAuthorizeResponse:
+    """Generate a PKCE state + code_challenge and return the authorization URL.
+
+    The client opens this URL in the system browser.  After the user grants
+    consent, the provider redirects to the deep-link URI (adiuvai://oauth/callback)
+    with ``code`` and ``state`` query params.  The client then calls
+    ``POST /auth/oauth/{provider}/callback`` with those values.
+    """
+    provider_factory = _PROVIDERS.get(provider)
+    if provider_factory is None:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, f"Unknown provider: {provider}")
+
+    oauth_provider = provider_factory()
+    state = str(uuid.uuid4())
+    code_verifier, code_challenge = generate_pkce_pair()
+
+    # Purge expired states to prevent unbounded growth.
+    now = time.time()
+    expired = [s for s, (_, exp) in _pending_states.items() if exp < now]
+    for s in expired:
+        del _pending_states[s]
+
+    _pending_states[state] = (code_verifier, now + _STATE_TTL_SECONDS)
+
+    url = oauth_provider.get_authorization_url(state=state, code_challenge=code_challenge)
+    return _OAuthAuthorizeResponse(url=url, state=state)
+
+
+@router.post(
+    "/oauth/{provider}/callback",
+    response_model=AuthTokens,
+    summary="Complete OAuth flow — exchange code and issue JWT tokens",
+)
+async def oauth_callback(
+    provider: Literal["google"],
+    body: _OAuthCallbackRequest,
+    db: AsyncSession = Depends(get_session),
+) -> AuthTokens:
+    """Validate state, exchange the authorization code, and sign in (or register) the user.
+
+    Resolution order:
+      1. ``oauth_accounts`` row match → existing user, log in.
+      2. Email match + ``email_verified=True`` → link OAuth account to existing user.
+      3. No match → create new user (password_hash=None, avatar from provider).
+    """
+    provider_factory = _PROVIDERS.get(provider)
+    if provider_factory is None:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, f"Unknown provider: {provider}")
+
+    # Validate state (CSRF protection).
+    now = time.time()
+    entry = _pending_states.pop(body.state, None)
+    if entry is None or entry[1] < now:
+        raise HTTPException(status.HTTP_401_UNAUTHORIZED, "Invalid or expired OAuth state")
+
+    code_verifier, _ = entry
+
+    oauth_provider = provider_factory()
+
+    # Exchange code for tokens.
+    try:
+        token_data = await oauth_provider.exchange_code(
+            code=body.code,
+            code_verifier=code_verifier,
+            redirect_uri=settings.OAUTH_REDIRECT_URI,
+        )
+    except Exception:
+        raise HTTPException(
+            status.HTTP_400_BAD_REQUEST, "Failed to exchange authorization code"
+        )
+
+    access_token_google = token_data.get("access_token")
+    if not access_token_google:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "No access token in provider response")
+
+    # Fetch user identity.
+    try:
+        userinfo = await oauth_provider.get_userinfo(access_token_google)
+    except Exception:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "Failed to fetch user info from provider")
+
+    # ── Resolution order ──────────────────────────────────────────────
+
+    # 1. Existing OAuth link?
+    oauth_result = await db.execute(
+        select(OAuthAccount).where(
+            OAuthAccount.provider == provider,
+            OAuthAccount.provider_user_id == userinfo.provider_user_id,
+        )
+    )
+    oauth_account = oauth_result.scalar_one_or_none()
+
+    if oauth_account is not None:
+        user_result = await db.execute(select(User).where(User.id == oauth_account.user_id))
+        user = user_result.scalar_one()
+        # Backfill avatar if the user doesn't have one yet.
+        if user.avatar_url is None and userinfo.avatar_url:
+            user.avatar_url = userinfo.avatar_url
+            await db.commit()
+        plain_token, tokens = await _issue_refresh_token(user, db)
+        await db.commit()
+        return tokens
+
+    # 2. Email match with a verified Google email → link accounts.
+    if userinfo.email_verified:
+        email_result = await db.execute(select(User).where(User.email == userinfo.email))
+        existing_user = email_result.scalar_one_or_none()
+
+        if existing_user is not None:
+            new_link = OAuthAccount(
+                user_id=existing_user.id,
+                provider=provider,
+                provider_user_id=userinfo.provider_user_id,
+                provider_email=userinfo.email,
+            )
+            db.add(new_link)
+            if existing_user.avatar_url is None and userinfo.avatar_url:
+                existing_user.avatar_url = userinfo.avatar_url
+            plain_token, tokens = await _issue_refresh_token(existing_user, db)
+            await db.commit()
+            return tokens
+
+    # Guard: if the email is already taken but we couldn't auto-link (e.g.
+    # email_verified=False), refuse with 409 instead of hitting a DB constraint.
+    if not userinfo.email_verified:
+        conflict = await db.execute(select(User).where(User.email == userinfo.email))
+        if conflict.scalar_one_or_none() is not None:
+            raise HTTPException(
+                status.HTTP_409_CONFLICT,
+                "An account with this email already exists. "
+                "Please sign in with your password.",
+            )
+
+    # 3. New user — social-only account (no password).
+    new_user = User(
+        id=str(uuid.uuid4()),
+        email=userinfo.email,
+        name=userinfo.name,
+        password_hash=None,
+        avatar_url=userinfo.avatar_url,
+        tier="free",
+        encryption_key=Fernet.generate_key().decode(),
+    )
+    db.add(new_user)
+    await db.flush()  # populate new_user.id
+
+    new_oauth = OAuthAccount(
+        user_id=new_user.id,
+        provider=provider,
+        provider_user_id=userinfo.provider_user_id,
+        provider_email=userinfo.email,
+    )
+    db.add(new_oauth)
+
+    plain_token, tokens = await _issue_refresh_token(new_user, db)
+    await db.commit()
+    return tokens
+
+
+# ── Onboarding helpers ────────────────────────────────────────────────
+
+
+async def _build_profile(user_id: str, email: str, db: AsyncSession) -> UserProfile:
+    """Re-fetch and return a full UserProfile (reuses get_current_user logic)."""
+
+    # We can't call the FastAPI dependency directly, but we can replicate
+    # the core logic inline.  Instead, we just re-query the same way.
+    from app.models import Subscription  # noqa: PLC0415
+
+    result = await db.execute(
+        select(Subscription.tier).where(Subscription.user_id == user_id)
+    )
+    default_tier = "power" if settings.ENV == "dev" else "free"
+    tier: str = result.scalar_one_or_none() or default_tier
+
+    user_result = await db.execute(
+        select(
+            User.name, User.surname, User.avatar_url, User.onboarding_completed_at,
+            User.password_hash,
+        ).where(User.id == user_id)
+    )
+    user_row = user_result.one_or_none()
+
+    onboarding_ms: int | None = None
+    if user_row and user_row.onboarding_completed_at is not None:
+        onboarding_ms = int(user_row.onboarding_completed_at.timestamp() * 1000)
+
+    memory_dict: dict[str, str] = {}
+    try:
+        mw = MemoryMiddleware(db)
+        blocks = await mw.list_core_blocks(user_id)
+        memory_dict = {b["label"]: b["value"] for b in blocks}
+    except Exception:
+        pass
+
+    return UserProfile(
+        id=user_id,
+        email=email,
+        name=user_row.name if user_row else None,
+        surname=user_row.surname if user_row else None,
+        avatar_url=user_row.avatar_url if user_row else None,
+        has_password=bool(user_row.password_hash) if user_row else False,
+        tier=tier,
+        onboarding_completed_at=onboarding_ms,
+        memory=memory_dict,
+    )
+
+
+# ── Onboarding routes ────────────────────────────────────────────────
+
+
+class _UpdateMemoryRequest(BaseModel):
+    memory: dict[str, str] = Field(default_factory=dict)
+    mark_onboarded: bool = False
+
+
+@router.put("/me/memory", response_model=UserProfile)
+async def update_memory(
+    body: _UpdateMemoryRequest,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> UserProfile:
+    """Update core memory key/value pairs and optionally mark onboarding complete."""
+    mw = MemoryMiddleware(db)
+    for key, value in body.memory.items():
+        await mw.update_core(current_user.id, key, value)
+    if body.mark_onboarded:
+        result = await db.execute(select(User).where(User.id == current_user.id))
+        user = result.scalar_one()
+        user.onboarding_completed_at = datetime.now(timezone.utc)
+        await db.commit()
+    return await _build_profile(current_user.id, current_user.email, db)
+
+
+@router.post("/me/onboarding/reset")
+async def reset_onboarding(
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+):
+    """Reset onboarding so the wizard runs again on next login."""
+    result = await db.execute(select(User).where(User.id == current_user.id))
+    user = result.scalar_one()
+    user.onboarding_completed_at = None
+    await db.commit()
+    return {"status": "reset"}
+
+
+class _NormalizeRequest(BaseModel):
+    inputs: dict[str, str]
+
+
+class _NormalizeResponse(BaseModel):
+    normalized: dict[str, str]
+
+
+@router.post("/onboarding/normalize", response_model=_NormalizeResponse)
+async def normalize_onboarding(
+    body: _NormalizeRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> _NormalizeResponse:
+    """One-shot LLM normalization for free-text onboarding answers."""
+    if not body.inputs:
+        return _NormalizeResponse(normalized={})
+    try:
+        llm = get_llm(model="gpt-4o-mini", temperature=0)
+        prompt = (
+            "You normalize user onboarding answers into clean, ≤3-word canonical labels.\n"
+            "Return a JSON object with the same keys and normalized values.\n"
+            "Examples: 'i build websites' → 'Web Developer', 'tech-ish stuff' → 'Technology'\n"
+            f"Input: {json.dumps(body.inputs)}"
+        )
+        response = await llm.ainvoke(
+            [
+                {"role": "system", "content": "You normalize user inputs. Return JSON only."},
+                {"role": "user", "content": prompt},
+            ],
+        )
+        normalized = json.loads(response.content)
+        return _NormalizeResponse(normalized=normalized)
+    except Exception:
+        # LLM failure must never block onboarding — return inputs unchanged
+        return _NormalizeResponse(normalized=body.inputs)
+
+
+# ── Password management ───────────────────────────────────────────────
+
+
+class _ChangePasswordRequest(BaseModel):
+    current_password: str = Field(min_length=1)
+    new_password: str = Field(min_length=8)
+
+
+@router.put("/me/password", status_code=status.HTTP_200_OK)
+async def change_password(
+    body: _ChangePasswordRequest,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Change the authenticated user's password.
+
+    Requires the current password for verification.
+    Returns 400 for social-only users (no password set).
+    """
+    result = await db.execute(select(User).where(User.id == current_user.id))
+    user = result.scalar_one()
+
+    if user.password_hash is None:
+        raise HTTPException(
+            status.HTTP_400_BAD_REQUEST,
+            "This account uses social login and has no password to change",
+        )
+
+    if not _verify_password(body.current_password, user.password_hash):
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "Current password is incorrect")
+
+    user.password_hash = _hash_password(body.new_password)
+    await db.commit()
+    return {"ok": True}
+
+
+# ── OAuth account management ─────────────────────────────────────────
+
+
+@router.get("/me/oauth-accounts", response_model=list[dict])
+async def list_oauth_accounts(
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> list[dict]:
+    """List all OAuth providers linked to the authenticated user."""
+    result = await db.execute(
+        select(OAuthAccount).where(OAuthAccount.user_id == current_user.id)
+    )
+    accounts = result.scalars().all()
+    return [
+        {
+            "provider": a.provider,
+            "provider_email": a.provider_email,
+            "created_at": int(a.created_at.timestamp() * 1000),
+        }
+        for a in accounts
+    ]
+
+
+@router.delete("/me/oauth-accounts/{provider}", status_code=status.HTTP_200_OK)
+async def unlink_oauth_account(
+    provider: str,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Unlink an OAuth provider from the authenticated user.
+
+    Refuses if the user has no password and this is their only login method.
+    """
+    result = await db.execute(select(User).where(User.id == current_user.id))
+    user = result.scalar_one()
+
+    oauth_result = await db.execute(
+        select(OAuthAccount).where(
+            OAuthAccount.user_id == current_user.id,
+            OAuthAccount.provider == provider,
+        )
+    )
+    account = oauth_result.scalar_one_or_none()
+    if account is None:
+        raise HTTPException(status.HTTP_404_NOT_FOUND, f"No linked {provider} account found")
+
+    # Safety: don't let users lock themselves out.
+    all_oauth = await db.execute(
+        select(OAuthAccount).where(OAuthAccount.user_id == current_user.id)
+    )
+    oauth_count = len(all_oauth.scalars().all())
+
+    if user.password_hash is None and oauth_count <= 1:
+        raise HTTPException(
+            status.HTTP_400_BAD_REQUEST,
+            "Cannot unlink the only login method. Set a password first.",
+        )
+
+    await db.delete(account)
+    await db.commit()
+    return {"ok": True}
+
+
+# ── Avatar update ─────────────────────────────────────────────────────
+
+
+class _UpdateAvatarRequest(BaseModel):
+    avatar_url: str = Field(min_length=1)
+
+
+@router.put("/me/avatar", response_model=UserProfile)
+async def update_avatar(
+    body: _UpdateAvatarRequest,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> UserProfile:
+    """Update the authenticated user's avatar URL.
+
+    Accepts {"avatar_url": "https://..."} — the client uploads the image
+    to its own storage and passes the resulting URL here.
+    """
+    if not body.avatar_url.startswith(("https://", "http://", "data:image/")):
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "Invalid avatar URL")
+
+    result = await db.execute(select(User).where(User.id == current_user.id))
+    user = result.scalar_one()
+    user.avatar_url = body.avatar_url
+    await db.commit()
+
+    return await _build_profile(current_user.id, current_user.email, db)
+
+
+# ── Account deletion ─────────────────────────────────────────────────
+
+
+@router.delete("/me", status_code=status.HTTP_200_OK)
+async def delete_account(
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Permanently delete the authenticated user's account.
+
+    Cascades: refresh tokens, OAuth accounts, subscription, and all memory
+    rows are deleted via SQLAlchemy relationship cascades.  Stripe subscription
+    is cancelled if active.
+    """
+    # Cancel Stripe subscription if present.
+    try:
+        from app.billing.stripe_service import stripe_service  # noqa: PLC0415
+        await stripe_service.cancel_subscription(current_user.id, db)
+    except HTTPException:
+        pass  # No subscription — that's fine
+
+    # Delete all memory rows (core, associative, episodic, proactive).
+    try:
+        from app.models import (  # noqa: PLC0415
+            MemoryAssociative, MemoryCore, MemoryEpisodic, MemoryProactive,
+        )
+        for model in (MemoryCore, MemoryAssociative, MemoryEpisodic, MemoryProactive):
+            await db.execute(
+                model.__table__.delete().where(model.user_id == current_user.id)
+            )
+    except Exception:
+        pass  # Non-critical — cascade on User will handle most
+
+    # Delete the user row — cascades handle refresh_tokens, oauth_accounts, subscription.
+    result = await db.execute(select(User).where(User.id == current_user.id))
+    user = result.scalar_one()
+    await db.delete(user)
+    await db.commit()
+
+    return {"ok": True}

app/api/routes/billing.py

@@ -0,0 +1,132 @@
+"""Billing routes: Stripe checkout, webhook, subscription management.
+
+Business logic lives in ``app.billing.stripe_service.StripeService``.
+The route layer handles HTTP concerns (request parsing, response shaping)
+and delegates everything else to the service singleton.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from fastapi import APIRouter, Depends, Header, HTTPException, Request, status
+from pydantic import BaseModel
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.api.deps import get_current_user
+from app.billing.stripe_service import stripe_service
+from app.db import get_session
+from app.schemas import BillingTier, UserProfile
+
+router = APIRouter(prefix="/billing", tags=["billing"])
+
+
+# ── Request bodies ─────────────────────────────────────────────────────
+
+class _CheckoutRequest(BaseModel):
+    tier: BillingTier
+
+
+# ── Routes ─────────────────────────────────────────────────────────────
+
+@router.post("/checkout", response_model=dict)
+async def create_checkout(
+    body: _CheckoutRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> dict[str, str]:
+    """Create a Stripe checkout session for a tier upgrade.
+
+    Returns a stub URL when ``STRIPE_SECRET_KEY`` is not configured.
+    """
+    url = stripe_service.create_checkout_session(current_user.id, body.tier)
+    return {"checkout_url": url}
+
+
+@router.post("/webhook", response_model=dict)
+async def stripe_webhook(
+    request: Request,
+    stripe_signature: str = Header(default="", alias="Stripe-Signature"),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Handle Stripe webhook events.
+
+    No JWT auth — authenticated via Stripe signature verification instead.
+    Returns 200 immediately when Stripe is not configured (local dev).
+    """
+    payload = await request.body()
+    await stripe_service.handle_webhook(payload, stripe_signature, db)
+    return {"ok": True}
+
+
+@router.get("/subscription", response_model=dict)
+async def get_subscription(
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, Any]:
+    """Return the current subscription info for the authenticated user."""
+    sub = await stripe_service.get_subscription(current_user.id, db)
+    if sub is None:
+        return {
+            "tier": current_user.tier,
+            "status": "free",
+            "stripe_subscription_id": None,
+            "current_period_end": None,
+        }
+    return sub
+
+
+@router.delete("/subscription", response_model=dict, status_code=status.HTTP_200_OK)
+async def cancel_subscription(
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Cancel the active subscription."""
+    await stripe_service.cancel_subscription(current_user.id, db)
+    return {"ok": True}
+
+
+@router.get("/invoices", response_model=list[dict])
+async def list_invoices(
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> list[dict[str, Any]]:
+    """Return billing history (invoices) from Stripe.
+
+    Returns an empty list when Stripe is not configured.
+    """
+    invoices = await stripe_service.list_invoices(current_user.id, db)
+    return invoices
+
+
+# ── Quota check ────────────────────────────────────────────────────────
+
+from app.billing.quota import check_folder_quota, QuotaExceeded  # noqa: E402
+
+
+class QuotaCheckRequest(BaseModel):
+    feature: str
+    estimated_files: int
+
+
+@router.post("/quota/check")
+async def quota_check(
+    payload: QuotaCheckRequest,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict:
+    """Pre-flight folder quota check. 402 if tier limits would be exceeded."""
+    if payload.feature != "folder_index":
+        raise HTTPException(status_code=400, detail="Unknown feature")
+    try:
+        await check_folder_quota(
+            user_id=current_user.id,
+            tier=current_user.tier,
+            estimated_files=payload.estimated_files,
+            db=db,
+        )
+    except QuotaExceeded as exc:
+        raise HTTPException(
+            status_code=402,
+            detail={"reason": exc.reason, "message": str(exc)},
+        )
+    return {"ok": True}

app/api/routes/chat.py

@@ -0,0 +1,116 @@
+"""Chat routes: POST /chat (REST fallback) and POST /chat/embed (text → vector).
+
+WebSocket chat is handled by the unified device WS endpoint (/api/v1/ws/device).
+"""
+
+from __future__ import annotations
+
+import uuid
+from typing import Literal
+
+from fastapi import APIRouter, Depends, HTTPException
+from fastapi.responses import JSONResponse
+from pydantic import BaseModel
+
+from app.api.deps import get_current_user
+from app.core.brief_agent import run_home_brief, run_project_brief
+from app.core.deep_agent import run_home
+from app.core.llm import embed
+from app.core.memory_middleware import MemoryMiddleware
+from app.db import async_session
+from app.schemas import ChatRequest, UserProfile
+
+router = APIRouter(prefix="/chat", tags=["chat"])
+
+
+# ── Embed helpers ─────────────────────────────────────────────────────────
+
+
+class _EmbedRequest(BaseModel):
+    text: str
+
+
+class _EmbedResponse(BaseModel):
+    vector: list[float]
+
+
+# ── Endpoints ─────────────────────────────────────────────────────────────
+
+
+@router.post("")
+async def chat(
+    body: ChatRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> JSONResponse:
+    """REST fallback for home chat when websocket streaming is unavailable."""
+    response = await run_home(
+        user_id=current_user.id,
+        message=body.message,
+        context=body.context.model_dump(),
+    )
+    return JSONResponse(content={"response": response})
+
+
+class _BriefRequest(BaseModel):
+    mode: Literal["home", "project"]
+    project_id: str | None = None
+
+
+class _BriefResponse(BaseModel):
+    response: str
+
+
+@router.post("/brief", response_model=_BriefResponse)
+async def brief(
+    body: _BriefRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> _BriefResponse:
+    """REST fallback for brief when the device WebSocket is not ready."""
+    if body.mode == "project":
+        if not body.project_id:
+            raise HTTPException(status_code=422, detail="project_id required for project mode")
+        try:
+            uuid.UUID(body.project_id)
+        except ValueError:
+            raise HTTPException(status_code=422, detail="project_id must be a valid UUID")
+
+    request_id = str(uuid.uuid4())
+    async with async_session() as db:
+        memory = MemoryMiddleware(db)
+        memory_context = await memory.enrich_context(
+            current_user.id,
+            "",
+            trace_id=request_id,
+            session_id=request_id,
+        )
+
+    context: dict = {
+        "_debug": {"request_id": request_id, "user_id": current_user.id},
+        **memory_context,
+    }
+
+    chunks: list[str] = []
+    if body.mode == "project":
+        stream = run_project_brief(current_user.id, body.project_id, context)  # type: ignore[arg-type]
+    else:
+        stream = run_home_brief(current_user.id, context)
+
+    async for event_type, data in stream:
+        if event_type == "token" and data:
+            chunks.append(str(data))
+
+    return _BriefResponse(response="".join(chunks))
+
+
+@router.post("/embed", response_model=_EmbedResponse)
+async def embed_text(
+    body: _EmbedRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> _EmbedResponse:
+    """Generate a 1536-dim embedding vector for the given text.
+
+    Uses ``text-embedding-3-small`` via OpenAI.  Auth required (JWT).
+    Used by Electron (vectordb.ts) for local note search.
+    """
+    vector = await embed(body.text)
+    return _EmbedResponse(vector=vector)

app/api/routes/device_ws.py

@@ -0,0 +1,864 @@
+"""Device WebSocket endpoint.
+
+Persistent connection from Electron devices to the backend.
+
+  WS  /api/v1/ws/device?token=<jwt>
+
+Auth: JWT passed as ``?token=`` query parameter (Bearer header is not
+available during the WebSocket handshake).
+
+Protocol:
+  1. Client connects → JWT validated → connection accepted.
+  2. Client sends ``device_hello`` frame: ``{ type, device_id, scout_ids }``.
+  3. Backend registers the connection in ``DeviceConnectionManager``.
+  4. Session enters message dispatch loop + heartbeat.
+
+Incoming frame dispatch:
+  - ``tool_result``      → resolves a pending tool-call Future.
+  - ``journey_start``    → starts a guided setup journey session.
+  - ``journey_message``  → continues a journey conversation.
+  - ``pong``             → heartbeat acknowledgement (updates last-seen).
+  - unknown types        → logged, ignored.
+
+Outgoing heartbeat: ``{ "type": "ping" }`` every 30 s.
+
+On disconnect:
+  - Unregisters from DeviceConnectionManager.
+  - Marks all in-progress AgentRunLog rows for this user as ``error``
+    with message "device disconnected".
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+from uuid import uuid4
+
+from fastapi import APIRouter, WebSocket, WebSocketDisconnect
+from jose import JWTError, jwt
+from sqlalchemy import update
+
+from app.api.routes.scout_setup import handle_journey_message, handle_journey_start
+from app.config.settings import settings
+from app.scouts.engine import ScoutEngine
+from app.core.scout_runner import trigger_pending_runs
+from app.core.scout_session_buffer import session_buffer
+from app.core.brief_agent import run_home_brief, run_project_brief
+from app.core.deep_agent import run_contextual_stream, run_home_stream, run_task_brief_research_stream
+from app.core.output_formatter import extract_canvas_block
+from app.core.device_manager import device_manager
+from app.core.memory_middleware import MemoryMiddleware
+from app.core.output_formatter import StreamFormatter
+from app.core.ws_context import clear_client_executor, set_client_executor
+from app.db import async_session
+from app.models import ScoutRunLog
+from app.schemas import WsFrameType, WsStreamEnd
+from app.schemas.contextual import ContextualScope, render_scope_block
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/ws", tags=["device-ws"])
+
+# ── v7 folder index session state ─────────────────────────────────────
+# Keyed by sessionId; value: { user_id, project_id, processed, total, cancelled }
+_index_sessions: dict[str, dict] = {}
+
+_HEARTBEAT_INTERVAL = 30  # seconds
+_PONG_TIMEOUT = 10  # seconds — grace window after a ping
+
+
+@router.websocket("/device")
+async def device_ws(websocket: WebSocket) -> None:
+    """Persistent WebSocket endpoint for Electron device connections.
+
+    Authentication is via ``?token=<jwt>`` query parameter.
+    """
+    # ── 1. Authenticate before accepting ─────────────────────────────
+    token = websocket.query_params.get("token", "")
+    try:
+        payload = jwt.decode(
+            token, settings.JWT_SECRET, algorithms=[settings.JWT_ALGORITHM]
+        )
+        user_id: str | None = payload.get("sub")
+        if not user_id:
+            raise JWTError("missing sub")
+    except JWTError:
+        await websocket.close(code=1008)  # Policy Violation
+        return
+
+    await websocket.accept()
+
+    # ── 2. Await device_hello frame ───────────────────────────────────
+    try:
+        raw = await asyncio.wait_for(websocket.receive_text(), timeout=15.0)
+    except (asyncio.TimeoutError, WebSocketDisconnect):
+        await websocket.close(code=1008)
+        return
+
+    try:
+        hello = json.loads(raw)
+        if hello.get("type") != WsFrameType.device_hello:
+            raise ValueError("expected device_hello as first frame")
+        device_id: str = hello["device_id"]
+        scout_ids: list[str] = hello.get("scout_ids", [])
+    except (KeyError, ValueError, json.JSONDecodeError) as exc:
+        logger.warning("device_ws: invalid device_hello from user=%s: %s", user_id, exc)
+        await websocket.close(code=1008)
+        return
+
+    # ── 3. Register connection ────────────────────────────────────────
+    device_manager.register(user_id, device_id, websocket)
+    logger.info(
+        "device_ws: connected user=%s device=%s scouts=%s",
+        user_id,
+        device_id,
+        scout_ids,
+    )
+
+    # Trigger any overdue agent runs now that the device is connected.
+    asyncio.create_task(trigger_pending_runs(user_id, device_id, device_manager))
+
+    # Drain any queued scout proposals and deliver to the client (non-blocking).
+    async def _deliver_pending_safe() -> None:
+        import uuid as _uuid  # noqa: PLC0415
+        try:
+            await ScoutEngine().deliver_pending(_uuid.UUID(user_id), websocket)
+        except Exception:
+            logger.exception("scout deliver_pending failed for user %s", user_id)
+
+    asyncio.create_task(_deliver_pending_safe())
+
+    # ── 4. Concurrent message loop + heartbeat ────────────────────────
+    try:
+        await asyncio.gather(
+            _message_loop(websocket, user_id),
+            _heartbeat_loop(websocket),
+        )
+    except WebSocketDisconnect:
+        pass
+    except Exception as exc:
+        logger.warning("device_ws: unhandled exception user=%s: %s", user_id, exc)
+    finally:
+        device_manager.unregister(user_id)
+        logger.info("device_ws: disconnected user=%s device=%s", user_id, device_id)
+        await _mark_runs_disconnected(user_id)
+
+
+# ── Message dispatch loop ─────────────────────────────────────────────
+
+async def _message_loop(websocket: WebSocket, user_id: str) -> None:
+    """Receive frames from Electron and dispatch to the appropriate handler."""
+    async for raw in websocket.iter_text():
+        try:
+            frame: dict = json.loads(raw)
+        except json.JSONDecodeError:
+            logger.warning("device_ws: invalid JSON from user=%s", user_id)
+            continue
+
+        frame_type = frame.get("type")
+
+        if frame_type == WsFrameType.tool_result:
+            call_id = frame.get("id")
+            if call_id:
+                device_manager.resolve_pending_call(user_id, call_id, frame)
+            else:
+                logger.warning(
+                    "device_ws: tool_result missing id from user=%s", user_id
+                )
+
+        elif frame_type == WsFrameType.home_request:
+            asyncio.create_task(
+                _handle_home_request(websocket, user_id, frame)
+            )
+
+        elif frame_type == WsFrameType.brief_request:
+            asyncio.create_task(
+                _handle_brief_request(websocket, user_id, frame)
+            )
+
+        elif frame_type == WsFrameType.task_brief_request:
+            asyncio.create_task(
+                _handle_task_brief_request(websocket, user_id, frame)
+            )
+
+        elif frame_type == WsFrameType.journey_start:
+            asyncio.create_task(
+                _handle_journey_start(websocket, user_id, frame)
+            )
+
+        elif frame_type == WsFrameType.journey_message:
+            asyncio.create_task(
+                _handle_journey_message(websocket, user_id, frame)
+            )
+
+        elif frame_type == WsFrameType.index_session_start:
+            asyncio.create_task(
+                _handle_index_session_start(websocket, user_id, frame)
+            )
+
+        elif frame_type == WsFrameType.index_file_batch:
+            asyncio.create_task(
+                _handle_index_file_batch(websocket, user_id, frame)
+            )
+
+        elif frame_type == WsFrameType.index_session_cancel:
+            await _handle_index_session_cancel(websocket, frame)
+
+        elif frame_type == WsFrameType.contextual_request:
+            asyncio.create_task(
+                _handle_contextual_request(websocket, user_id, frame)
+            )
+
+        elif frame_type == WsFrameType.contextual_scope_update:
+            asyncio.create_task(
+                _handle_contextual_scope_update(websocket, user_id, frame)
+            )
+
+        elif frame_type == "scout_proposal_ack":
+            proposal_id = frame.get("proposal_id")
+            if proposal_id:
+                try:
+                    await ScoutEngine().ack_proposal(proposal_id)
+                except Exception:
+                    logger.exception("scout ack_proposal failed for %s", proposal_id)
+
+        elif frame_type == "pong":
+            # Heartbeat ack — nothing to do, connection is alive.
+            pass
+
+        else:
+            logger.debug(
+                "device_ws: unknown frame type %r from user=%s", frame_type, user_id
+            )
+
+
+# ── v3 Chat Handlers ──────────────────────────────────────────────────
+
+async def _make_ws_executor(websocket: WebSocket, user_id: str):
+    """Return a callback that sends tool_call frames and awaits tool_result."""
+    async def _executor(payload: dict) -> dict:
+        payload["type"] = WsFrameType.tool_call
+        await websocket.send_text(json.dumps(payload))
+        future = device_manager.create_pending_call(user_id, payload["id"])
+        return await future
+    return _executor
+
+
+async def _handle_home_request(
+    websocket: WebSocket,
+    user_id: str,
+    frame: dict,
+) -> None:
+    """Handle a home_request frame — streams HomeFormatter output back on the socket."""
+    request_id = frame.get("request_id") or str(uuid4())
+    message: str = frame.get("message", "")
+    session_id: str = frame.get("session_id") or str(uuid4())
+    project_id: str | None = frame.get("project_id") or frame.get("projectId") or None
+    logger.info(
+        "device_ws: home_request_start user=%s req=%s session=%s project=%s msg=%s",
+        user_id,
+        request_id,
+        session_id,
+        project_id,
+        message[:200],
+    )
+
+    # ── Memory: enrich context before LLM call ────────────────────────
+    async with async_session() as db:
+        memory = MemoryMiddleware(db)
+        memory_context = await memory.enrich_context(
+            user_id,
+            message,
+            trace_id=request_id,
+            session_id=session_id,
+        )
+
+    context: dict = {
+        "conversation_history": frame.get("conversation_history", []),
+        "_debug": {"request_id": request_id, "session_id": session_id, "user_id": user_id},
+        "format_prefs": frame.get("format_prefs"),
+        **memory_context,
+    }
+
+    executor = await _make_ws_executor(websocket, user_id)
+    set_client_executor(executor)
+    response_chunks: list[str] = []
+    try:
+        event_stream = run_home_stream(user_id, message, context, project_id=project_id)
+        formatter = StreamFormatter(request_id=request_id)
+        async for ws_frame in formatter.format(event_stream):
+            await websocket.send_text(ws_frame.model_dump_json())
+            # Collect text chunks to build the full response for episode storage
+            if ws_frame.type == "stream_text":  # type: ignore[union-attr]
+                response_chunks.append(ws_frame.chunk)  # type: ignore[union-attr]
+    except Exception as exc:
+        logger.error(
+            "device_ws: home_request failed user=%s req=%s: %s",
+            user_id, request_id, exc,
+        )
+    finally:
+        clear_client_executor()
+
+    # ── Memory: store episode after response ──────────────────────────
+    async with async_session() as db:
+        memory = MemoryMiddleware(db)
+        await memory.store_episode(
+            user_id, session_id, message, "".join(response_chunks), trace_id=request_id
+        )
+    logger.info(
+        "device_ws: home_request_end user=%s req=%s session=%s response_chars=%d",
+        user_id,
+        request_id,
+        session_id,
+        len("".join(response_chunks)),
+    )
+
+
+# ── v8 Contextual Sidebar Handlers ───────────────────────────────────
+
+
+def get_session_buffer(user_id: str, session_id: str, channel: str = "contextual"):
+    """Return a session-scoped buffer proxy for the given user+session.
+
+    Returns a _ContextualBufferProxy that exposes append_system_message().
+    Defined at module level so tests can monkeypatch it.
+    The channel kwarg is accepted for forward-compatibility.
+    """
+    from app.core.scout_session_buffer import ContextualBufferProxy  # noqa: PLC0415
+    return ContextualBufferProxy(session_buffer, user_id, session_id)
+
+
+async def _handle_contextual_request(
+    websocket: WebSocket,
+    user_id: str,
+    frame: dict,
+) -> None:
+    """Handle a contextual_request frame — runs the contextual agent and streams frames."""
+    request_id = frame.get("request_id") or str(uuid4())
+    message: str = frame.get("message", "")
+    session_id: str = frame.get("session_id") or str(uuid4())
+    scope_payload: dict = frame.get("scope", {})
+    logger.info(
+        "device_ws: contextual_request_start user=%s req=%s session=%s msg=%s",
+        user_id,
+        request_id,
+        session_id,
+        message[:200],
+    )
+
+    scope = ContextualScope.model_validate(scope_payload)
+
+    # Enrich context with memory before the LLM call.
+    async with async_session() as db:
+        memory = MemoryMiddleware(db)
+        memory_context = await memory.enrich_context(
+            user_id,
+            message,
+            trace_id=request_id,
+            session_id=session_id,
+        )
+
+    context: dict = {
+        "conversation_history": frame.get("conversation_history", []),
+        "format_prefs": frame.get("format_prefs"),
+        "_debug": {"request_id": request_id, "session_id": session_id, "user_id": user_id},
+        **memory_context,
+    }
+
+    executor = await _make_ws_executor(websocket, user_id)
+    set_client_executor(executor)
+    response_chunks: list[str] = []
+    try:
+        event_stream = run_contextual_stream(
+            user_id=user_id,
+            message=message,
+            context=context,
+            scope=scope,
+        )
+        formatter = StreamFormatter(request_id=request_id)
+        async for ws_frame in formatter.format(event_stream):
+            await websocket.send_text(ws_frame.model_dump_json())
+            if ws_frame.type == "stream_text":  # type: ignore[union-attr]
+                response_chunks.append(ws_frame.chunk)  # type: ignore[union-attr]
+    except Exception as exc:
+        logger.error(
+            "device_ws: contextual_request failed user=%s req=%s: %s",
+            user_id, request_id, exc,
+        )
+    finally:
+        clear_client_executor()
+
+    # Store episode so the contextual agent can recall prior turns.
+    async with async_session() as db:
+        memory = MemoryMiddleware(db)
+        await memory.store_episode(
+            user_id, session_id, message, "".join(response_chunks), trace_id=request_id
+        )
+    logger.info(
+        "device_ws: contextual_request_end user=%s req=%s session=%s response_chars=%d",
+        user_id,
+        request_id,
+        session_id,
+        len("".join(response_chunks)),
+    )
+
+
+async def _handle_contextual_scope_update(
+    websocket: WebSocket,
+    user_id: str,
+    frame: dict,
+) -> None:
+    """Handle a contextual_scope_update frame.
+
+    Injects a synthetic system message into the session buffer so the next
+    agent turn knows the user navigated.  No LLM call is made.
+    """
+    session_id: str = frame.get("session_id") or str(uuid4())
+    scope = ContextualScope.model_validate(frame.get("scope", {}))
+    block = render_scope_block(scope)
+    buf = get_session_buffer(user_id, session_id, channel="contextual")
+    buf.append_system_message(
+        f"User navigated to a new view. {block} Treat this as the new active context."
+    )
+    await websocket.send_text(json.dumps({
+        "type": WsFrameType.contextual_scope_ack,
+        "session_id": session_id,
+    }))
+    logger.info(
+        "device_ws: contextual_scope_update user=%s session=%s page=%s",
+        user_id, session_id, scope.page,
+    )
+
+
+async def _handle_brief_request(
+    websocket: WebSocket,
+    user_id: str,
+    frame: dict,
+) -> None:
+    """Handle a brief_request frame — streams plain-text brief back on the socket.
+
+    No episode storage — briefs are not conversations.
+    """
+    import uuid as _uuid
+
+    request_id = frame.get("request_id") or str(uuid4())
+    session_id = frame.get("session_id") or str(uuid4())
+    mode: str = frame.get("mode", "home")
+    project_id: str | None = frame.get("project_id")
+
+    logger.info(
+        "device_ws: brief_request_start user=%s req=%s mode=%s project_id=%s",
+        user_id, request_id, mode, project_id,
+    )
+
+    # Validate project_id for project mode before touching LLM.
+    if mode == "project":
+        try:
+            if not project_id:
+                raise ValueError("project_id required for project mode")
+            _uuid.UUID(project_id)
+        except (ValueError, AttributeError) as exc:
+            logger.warning(
+                "device_ws: brief_request invalid project_id user=%s req=%s: %s",
+                user_id, request_id, exc,
+            )
+            await websocket.send_text(
+                WsStreamEnd(request_id=request_id, error=str(exc)).model_dump_json()
+            )
+            return
+
+    # Enrich context with memory (no user message — use empty string as probe).
+    async with async_session() as db:
+        memory = MemoryMiddleware(db)
+        memory_context = await memory.enrich_context(
+            user_id,
+            "",
+            trace_id=request_id,
+            session_id=session_id,
+        )
+
+    context: dict = {
+        "_debug": {"request_id": request_id, "session_id": session_id, "user_id": user_id},
+        "format_prefs": frame.get("format_prefs"),
+        **memory_context,
+    }
+
+    executor = await _make_ws_executor(websocket, user_id)
+    set_client_executor(executor)
+    try:
+        if mode == "project":
+            event_stream = run_project_brief(user_id, project_id, context)  # type: ignore[arg-type]
+        else:
+            event_stream = run_home_brief(user_id, context)
+
+        formatter = StreamFormatter(request_id=request_id)
+        async for ws_frame in formatter.format(event_stream):
+            await websocket.send_text(ws_frame.model_dump_json())
+    except Exception as exc:
+        logger.error(
+            "device_ws: brief_request failed user=%s req=%s: %s",
+            user_id, request_id, exc,
+        )
+        await websocket.send_text(
+            WsStreamEnd(request_id=request_id, error=str(exc)).model_dump_json()
+        )
+    finally:
+        clear_client_executor()
+
+    logger.info(
+        "device_ws: brief_request_end user=%s req=%s mode=%s",
+        user_id, request_id, mode,
+    )
+
+
+# ── v6 Task Brief Handler ────────────────────────────────────────────
+
+
+async def _handle_task_brief_request(
+    websocket: WebSocket,
+    user_id: str,
+    frame: dict,
+) -> None:
+    """Handle a task_brief_request frame — Stage-1 executive assistant deep research.
+
+    Streams the briefing markdown back to the client.
+    On stream_end, emits a ``canvas_draft`` mutation if the agent produced one.
+    """
+    request_id = frame.get("request_id") or str(uuid4())
+    session_id = frame.get("session_id") or str(uuid4())
+    task_id: str = frame.get("task_id") or frame.get("taskId") or ""
+    project_id: str | None = frame.get("project_id") or frame.get("projectId") or None
+
+    logger.info(
+        "device_ws: task_brief_request_start user=%s req=%s task=%s project=%s [cache_miss]",
+        user_id, request_id, task_id, project_id,
+    )
+
+    if not task_id:
+        await websocket.send_text(
+            WsStreamEnd(request_id=request_id, error="task_id is required").model_dump_json()
+        )
+        return
+
+    async with async_session() as db:
+        memory = MemoryMiddleware(db)
+        memory_context = await memory.enrich_context(
+            user_id,
+            f"task brief: {task_id}",
+            trace_id=request_id,
+            session_id=session_id,
+        )
+
+    context: dict = {
+        "_debug": {"request_id": request_id, "session_id": session_id, "user_id": user_id},
+        "format_prefs": frame.get("format_prefs"),
+        **memory_context,
+    }
+
+    executor = await _make_ws_executor(websocket, user_id)
+    set_client_executor(executor)
+    response_chunks: list[str] = []
+
+    try:
+        event_stream = run_task_brief_research_stream(user_id, task_id, context, project_id=project_id)
+        formatter = StreamFormatter(request_id=request_id)
+        async for ws_frame in formatter.format(event_stream):
+            if ws_frame.type == "stream_text":  # type: ignore[union-attr]
+                response_chunks.append(ws_frame.chunk)  # type: ignore[union-attr]
+                await websocket.send_text(ws_frame.model_dump_json())
+            elif ws_frame.type == "stream_start":
+                await websocket.send_text(ws_frame.model_dump_json())
+            # stream_end is emitted below with mutations — skip formatter's version
+    except Exception as exc:
+        logger.error(
+            "device_ws: task_brief_request failed user=%s req=%s task=%s: %s",
+            user_id, request_id, task_id, exc,
+        )
+        await websocket.send_text(
+            WsStreamEnd(request_id=request_id, error=str(exc)).model_dump_json()
+        )
+        return
+    finally:
+        clear_client_executor()
+
+    # Extract canvas block then emit stream_end with optional mutations.
+    full_response = "".join(response_chunks)
+    _visible, canvas_content, canvas_kind = extract_canvas_block(full_response)
+
+    mutations: list[dict] = []
+    if canvas_content:
+        mutations.append({
+            "type": "canvas_draft",
+            "content": canvas_content,
+            "kind": canvas_kind,
+        })
+
+    await websocket.send_text(
+        WsStreamEnd(request_id=request_id, mutations=mutations or None).model_dump_json()
+    )
+
+    logger.info(
+        "device_ws: task_brief_request_end user=%s req=%s task=%s response_chars=%d canvas=%s",
+        user_id, request_id, task_id, len(full_response), canvas_kind or "none",
+    )
+
+
+# ── v4 Journey Handlers ─────────────────────────────────────────────
+
+
+async def _handle_journey_start(
+    websocket: WebSocket,
+    user_id: str,
+    frame: dict,
+) -> None:
+    """Handle a journey_start frame — explores directory and sends first question."""
+    executor = await _make_ws_executor(websocket, user_id)
+    set_client_executor(executor)
+    try:
+        reply = await handle_journey_start(user_id, frame)
+        await websocket.send_text(json.dumps(reply))
+    except Exception as exc:
+        logger.error(
+            "device_ws: journey_start failed user=%s: %s", user_id, exc
+        )
+        await websocket.send_text(json.dumps({
+            "type": "journey_reply",
+            "session_id": frame.get("session_id", ""),
+            "message": f"Failed to start journey: {exc}",
+            "done": True,
+            "prompt_template": None,
+        }))
+    finally:
+        clear_client_executor()
+
+
+async def _handle_journey_message(
+    websocket: WebSocket,
+    user_id: str,
+    frame: dict,
+) -> None:
+    """Handle a journey_message frame — continues the journey conversation."""
+    executor = await _make_ws_executor(websocket, user_id)
+    set_client_executor(executor)
+    try:
+        reply = await handle_journey_message(user_id, frame)
+        await websocket.send_text(json.dumps(reply))
+    except Exception as exc:
+        session_id = frame.get("session_id", "")
+        logger.error(
+            "device_ws: journey_message failed user=%s session=%s: %s",
+            user_id, session_id, exc,
+        )
+        await websocket.send_text(json.dumps({
+            "type": "journey_reply",
+            "session_id": session_id,
+            "message": f"Journey error: {exc}",
+            "done": True,
+            "prompt_template": None,
+        }))
+    finally:
+        clear_client_executor()
+
+
+# ── v7 Folder Index Handlers ──────────────────────────────────────────
+
+
+async def _handle_index_session_start(
+    websocket: WebSocket,
+    user_id: str,
+    frame: dict,
+) -> None:
+    """Register a new folder index session.  No response sent — client is declaring intent."""
+    session_id: str = frame.get("sessionId") or frame.get("session_id") or ""
+    project_id: str | None = frame.get("projectId") or frame.get("project_id")
+    total: int = int(frame.get("totalFiles") or frame.get("total_files") or 0)
+
+    if not session_id:
+        logger.warning("device_ws: index_session_start missing sessionId user=%s", user_id)
+        return
+
+    _index_sessions[session_id] = {
+        "user_id": user_id,
+        "project_id": project_id,
+        "processed": 0,
+        "total": total,
+        "cancelled": False,
+    }
+    logger.info(
+        "device_ws: index_session_start user=%s session=%s project=%s total=%d",
+        user_id, session_id, project_id, total,
+    )
+
+
+async def _handle_index_session_cancel(
+    websocket: WebSocket,
+    frame: dict,
+) -> None:
+    """Mark a session as cancelled and emit index_session_done(cancelled)."""
+    session_id: str = frame.get("sessionId") or frame.get("session_id") or ""
+    session = _index_sessions.get(session_id)
+    if session:
+        session["cancelled"] = True
+
+    await websocket.send_text(json.dumps({
+        "type": WsFrameType.index_session_done,
+        "sessionId": session_id,
+        "status": "cancelled",
+    }))
+    _index_sessions.pop(session_id, None)
+    logger.info("device_ws: index_session_cancel session=%s", session_id)
+
+
+async def _handle_index_file_batch(
+    websocket: WebSocket,
+    user_id: str,
+    frame: dict,
+) -> None:
+    """Process a batch of files for an index session, streaming results back."""
+    # Lazy imports to avoid heavy load at module startup.
+    from app.core.folder_indexer import (  # noqa: PLC0415
+        summarize_image,
+        summarize_pdf,
+        summarize_docx,
+        summarize_text,
+    )
+    from app.billing.tier_manager import tier_manager  # noqa: PLC0415
+    from app.billing.quota import add_token_usage  # noqa: PLC0415
+
+    session_id: str = frame.get("sessionId") or frame.get("session_id") or ""
+    files: list[dict] = frame.get("files", [])
+
+    session = _index_sessions.get(session_id)
+    if not session or session.get("cancelled"):
+        return
+
+    async with async_session() as db:
+        tier = await tier_manager.get_tier(user_id, db)
+        raw_cap = tier_manager.get_feature_value(tier, "folder_monthly_tokens")
+        cap: int | None = None if raw_cap == -1 else raw_cap
+
+        for file_info in files:
+            if session.get("cancelled"):
+                return
+
+            # Electron's toSnakeCase converts payload keys, so accept both forms.
+            rel_path: str = file_info.get("relPath") or file_info.get("rel_path") or ""
+            kind: str = file_info.get("kind") or "text"
+            content: str = file_info.get("content") or ""
+            ext: str = file_info.get("ext") or ""
+            mime: str = file_info.get("mime") or "application/octet-stream"
+            name: str = rel_path.split("/")[-1] or rel_path
+
+            try:
+                if kind == "image":
+                    res = await summarize_image(image_b64=content, mime=mime)
+                elif kind == "pdf":
+                    res = await summarize_pdf(pdf_b64=content, name=name)
+                elif kind == "docx":
+                    res = await summarize_docx(docx_b64=content, name=name)
+                else:
+                    res = await summarize_text(content=content, ext=ext, name=name)
+            except Exception as exc:
+                logger.warning(
+                    "device_ws: index_file_batch summarize failed session=%s path=%s: %s",
+                    session_id, rel_path, exc,
+                )
+                await websocket.send_text(json.dumps({
+                    "type": WsFrameType.index_file_result,
+                    "sessionId": session_id,
+                    "relPath": rel_path,
+                    "summary": None,
+                    "tokensUsed": 0,
+                    "error": str(exc),
+                }))
+                session["processed"] += 1
+                continue
+
+            # Account for token usage and check cap.
+            usage = await add_token_usage(
+                user_id=user_id,
+                feature="folder_index",
+                tokens=res.tokens_used,
+                db=db,
+                cap=cap,
+            )
+
+            await websocket.send_text(json.dumps({
+                "type": WsFrameType.index_file_result,
+                "sessionId": session_id,
+                "relPath": rel_path,
+                "summary": res.summary,
+                "tokensUsed": res.tokens_used,
+            }))
+            session["processed"] += 1
+
+            if usage.exhausted:
+                await websocket.send_text(json.dumps({
+                    "type": WsFrameType.index_session_done,
+                    "sessionId": session_id,
+                    "status": "quota_exceeded",
+                }))
+                _index_sessions.pop(session_id, None)
+                logger.info(
+                    "device_ws: index_session quota_exceeded user=%s session=%s",
+                    user_id, session_id,
+                )
+                return
+
+        # After processing the batch, emit progress.
+        processed = session["processed"]
+        total = session["total"]
+        await websocket.send_text(json.dumps({
+            "type": WsFrameType.index_session_progress,
+            "sessionId": session_id,
+            "processed": processed,
+            "total": total,
+        }))
+
+        if processed >= total:
+            await websocket.send_text(json.dumps({
+                "type": WsFrameType.index_session_done,
+                "sessionId": session_id,
+                "status": "completed",
+            }))
+            _index_sessions.pop(session_id, None)
+            logger.info(
+                "device_ws: index_session_done completed user=%s session=%s processed=%d",
+                user_id, session_id, processed,
+            )
+
+
+# ── Heartbeat ─────────────────────────────────────────────────────────
+
+async def _heartbeat_loop(websocket: WebSocket) -> None:
+    """Send a ping frame every 30 s to keep the connection alive."""
+    while True:
+        await asyncio.sleep(_HEARTBEAT_INTERVAL)
+        await websocket.send_text(json.dumps({"type": "ping"}))
+
+
+# ── Disconnect cleanup ────────────────────────────────────────────────
+
+async def _mark_runs_disconnected(user_id: str) -> None:
+    """Mark all in-progress ScoutRunLog rows as 'error' for this user."""
+    try:
+        async with async_session() as db:
+            await db.execute(
+                update(ScoutRunLog)
+                .where(
+                    ScoutRunLog.user_id == user_id,
+                    ScoutRunLog.status == "running",
+                )
+                .values(
+                    status="error",
+                    errors=["device disconnected"],
+                )
+            )
+            await db.commit()
+    except Exception as exc:
+        logger.error(
+            "device_ws: failed to mark runs as disconnected for user=%s: %s",
+            user_id,
+            exc,
+        )

app/api/routes/memory.py

@@ -0,0 +1,225 @@
+"""Memory management routes — view/edit/delete user memory tiers.
+
+All routes require authentication. Data is always user-scoped.
+"""
+
+from __future__ import annotations
+
+import logging
+from datetime import datetime, timezone
+from typing import Annotated
+
+from fastapi import APIRouter, Depends, Header, HTTPException, status
+from pydantic import BaseModel, Field
+from sqlalchemy import delete, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.api.deps import get_current_user
+from app.core.memory_middleware import MemoryMiddleware
+from app.db import get_session
+from app.models import (
+    ExtractionQueue,
+    MemoryAssociative,
+    MemoryCore,
+    MemoryEpisodic,
+    MemoryProactive,
+    MemoryRelation,
+)
+from app.schemas import UserProfile
+
+router = APIRouter(prefix="/memory", tags=["memory"])
+
+logger = logging.getLogger(__name__)
+
+_ALLOWED_PREDICATES = {
+    "works_at",
+    "reports_to",
+    "stakeholder_of",
+    "last_contacted_on",
+    "owes_followup",
+    "manages",
+    "collaborates_with",
+    "owns",
+    "member_of",
+    "custom",
+}
+
+
+# ── Response schemas ─────────────────────────────────────────────────────────
+
+class RelationOut(BaseModel):
+    id: str
+    subject_label: str
+    subject_type: str
+    predicate: str
+    object_label: str
+    object_type: str
+    confidence: float
+    last_confirmed_at: int | None = None  # epoch ms
+
+
+class RelationPatch(BaseModel):
+    subject_label: str | None = None
+    object_label: str | None = None
+    predicate: str | None = None
+    confidence: float | None = Field(None, ge=0.0, le=1.0)
+
+
+class CoreAddBody(BaseModel):
+    key: str = Field(..., min_length=1, max_length=255)
+    value: str = Field(..., min_length=1)
+
+
+# ── Helpers ──────────────────────────────────────────────────────────────────
+
+def _relation_to_out(row: MemoryRelation) -> RelationOut:
+    last_ms: int | None = None
+    if row.last_confirmed_at is not None:
+        last_ms = int(row.last_confirmed_at.timestamp() * 1000)
+    return RelationOut(
+        id=row.id,
+        subject_label=row.subject_label,
+        subject_type=row.subject_type,
+        predicate=row.predicate,
+        object_label=row.object_label,
+        object_type=row.object_type,
+        confidence=row.confidence,
+        last_confirmed_at=last_ms,
+    )
+
+
+# ── Routes ───────────────────────────────────────────────────────────────────
+
+@router.get("/core", response_model=dict[str, str])
+async def get_core_memory(
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, str]:
+    """Return all core memory k/v pairs (plaintext) for the current user."""
+    mw = MemoryMiddleware(db)
+    blocks = await mw.list_core_blocks(current_user.id)
+    return {b["label"]: b["value"] for b in blocks}
+
+
+@router.delete("/core/{key}", status_code=status.HTTP_204_NO_CONTENT)
+async def delete_core_key(
+    key: str,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> None:
+    """Delete a single core memory key (GDPR Art. 17)."""
+    mw = MemoryMiddleware(db)
+    deleted = await mw.delete_core(current_user.id, key)
+    if not deleted:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Key not found")
+
+
+@router.post("/core", status_code=status.HTTP_201_CREATED, response_model=dict[str, str])
+async def add_core_key(
+    body: CoreAddBody,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, str]:
+    """Add or overwrite a core memory key/value pair."""
+    mw = MemoryMiddleware(db)
+    await mw.update_core(current_user.id, body.key, body.value)
+    return {body.key: body.value}
+
+
+@router.get("/relational", response_model=list[RelationOut])
+async def get_relational_memory(
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> list[RelationOut]:
+    """Return all relational memory rows for the current user."""
+    mw = MemoryMiddleware(db)
+    rows = await mw.query_relations(current_user.id, limit=200)
+    return [_relation_to_out(r) for r in rows]
+
+
+@router.patch("/relational/{relation_id}", response_model=RelationOut)
+async def patch_relation(
+    relation_id: str,
+    body: RelationPatch,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> RelationOut:
+    """Edit a relation row's labels, predicate, or confidence."""
+    if body.predicate is not None and body.predicate not in _ALLOWED_PREDICATES:
+        raise HTTPException(
+            status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+            detail=f"predicate must be one of: {sorted(_ALLOWED_PREDICATES)}",
+        )
+
+    result = await db.execute(
+        select(MemoryRelation).where(
+            MemoryRelation.id == relation_id,
+            MemoryRelation.user_id == current_user.id,
+        )
+    )
+    row = result.scalar_one_or_none()
+    if row is None:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Relation not found")
+
+    if body.subject_label is not None:
+        row.subject_label = body.subject_label
+    if body.object_label is not None:
+        row.object_label = body.object_label
+    if body.predicate is not None:
+        row.predicate = body.predicate
+    if body.confidence is not None:
+        row.confidence = body.confidence
+        row.last_confirmed_at = datetime.now(timezone.utc)
+
+    await db.commit()
+    await db.refresh(row)
+    logger.info("memory: patch_relation user=%s relation=%s", current_user.id, relation_id)
+    return _relation_to_out(row)
+
+
+@router.delete("/relational/{relation_id}", status_code=status.HTTP_204_NO_CONTENT)
+async def delete_relation(
+    relation_id: str,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> None:
+    """Hard-delete a relation row (GDPR Art. 17)."""
+    result = await db.execute(
+        select(MemoryRelation).where(
+            MemoryRelation.id == relation_id,
+            MemoryRelation.user_id == current_user.id,
+        )
+    )
+    row = result.scalar_one_or_none()
+    if row is None:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Relation not found")
+    await db.delete(row)
+    await db.commit()
+    logger.info("memory: delete_relation user=%s relation=%s", current_user.id, relation_id)
+
+
+@router.post("/forget-all", status_code=status.HTTP_204_NO_CONTENT)
+async def forget_all(
+    x_confirm: Annotated[str | None, Header(alias="X-Confirm")] = None,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> None:
+    """Wipe all memory tiers for the current user (GDPR Art. 17).
+
+    Requires ``X-Confirm: true`` header. Does NOT delete the user account.
+    """
+    if x_confirm != "true":
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail="Missing or invalid X-Confirm header. Send X-Confirm: true to confirm.",
+        )
+
+    uid = current_user.id
+    await db.execute(delete(MemoryCore).where(MemoryCore.user_id == uid))
+    await db.execute(delete(MemoryAssociative).where(MemoryAssociative.user_id == uid))
+    await db.execute(delete(MemoryEpisodic).where(MemoryEpisodic.user_id == uid))
+    await db.execute(delete(MemoryProactive).where(MemoryProactive.user_id == uid))
+    await db.execute(delete(MemoryRelation).where(MemoryRelation.user_id == uid))
+    await db.execute(delete(ExtractionQueue).where(ExtractionQueue.user_id == uid))
+    await db.commit()
+    logger.warning("memory: forget_all GDPR wipe user=%s", uid)

app/api/routes/scout_setup.py

@@ -0,0 +1,513 @@
+"""Chatbot Journey — WS-based guided conversation to build an ScoutConfig.
+
+The journey is driven entirely through WebSocket frames (no REST endpoints).
+The device WS handler dispatches ``journey_start`` and ``journey_message``
+frames to the functions exported here.
+
+Journey flow:
+  1. FE sends ``journey_start`` frame with basic agent info (directory,
+     data_types, schedule).
+  2. Server creates an in-memory session, sets up a WS executor so the
+     setup LLM can use file-system tools, does a first directory scrape,
+     and sends back a ``journey_reply`` with the first question.
+  3. FE sends ``journey_message`` frames for each user reply.
+  4. Server appends the user message, calls the LLM (which may read files
+     via tools), and sends back a ``journey_reply``.
+  5. After 3-5 turns the LLM wraps up by emitting an ``ScoutConfig`` JSON
+     block delimited by ``AGENT_CONFIG_START`` / ``AGENT_CONFIG_END``.
+  6. Server parses and validates the JSON with Pydantic, sends
+     ``journey_reply`` with ``done=True`` and the serialised config.
+     FE stores it locally.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import time
+import uuid
+from dataclasses import dataclass, field
+from typing import Any
+
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
+
+from app.agents.filesystem_agent import make_directory_tools
+from app.core.langfuse_client import compile_prompt, extract_usage, get_langfuse, get_prompt_or_fallback, langfuse_context
+from app.core.llm import get_agent_llm, model_for_agent
+from app.schemas import ScoutConfig
+
+logger = logging.getLogger(__name__)
+
+# ── Session TTL ───────────────────────────────────────────────────────────
+
+_SESSION_TTL_SECONDS: int = 1800  # 30 minutes
+
+# Sentinel strings used to delimit the LLM-produced ScoutConfig JSON.
+_CONFIG_START = "AGENT_CONFIG_START"
+_CONFIG_END = "AGENT_CONFIG_END"
+
+# Minimum turns before we consider nudging the LLM to wrap up.
+_MIN_TURNS_BEFORE_NUDGE: int = 3
+# Hard cap to avoid infinite loops (safety net, not the primary stopping criterion).
+_MAX_TURNS: int = 15
+# Max tool-calling steps per LLM invocation.
+_MAX_TOOL_STEPS: int = 6
+
+# ── In-memory session store ───────────────────────────────────────────────
+
+
+@dataclass
+class JourneySession:
+    session_id: str
+    user_id: str
+    agent_type: str  # "local" | "cloud"
+    directory: str
+    data_types: list[str]
+    history: list[dict[str, Any]] = field(default_factory=list)
+    system_prompt: str = ""
+    langfuse_prompt: Any = None
+    created_at: float = field(default_factory=time.monotonic)
+
+    def is_expired(self) -> bool:
+        return (time.monotonic() - self.created_at) > _SESSION_TTL_SECONDS
+
+
+# session_id → session
+_sessions: dict[str, JourneySession] = {}
+
+
+def get_journey_session(session_id: str, user_id: str) -> JourneySession | None:
+    """Retrieve session; return None on missing, expired, or wrong owner."""
+    s = _sessions.get(session_id)
+    if s is None or s.is_expired():
+        _sessions.pop(session_id, None)
+        return None
+    if s.user_id != user_id:
+        return None
+    return s
+
+
+# ── System prompt ─────────────────────────────────────────────────────────
+
+_JOURNEY_SYSTEM_PROMPT = """\
+You are a friendly assistant helping a freelancer configure a data-extraction agent.
+Your job is to understand what files the user has in their directory and produce a
+structured ScoutConfig JSON that the extraction agent will use as its instruction set.
+
+You have access to file-system tools to explore the user's directory:
+- list_directory: see folder structure and file names
+- read_file_content: peek at a file's content
+- get_file_metadata: check file size, extension, dates
+
+The user's configured directory is: {directory}
+Target data types: {data_types}
+
+## Your process
+
+### Step 1 — Explore the directory
+Use list_directory and read_file_content to understand what types of files are present
+(HTML emails, plain-text documents, CSVs, etc.).
+
+### Step 2 — Identify content types
+For each distinct file type found, decide:
+- A short id (e.g. "email_html", "plain_text", "csv")
+- Which preprocessing handler to use: "email_html" for HTML emails, "generic" for everything else
+- A human-readable label and optional detection_hint
+
+### Step 3 — Ask focused questions (one at a time)
+Cover these topics based on what you discovered:
+1. How to map content to entity types (task / note / timeline entry)
+2. Field mapping rules (e.g. email Subject → task title, filename → note title)
+3. Priority or status rules (e.g. "urgent" in subject → high priority)
+4. Date extraction (e.g. "by Friday" → dueDate)
+5. Exclusion rules (e.g. skip newsletters, skip files with no project match)
+
+### Step 4 — Produce the ScoutConfig JSON
+Once you are ≥ 90% confident, output the final config between these exact markers
+(each on its own line):
+
+{config_start}
+{{
+  "content_types": [
+    {{
+      "id": "email_html",
+      "label": "Email HTML",
+      "detection_hint": "HTML file with From/To/Subject headers",
+      "preprocessing": "email_html",
+      "extraction_prompt": "Detailed extraction instructions for this content type..."
+    }}
+  ],
+  "global_rules": [
+    "If the file cannot be matched to any project, do not create any entity."
+  ],
+  "data_types": {data_types_json}
+}}
+{config_end}
+
+## Rules for the extraction_prompt field
+- Describe when to create a task vs note vs timeline entry (be specific and concrete)
+- Include field mapping rules based on what you found in the directory
+- Include priority/status/date rules if applicable
+- Do NOT include projectId logic — the runner handles project assignment automatically
+- Do NOT mention isAiSuggested — the runner always sets it to 1
+
+## Constraints
+- Never ask about projects, projectId, or how to link records to projects
+- Never include projectId or project creation logic in the generated config
+- Keep asking questions until ≥ 90% confident, then output the JSON immediately
+
+{existing_section}\
+Begin by exploring the directory, then ask your first question.\
+"""
+
+
+def _build_system_prompt(
+    directory: str,
+    data_types: list[str],
+    existing_config: str | None = None,
+) -> tuple[str, Any]:
+    """Return ``(compiled_system_prompt, langfuse_prompt_obj_or_None)``."""
+    existing_section = (
+        "\nThe user already has the following ScoutConfig — refine it based on their answers:\n"
+        f"```json\n{existing_config}\n```\n"
+        if existing_config
+        else ""
+    )
+    template, prompt_obj = get_prompt_or_fallback(
+        "journey_system", _JOURNEY_SYSTEM_PROMPT
+    )
+    compiled = compile_prompt(
+        template,
+        prompt_obj,
+        directory=directory,
+        data_types=", ".join(data_types),
+        data_types_json=json.dumps(data_types),
+        config_start=_CONFIG_START,
+        config_end=_CONFIG_END,
+        existing_section=existing_section,
+    )
+    return compiled, prompt_obj
+
+
+# ── ScoutConfig extraction ────────────────────────────────────────────────
+
+
+def _extract_agent_config(text: str) -> str | None:
+    """Return validated ScoutConfig JSON string from between markers, or None.
+
+    Parses the JSON with Pydantic to ensure it conforms to the schema before
+    returning.  Returns None if markers are absent or JSON is invalid.
+    """
+    if _CONFIG_START not in text or _CONFIG_END not in text:
+        return None
+    start_idx = text.index(_CONFIG_START) + len(_CONFIG_START)
+    end_idx = text.index(_CONFIG_END)
+    raw = text[start_idx:end_idx].strip()
+    if not raw:
+        return None
+    try:
+        parsed = ScoutConfig.model_validate_json(raw)
+        return parsed.model_dump_json()
+    except Exception as exc:
+        logger.warning("agent_setup: failed to parse ScoutConfig JSON: %s", exc)
+        return None
+
+
+# ── LLM call with tool support ───────────────────────────────────────────
+
+
+def _as_text(content: Any) -> str:
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts: list[str] = []
+        for item in content:
+            if isinstance(item, str):
+                parts.append(item)
+            elif isinstance(item, dict):
+                text = item.get("text")
+                if isinstance(text, str):
+                    parts.append(text)
+        return "".join(parts)
+    return str(content)
+
+
+async def _call_llm_with_tools(
+    system_prompt: str,
+    history: list[dict[str, Any]],
+    tools: list[Any],
+    *,
+    user_id: str = "",
+    session_id: str = "",
+    langfuse_prompt: Any = None,
+) -> str:
+    """Build LangChain messages from history and invoke the LLM with tools.
+
+    Handles tool-calling loops: if the LLM calls tools, execute them and
+    continue until a final text response is produced.
+    """
+    lf = get_langfuse()
+    messages: list[Any] = [SystemMessage(content=system_prompt)]
+    for turn in history:
+        if turn["role"] == "user":
+            messages.append(HumanMessage(content=turn["content"]))
+        else:
+            messages.append(AIMessage(content=turn["content"]))
+
+    llm = get_agent_llm("setup", temperature=0.4)
+    llm_with_tools = llm.bind_tools(tools)
+    tool_map = {tool_def.name: tool_def for tool_def in tools}
+
+    _lf_ctx = langfuse_context(user_id=user_id or None, session_id=session_id or None)
+    _lf_ctx.__enter__()
+
+    _span_ctx = (
+        lf.start_as_current_observation(
+            as_type="span",
+            name="journey-setup",
+            input=history[-1]["content"] if history else "",
+        )
+        if lf else None
+    )
+    _span = _span_ctx.__enter__() if _span_ctx else None
+
+    try:
+        for step in range(_MAX_TOOL_STEPS):
+            _gen_ctx = (
+                lf.start_as_current_observation(
+                    as_type="generation",
+                    name="journey-setup-llm",
+                    model=model_for_agent("setup"),
+                    prompt=langfuse_prompt,
+                    input=messages,
+                )
+                if lf else None
+            )
+            _gen = _gen_ctx.__enter__() if _gen_ctx else None
+            response: AIMessage = await llm_with_tools.ainvoke(messages)
+            if _gen_ctx:
+                _gen.update(output=_as_text(response.content), usage_details=extract_usage(response))
+                _gen_ctx.__exit__(None, None, None)
+
+            resp_text = _as_text(response.content)
+
+            # Guard against empty responses (e.g. model returned finish_reason
+            # 'error' which LiteLLM maps to 'stop' with empty content).
+            if not response.tool_calls and not resp_text.strip():
+                logger.warning(
+                    "agent_setup: journey LLM returned empty response at step %d — retrying",
+                    step,
+                )
+                # Drop the empty AIMessage so we don't pollute history, and retry.
+                continue
+
+            messages.append(response)
+
+            if not response.tool_calls:
+                if _span:
+                    _span.update(output=resp_text)
+                return resp_text
+
+            for call in response.tool_calls:
+                call_name = str(call.get("name", ""))
+                call_args = call.get("args", {})
+                logger.info(
+                    "agent_setup: journey tool_call name=%s args=%s",
+                    call_name,
+                    json.dumps(call_args, ensure_ascii=True)[:500],
+                )
+
+                tool_fn = tool_map.get(call_name)
+                if tool_fn is None:
+                    tool_output = f"Unknown tool: {call_name}"
+                else:
+                    tool_output = await tool_fn.ainvoke(call_args)
+
+                logger.info(
+                    "agent_setup: journey tool_result name=%s output=%s",
+                    call_name,
+                    str(tool_output)[:800],
+                )
+                messages.append(ToolMessage(content=str(tool_output), tool_call_id=call["id"]))
+
+        # Fallback: exceeded max steps.
+        final = await llm.ainvoke(messages)
+        final_text = _as_text(final.content)
+        if _span:
+            _span.update(output=final_text)
+        return final_text or (
+            "Sorry, I had trouble processing the files. "
+            "Could you try again? If the issue persists, the files might be too large for me to analyse."
+        )
+    finally:
+        if _span_ctx:
+            _span_ctx.__exit__(None, None, None)
+        _lf_ctx.__exit__(None, None, None)
+        if lf:
+            lf.flush()
+
+
+# ── Journey handlers (called from device_ws.py) ──────────────────────────
+
+
+async def handle_journey_start(
+    user_id: str,
+    frame: dict[str, Any],
+) -> dict[str, Any]:
+    """Handle a ``journey_start`` WS frame.
+
+    Creates a session, runs the setup LLM with directory exploration,
+    and returns the ``journey_reply`` payload.
+    """
+    agent_type = frame.get("agent_type", "local")
+    directory = frame.get("directory", "")
+    data_types = frame.get("data_types", [])
+    existing_config = frame.get("existing_config")
+
+    # Use the session_id provided by the FE so the reply matches the
+    # listener key; fall back to a generated one if absent.
+    session_id = frame.get("session_id") or str(uuid.uuid4())
+    system_prompt, langfuse_prompt = _build_system_prompt(directory, data_types, existing_config)
+
+    session = JourneySession(
+        session_id=session_id,
+        user_id=user_id,
+        agent_type=agent_type,
+        directory=directory,
+        data_types=data_types,
+        system_prompt=system_prompt,
+        langfuse_prompt=langfuse_prompt,
+    )
+
+    # Seed with an initial user message — some providers require at least one
+    # user/input message to be present.
+    seed_history: list[dict[str, Any]] = [
+        {"role": "user", "content": "Hi, I'm ready to set up my agent. Please explore my directory and ask me your first question."},
+    ]
+    ai_reply = await _call_llm_with_tools(
+        system_prompt=system_prompt,
+        history=seed_history,
+        tools=make_directory_tools(directory),
+        user_id=user_id,
+        session_id=session_id,
+        langfuse_prompt=langfuse_prompt,
+    )
+
+    session.history.extend(seed_history)
+    session.history.append({"role": "assistant", "content": ai_reply})
+    _sessions[session_id] = session
+
+    logger.info(
+        "agent_setup: journey session %s started for user %s (directory=%s)",
+        session_id,
+        user_id,
+        directory,
+    )
+
+    # Check if the LLM produced the config on the first turn (unlikely but possible).
+    agent_config = _extract_agent_config(ai_reply)
+    done = agent_config is not None
+
+    display_message = ai_reply
+    if done:
+        display_message = (
+            ai_reply[: ai_reply.index(_CONFIG_START)].strip()
+            or "Here is your agent configuration. You can save it or continue refining."
+        )
+        _sessions.pop(session_id, None)
+
+    return {
+        "type": "journey_reply",
+        "session_id": session_id,
+        "message": display_message,
+        "done": done,
+        "agent_config": agent_config,
+    }
+
+
+async def handle_journey_message(
+    user_id: str,
+    frame: dict[str, Any],
+) -> dict[str, Any]:
+    """Handle a ``journey_message`` WS frame.
+
+    Appends the user message, calls the LLM, and returns the
+    ``journey_reply`` payload.
+    """
+    session_id = frame.get("session_id", "")
+    message = frame.get("message", "")
+
+    session = get_journey_session(session_id, user_id)
+    if session is None:
+        return {
+            "type": "journey_reply",
+            "session_id": session_id,
+            "message": "Journey session not found or expired. Please start a new setup.",
+            "done": True,
+            "agent_config": None,
+        }
+
+    # Append user turn.
+    session.history.append({"role": "user", "content": message})
+
+    # Call the LLM with tools.
+    session_tools = make_directory_tools(session.directory)
+    ai_reply = await _call_llm_with_tools(
+        system_prompt=session.system_prompt,
+        history=session.history,
+        tools=session_tools,
+        user_id=session.user_id,
+        session_id=session_id,
+        langfuse_prompt=session.langfuse_prompt,
+    )
+
+    session.history.append({"role": "assistant", "content": ai_reply})
+
+    # Check if the LLM produced the final config.
+    agent_config = _extract_agent_config(ai_reply)
+    done = agent_config is not None
+
+    # If the LLM didn't produce a config, nudge it once it hits the hard safety cap.
+    if not done:
+        turns = sum(1 for t in session.history if t["role"] == "user")
+        if turns >= _MAX_TURNS:
+            nudge_content = (
+                "[System: You have enough information. Please generate the final "
+                f"ScoutConfig JSON now, wrapped in {_CONFIG_START} / {_CONFIG_END} markers.]"
+            )
+            session.history.append({"role": "user", "content": nudge_content})
+
+            nudge_reply = await _call_llm_with_tools(
+                system_prompt=session.system_prompt,
+                history=session.history,
+                tools=session_tools,
+                user_id=session.user_id,
+                session_id=session_id,
+                langfuse_prompt=session.langfuse_prompt,
+            )
+            session.history.append({"role": "assistant", "content": nudge_reply})
+
+            agent_config = _extract_agent_config(nudge_reply)
+            if agent_config is not None:
+                done = True
+                ai_reply = nudge_reply
+
+    display_message = ai_reply
+    if done:
+        display_message = (
+            ai_reply[: ai_reply.index(_CONFIG_START)].strip()
+            if _CONFIG_START in ai_reply
+            else "Here is your agent configuration. You can save it or continue refining."
+        )
+        _sessions.pop(session_id, None)
+        logger.info("agent_setup: journey session %s completed for user %s", session_id, user_id)
+
+    return {
+        "type": "journey_reply",
+        "session_id": session_id,
+        "message": display_message,
+        "done": done,
+        "agent_config": agent_config,
+    }

app/api/routes/scout_webhooks.py

@@ -0,0 +1,120 @@
+"""Gmail Pub/Sub push receiver.
+
+Google Pub/Sub push subscriptions deliver Gmail watch notifications as POST
+requests with a JSON envelope. The body payload contains a base64-encoded
+JSON blob with ``emailAddress`` + ``historyId``. We resolve the user by
+email, look up their cloud_scout_configs row for provider='gmail', and
+hand off to ScoutEngine.trigger_scout.
+
+Authentication: Pub/Sub push includes an OIDC JWT in the Authorization
+header. We verify it against Google's public keys with the audience
+configured in our Pub/Sub subscription.
+
+Dev mode: when ``GMAIL_PUBSUB_AUDIENCE`` is empty, JWT verification is
+skipped and a warning is logged. Production must set this env var.
+"""
+
+from __future__ import annotations
+
+import base64
+import json
+import logging
+import uuid
+
+from fastapi import APIRouter, Header, HTTPException, Request, status
+from sqlalchemy import select
+
+from app.config.settings import settings
+from app.db import async_session
+from app.models import CloudScoutConfig, User
+from app.scouts.engine import ScoutEngine
+
+logger = logging.getLogger(__name__)
+router = APIRouter(prefix="/scouts/webhooks", tags=["scout-webhooks"])
+
+
+def _verify_pubsub_jwt(token: str) -> bool:
+    """Verify the Google Pub/Sub OIDC JWT.
+
+    Returns True when valid, False on any verification failure.
+
+    Dev skip: if ``settings.GMAIL_PUBSUB_AUDIENCE`` is empty, logs a
+    warning and returns True so local development works without a real
+    Pub/Sub subscription. Production must configure the audience.
+    """
+    if not token:
+        return False
+
+    if not settings.GMAIL_PUBSUB_AUDIENCE:
+        logger.warning(
+            "GMAIL_PUBSUB_AUDIENCE not set — skipping Pub/Sub JWT verification (dev mode only)"
+        )
+        return True
+
+    try:
+        from google.auth.transport import requests as g_requests  # noqa: PLC0415
+        from google.oauth2 import id_token  # noqa: PLC0415
+
+        id_token.verify_oauth2_token(
+            token,
+            g_requests.Request(),
+            audience=settings.GMAIL_PUBSUB_AUDIENCE,
+        )
+        return True
+    except Exception:
+        logger.warning("pubsub jwt verification failed", exc_info=True)
+        return False
+
+
+@router.post("/gmail", status_code=status.HTTP_204_NO_CONTENT)
+async def gmail_pubsub(
+    request: Request,
+    authorization: str = Header(default=""),
+) -> None:
+    """Receive a Gmail Pub/Sub push notification.
+
+    Verifies the OIDC JWT, decodes the Pub/Sub envelope, resolves the user
+    by email, and triggers ScoutEngine.trigger_scout for each enabled Gmail
+    scout belonging to that user.
+
+    Returns 204 No Content on success (including benign no-ops like unknown
+    email or empty message data). Returns 401 on JWT verification failure.
+    """
+    token = authorization.removeprefix("Bearer ").strip()
+    if not _verify_pubsub_jwt(token):
+        raise HTTPException(status.HTTP_401_UNAUTHORIZED, "Invalid Pub/Sub JWT")
+
+    body = await request.json()
+    msg = body.get("message") or {}
+    raw = msg.get("data")
+    if not raw:
+        return  # ack without action — empty message data
+
+    try:
+        decoded = json.loads(base64.b64decode(raw).decode())
+    except Exception:
+        logger.warning("pubsub payload decode failed")
+        return
+
+    email = decoded.get("emailAddress")
+    if not email:
+        return
+
+    async with async_session() as session:
+        user_q = await session.execute(select(User).where(User.email == email))
+        user = user_q.scalar_one_or_none()
+        if user is None:
+            logger.info("pubsub: no user for %s — ignoring", email)
+            return
+        scouts_q = await session.execute(
+            select(CloudScoutConfig).where(
+                CloudScoutConfig.user_id == user.id,
+                CloudScoutConfig.provider == "gmail",
+                CloudScoutConfig.enabled == True,  # noqa: E712
+            )
+        )
+        scouts = scouts_q.scalars().all()
+
+    engine = ScoutEngine()
+    for scout in scouts:
+        await engine.trigger_scout(uuid.UUID(str(scout.id)))

app/api/routes/scouts.py

@@ -0,0 +1,440 @@
+"""Scout routes.
+
+Backend responsibilities are intentionally minimal:
+    GET  /scouts/catalog         — static catalog for UI display
+    POST /scouts/can-create      — billing eligibility check
+    POST /scouts/trigger         — trigger a local scout run
+
+Scout configuration is owned by the Electron app and is not persisted
+in backend scout-config tables.
+
+Gmail OAuth setup (scout-specific consent):
+    GET  /scouts/oauth/gmail/authorize       — returns consent-screen URL
+    GET  /scouts/oauth/gmail/web-callback    — bounces to deep link (excluded from schema)
+    POST /scouts/oauth/gmail/callback        — exchanges code, stores encrypted token
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import secrets
+import time
+import urllib.parse
+import uuid
+from datetime import datetime, timezone
+
+from fastapi import APIRouter, Depends, HTTPException, status
+from fastapi.responses import RedirectResponse
+from sqlalchemy import func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from pydantic import BaseModel
+
+from app.api.deps import get_current_user
+from app.auth.oauth_providers import generate_pkce_pair
+from app.billing.tier_manager import FEATURES
+from app.config.settings import settings
+from app.core.scout_runner import is_agent_running, run_local_agent
+from app.core.device_manager import device_manager
+from app.core.note_summarizer import generate_note_summary
+from app.db import get_session
+from app.integrations import encrypt_token
+from app.models import CloudScoutConfig, ScoutRunLog, LocalScoutConfig
+from app.schemas import (
+    ScoutCatalogItem,
+    ScoutCreationCheckRequest,
+    ScoutCreationCheckResponse,
+    ScoutRunLogResponse,
+    ScoutTriggerRequest,
+    UserProfile,
+)
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/scouts", tags=["scouts"])
+
+
+# ── Datetime helpers ──────────────────────────────────────────────────
+
+def _dt_ms(dt: datetime) -> int:
+    return int(dt.timestamp() * 1000)
+
+
+def _dt_ms_opt(dt: datetime | None) -> int | None:
+    return int(dt.timestamp() * 1000) if dt else None
+
+
+def _to_data_types(values: list[str]) -> list[str]:
+    normalize = {
+        "task": "tasks",           "tasks": "tasks",
+        "note": "notes",           "notes": "notes",
+        "timeline": "timelines",   "timelines": "timelines",   "timelineEvents": "timelines",
+        "project": "projects",     "projects": "projects",
+    }
+    seen: set[str] = set()
+    result: list[str] = []
+    for v in values:
+        mapped = normalize.get(v)
+        if mapped and mapped not in seen:
+            seen.add(mapped)
+            result.append(mapped)
+    return result
+
+
+def _to_run_log_response(log: ScoutRunLog) -> ScoutRunLogResponse:
+    return ScoutRunLogResponse(
+        id=log.id,
+        agent_id=log.scout_id,
+        agent_type=log.scout_type,  # type: ignore[arg-type]
+        status=log.status,  # type: ignore[arg-type]
+        items_processed=log.items_processed,
+        items_created=log.items_created,
+        errors=log.errors or [],
+        started_at=_dt_ms(log.started_at),
+        completed_at=_dt_ms_opt(log.completed_at),
+    )
+
+
+def _enforce_agent_limit(tier: str, current_count: int) -> int:
+    limit: int = FEATURES.get(tier, FEATURES["free"])["batch_active"]
+    if limit != -1 and current_count >= limit:
+        raise HTTPException(
+            status_code=status.HTTP_403_FORBIDDEN,
+            detail=f"Agent limit ({limit}) reached for your tier. Upgrade to create more.",
+        )
+    return limit
+
+
+async def _enforce_run_frequency(
+    tier: str,
+    user_id: str,
+    db: AsyncSession,
+) -> None:
+    """Raise HTTP 402 if the user has exceeded their daily batch run limit."""
+    limit: int = FEATURES.get(tier, FEATURES["free"])["batch_runs_per_day"]
+    if limit == -1:
+        return  # unlimited
+
+    today_start = datetime.now(timezone.utc).replace(
+        hour=0, minute=0, second=0, microsecond=0
+    )
+    result = await db.execute(
+        select(func.count(ScoutRunLog.id)).where(
+            ScoutRunLog.user_id == user_id,
+            ScoutRunLog.started_at >= today_start,
+        )
+    )
+    runs_today: int = result.scalar_one()
+
+    if runs_today >= limit:
+        raise HTTPException(
+            status_code=status.HTTP_402_PAYMENT_REQUIRED,
+            detail=f"Daily batch run limit ({limit}) reached for your tier. Upgrade for more runs.",
+        )
+
+
+# ── Catalog ───────────────────────────────────────────────────────────
+
+@router.get("/catalog", response_model=list[ScoutCatalogItem])
+async def get_agent_catalog(
+    current_user: UserProfile = Depends(get_current_user),
+) -> list[ScoutCatalogItem]:
+    """Return the static list of available agent types and their descriptions."""
+    return [
+        ScoutCatalogItem(
+            type="local_directory",
+            name="Local Directory Monitor",
+            description="Watches local directories, extracts data from files using AI",
+        ),
+        ScoutCatalogItem(
+            type="gmail",
+            name="Gmail Connector",
+            description="Scans Gmail inbox, extracts tasks/notes from emails",
+        ),
+        ScoutCatalogItem(
+            type="teams",
+            name="Microsoft Teams Connector",
+            description="Monitors Teams messages, extracts action items",
+        ),
+        ScoutCatalogItem(
+            type="outlook",
+            name="Outlook Connector",
+            description="Scans Outlook inbox, extracts tasks/notes",
+        ),
+    ]
+
+
+@router.post("/can-create", response_model=ScoutCreationCheckResponse)
+async def can_create_agent(
+    body: ScoutCreationCheckRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> ScoutCreationCheckResponse:
+    """Check if the user can create one more agent based on billing tier.
+
+    Since configuration is client-owned, the Electron app sends its current
+    active agent count and the backend applies tier limits.
+    """
+    limit: int = FEATURES.get(current_user.tier, FEATURES["free"])["batch_active"]
+    allowed = limit == -1 or body.active_agents < limit
+    return ScoutCreationCheckResponse(
+        allowed=allowed,
+        tier=current_user.tier,
+        active_agents=body.active_agents,
+        limit=limit,
+    )
+
+
+@router.post("/trigger", response_model=ScoutRunLogResponse, status_code=status.HTTP_202_ACCEPTED)
+async def trigger_agent_run(
+    body: ScoutTriggerRequest,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> ScoutRunLogResponse:
+    """Trigger a local agent run using client-provided configuration."""
+    _enforce_agent_limit(current_user.tier, body.active_agents)
+    await _enforce_run_frequency(current_user.tier, current_user.id, db)
+
+    last_run_dt = (
+        datetime.fromtimestamp(body.last_run_at / 1000, tz=timezone.utc)
+        if body.last_run_at
+        else None
+    )
+    config = LocalScoutConfig(
+        id=str(uuid.uuid4()),
+        user_id=current_user.id,
+        device_id=body.device_id,
+        name="Local Directory Monitor",
+        directory_paths=[body.directory],
+        data_types=_to_data_types(body.what_to_extract),
+        prompt_template=body.custom_agent_prompt or "",
+        scout_config=body.agent_config,
+        file_extensions=[],
+        schedule_cron=body.batch_interval,
+        enabled=True,
+        last_run_at=last_run_dt,
+    )
+
+    # Use the FE's stable agent_id if provided, fall back to the ephemeral config id.
+    stable_agent_id = body.agent_id or config.id
+
+    if is_agent_running(stable_agent_id):
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="Agent is already running. Only one run per agent is allowed at a time.",
+        )
+
+    run_log = ScoutRunLog(
+        scout_id=stable_agent_id,
+        scout_type="local",
+        user_id=current_user.id,
+        status="running",
+    )
+    db.add(run_log)
+    await db.commit()
+    await db.refresh(run_log)
+
+    run_context = {
+        "type": "agent_batch",
+        "run_id": run_log.id,
+        "agent_id": stable_agent_id,
+    }
+
+    asyncio.create_task(
+        run_local_agent(current_user.id, config, run_log, device_manager, run_context)
+    )
+
+    return _to_run_log_response(run_log)
+
+
+# ── Note summary endpoint ──────────────────────────────────────────────────────
+
+
+class NoteSummarizeRequest(BaseModel):
+    title: str
+    content: str
+
+
+class NoteSummarizeResponse(BaseModel):
+    summary: str
+
+
+@router.post("/notes/summarize", response_model=NoteSummarizeResponse)
+async def summarize_note(
+    body: NoteSummarizeRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> NoteSummarizeResponse:
+    """Generate an AI summary for a note.  Used by the Electron backfill on startup."""
+    summary = await generate_note_summary(body.title, body.content)
+    return NoteSummarizeResponse(summary=summary)
+
+
+# ── Gmail OAuth setup (scout-specific) ───────────────────────────────────────
+
+# Scopes required for Gmail scout connectivity.
+_GMAIL_SCOUT_SCOPES = [
+    "openid",
+    "email",
+    "https://www.googleapis.com/auth/gmail.readonly",
+    "https://www.googleapis.com/auth/gmail.modify",
+]
+
+# Google OAuth endpoints.
+_GOOGLE_AUTH_URL = "https://accounts.google.com/o/oauth2/v2/auth"
+_GOOGLE_TOKEN_URL = "https://oauth2.googleapis.com/token"
+
+# In-memory pending OAuth states for scout Gmail consent:
+# state → (code_verifier, scout_id, user_id, expires_at_epoch_s)
+# Production note: replace with Redis for multi-process deployments.
+_pending_scout_oauth_states: dict[str, tuple[str, str, str, float]] = {}
+_SCOUT_OAUTH_TTL_SECONDS = 600  # 10 minutes
+
+
+def _scout_gmail_redirect_uri() -> str:
+    """Derive the scout Gmail web-callback URI from the configured base OAUTH_REDIRECT_URI.
+
+    ``OAUTH_REDIRECT_URI`` is the full path used for login OAuth
+    (e.g. http://localhost:8000/api/v1/auth/oauth/google/web-callback).
+    We strip the path to get the scheme+host base, then append the scout path.
+    """
+    parsed = urllib.parse.urlparse(settings.OAUTH_REDIRECT_URI)
+    base = f"{parsed.scheme}://{parsed.netloc}"
+    return f"{base}/api/v1/scouts/oauth/gmail/web-callback"
+
+
+class _ScoutGmailAuthorizeResponse(BaseModel):
+    authorize_url: str
+
+
+class _ScoutGmailCallbackBody(BaseModel):
+    code: str
+    state: str
+
+
+@router.get("/oauth/gmail/authorize", response_model=_ScoutGmailAuthorizeResponse)
+async def scout_gmail_oauth_authorize(
+    scout_id: str,
+    current_user: UserProfile = Depends(get_current_user),
+) -> _ScoutGmailAuthorizeResponse:
+    """Start the Gmail OAuth flow for a specific cloud scout.
+
+    Returns the Google consent-screen URL.  The client opens this URL in the
+    system browser; after consent Google redirects to web-callback which bounces
+    to the ``adiuvai://scout/oauth/gmail/callback`` deep link.
+    """
+    if not settings.GOOGLE_AUTH_CLIENT_ID or not settings.GOOGLE_AUTH_CLIENT_SECRET:
+        raise HTTPException(
+            status.HTTP_503_SERVICE_UNAVAILABLE,
+            "Google OAuth is not configured on this server",
+        )
+
+    code_verifier, code_challenge = generate_pkce_pair()
+    state = secrets.token_urlsafe(32)
+
+    # Purge expired states to prevent unbounded growth.
+    now = time.time()
+    expired = [s for s, (_, _, _, exp) in _pending_scout_oauth_states.items() if exp < now]
+    for s in expired:
+        del _pending_scout_oauth_states[s]
+
+    _pending_scout_oauth_states[state] = (code_verifier, scout_id, current_user.id, now + _SCOUT_OAUTH_TTL_SECONDS)
+
+    redirect_uri = _scout_gmail_redirect_uri()
+    params = {
+        "client_id": settings.GOOGLE_AUTH_CLIENT_ID,
+        "redirect_uri": redirect_uri,
+        "response_type": "code",
+        "scope": " ".join(_GMAIL_SCOUT_SCOPES),
+        "state": state,
+        "code_challenge": code_challenge,
+        "code_challenge_method": "S256",
+        "access_type": "offline",
+        "prompt": "consent",
+    }
+    authorize_url = f"{_GOOGLE_AUTH_URL}?{urllib.parse.urlencode(params)}"
+    return _ScoutGmailAuthorizeResponse(authorize_url=authorize_url)
+
+
+@router.get("/oauth/gmail/web-callback", include_in_schema=False)
+async def scout_gmail_oauth_web_callback(code: str, state: str) -> RedirectResponse:
+    """Google redirects here after Gmail consent.
+
+    Immediately bounces to the Electron deep link so the desktop app
+    receives the authorization code.
+    """
+    params = urllib.parse.urlencode({"code": code, "state": state})
+    deep_link = f"adiuvai://scout/oauth/gmail/callback?{params}"
+    return RedirectResponse(url=deep_link, status_code=302)
+
+
+@router.post("/oauth/gmail/callback")
+async def scout_gmail_oauth_callback(
+    body: _ScoutGmailCallbackBody,
+    db: AsyncSession = Depends(get_session),
+    current_user: UserProfile = Depends(get_current_user),
+) -> dict:
+    """Exchange the Gmail authorization code and store the encrypted token on the scout.
+
+    Called by the Electron app after it receives the deep-link callback with
+    the ``code`` and ``state`` params.
+    """
+    entry = _pending_scout_oauth_states.pop(body.state, None)
+    if entry is None or entry[3] < time.time() or entry[2] != current_user.id:
+        raise HTTPException(status.HTTP_401_UNAUTHORIZED, "Invalid or expired OAuth state")
+
+    code_verifier, scout_id, _, _ = entry
+
+    redirect_uri = _scout_gmail_redirect_uri()
+
+    import httpx
+    async with httpx.AsyncClient() as client:
+        response = await client.post(
+            _GOOGLE_TOKEN_URL,
+            data={
+                "client_id": settings.GOOGLE_AUTH_CLIENT_ID,
+                "client_secret": settings.GOOGLE_AUTH_CLIENT_SECRET,
+                "code": body.code,
+                "code_verifier": code_verifier,
+                "grant_type": "authorization_code",
+                "redirect_uri": redirect_uri,
+            },
+        )
+    try:
+        response.raise_for_status()
+    except httpx.HTTPStatusError as exc:
+        logger.error("Gmail token exchange failed: %s", exc.response.text)
+        raise HTTPException(status.HTTP_502_BAD_GATEWAY, "Failed to exchange Gmail authorization code")
+
+    token_data = response.json()
+
+    creds_dict: dict = {
+        "token": token_data["access_token"],
+        "refresh_token": token_data.get("refresh_token"),
+        "token_uri": _GOOGLE_TOKEN_URL,
+        "client_id": settings.GOOGLE_AUTH_CLIENT_ID,
+        "client_secret": settings.GOOGLE_AUTH_CLIENT_SECRET,
+        "scopes": [
+            "https://www.googleapis.com/auth/gmail.readonly",
+            "https://www.googleapis.com/auth/gmail.modify",
+        ],
+    }
+    encrypted = encrypt_token(creds_dict)
+
+    scout = await db.get(CloudScoutConfig, scout_id)
+    if scout is None or scout.user_id != current_user.id:
+        raise HTTPException(status.HTTP_404_NOT_FOUND, "Scout not found")
+    scout.oauth_token_encrypted = encrypted
+    await db.commit()
+
+    # Attempt to set up Gmail push watch so we start receiving Pub/Sub notifications.
+    from app.scouts.connectors.registry import get_connector
+    try:
+        connector = get_connector("gmail")
+        await connector.setup_watch(scout)
+        await db.commit()
+    except KeyError:
+        logger.warning("gmail connector not registered — skipping setup_watch for scout %s", scout_id)
+    except Exception:
+        logger.exception("setup_watch failed for scout %s", scout_id)
+
+    return {"ok": True}

app/auth/__init__.py

@@ -0,0 +1 @@
+"OAuth provider abstractions and utilities."

app/auth/oauth_providers.py

@@ -0,0 +1,135 @@
+"""OAuth 2.0 + PKCE provider abstractions.
+
+Each provider implements a three-step flow designed for a desktop (public) client:
+
+  1. get_authorization_url(state, code_challenge) → str
+       Build the provider's consent-screen URL.  State and code_challenge are
+       generated server-side; the client opens this URL in the system browser.
+
+  2. exchange_code(code, code_verifier, redirect_uri) → dict
+       Exchange the short-lived authorization code for an access token.
+       The code_verifier proves ownership of the PKCE challenge.
+
+  3. get_userinfo(access_token) → OAuthUserInfo
+       Fetch the canonical user identity from the provider.
+
+Currently supported providers:
+  - GoogleOAuthProvider  (scope: openid email profile)
+
+Adding a new provider:
+  - Implement the three methods above.
+  - Register in _PROVIDERS inside routes/auth.py.
+"""
+
+from __future__ import annotations
+
+import base64
+import hashlib
+import os
+import urllib.parse
+from dataclasses import dataclass
+
+import httpx
+
+
+# ── Data transfer objects ─────────────────────────────────────────────
+
+
+@dataclass
+class OAuthUserInfo:
+    """Normalized user identity returned by any provider."""
+
+    provider_user_id: str
+    email: str
+    email_verified: bool
+    avatar_url: str | None
+    name: str | None
+
+
+# ── PKCE helpers ──────────────────────────────────────────────────────
+
+
+def generate_pkce_pair() -> tuple[str, str]:
+    """Generate a (code_verifier, code_challenge) pair for PKCE S256.
+
+    The code_verifier is a random 32-byte URL-safe base64 string.
+    The code_challenge is SHA-256(code_verifier) base64url-encoded (no padding).
+    """
+    code_verifier = base64.urlsafe_b64encode(os.urandom(32)).rstrip(b"=").decode()
+    digest = hashlib.sha256(code_verifier.encode()).digest()
+    code_challenge = base64.urlsafe_b64encode(digest).rstrip(b"=").decode()
+    return code_verifier, code_challenge
+
+
+# ── Google provider ───────────────────────────────────────────────────
+
+
+class GoogleOAuthProvider:
+    """Google OAuth 2.0 provider (openid email profile scope).
+
+    Uses Google's standard authorization endpoint with PKCE S256.
+    Does NOT use google-auth-oauthlib to keep the flow generic and async.
+    """
+
+    name = "google"
+
+    _AUTH_URL = "https://accounts.google.com/o/oauth2/v2/auth"
+    _TOKEN_URL = "https://oauth2.googleapis.com/token"
+    _USERINFO_URL = "https://www.googleapis.com/oauth2/v3/userinfo"
+
+    def __init__(self, client_id: str, client_secret: str, redirect_uri: str) -> None:
+        self.client_id = client_id
+        self.client_secret = client_secret
+        self.redirect_uri = redirect_uri
+
+    def get_authorization_url(self, state: str, code_challenge: str) -> str:
+        """Build the Google consent-screen URL."""
+        params = {
+            "client_id": self.client_id,
+            "redirect_uri": self.redirect_uri,
+            "response_type": "code",
+            "scope": "openid email profile",
+            "state": state,
+            "code_challenge": code_challenge,
+            "code_challenge_method": "S256",
+            "access_type": "offline",
+            "prompt": "select_account",
+        }
+        return f"{self._AUTH_URL}?{urllib.parse.urlencode(params)}"
+
+    async def exchange_code(
+        self, code: str, code_verifier: str, redirect_uri: str
+    ) -> dict:
+        """Exchange authorization code for an access token."""
+        async with httpx.AsyncClient() as client:
+            response = await client.post(
+                self._TOKEN_URL,
+                data={
+                    "client_id": self.client_id,
+                    "client_secret": self.client_secret,
+                    "code": code,
+                    "code_verifier": code_verifier,
+                    "grant_type": "authorization_code",
+                    "redirect_uri": redirect_uri,
+                },
+            )
+        response.raise_for_status()
+        return response.json()
+
+    async def get_userinfo(self, access_token: str) -> OAuthUserInfo:
+        """Fetch the authenticated user's identity from Google."""
+        async with httpx.AsyncClient() as client:
+            response = await client.get(
+                self._USERINFO_URL,
+                headers={"Authorization": f"Bearer {access_token}"},
+            )
+        response.raise_for_status()
+        data = response.json()
+
+        return OAuthUserInfo(
+            provider_user_id=data["sub"],
+            email=data["email"],
+            email_verified=data.get("email_verified", False),
+            avatar_url=data.get("picture"),
+            name=data.get("name"),
+        )

app/billing/__init__.py

@@ -0,0 +1,4 @@
+from app.billing.stripe_service import stripe_service
+from app.billing.tier_manager import tier_manager
+
+__all__ = ["stripe_service", "tier_manager"]

app/billing/quota.py

@@ -0,0 +1,139 @@
+"""Quota checks and atomic token-usage accounting for folder integration."""
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime, timezone
+
+from sqlalchemy import select, update
+from sqlalchemy.dialects.postgresql import insert as pg_insert
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.billing.tier_manager import TierManager
+from app.models import MonthlyTokenUsage
+from app.schemas import BillingTier
+
+
+class QuotaExceeded(Exception):
+    """Raised when a folder operation cannot proceed under the user's tier."""
+
+    def __init__(self, reason: str, message: str) -> None:
+        super().__init__(message)
+        self.reason = reason  # "max_files" | "monthly_tokens"
+
+
+@dataclass
+class TokenUsageResult:
+    tokens_used: int
+    exhausted: bool
+
+
+def _current_year_month() -> str:
+    return datetime.now(timezone.utc).strftime("%Y-%m")
+
+
+_tier_manager = TierManager()
+
+
+async def check_folder_quota(
+    *,
+    user_id: str,
+    tier: BillingTier,
+    estimated_files: int,
+    db: AsyncSession,
+) -> None:
+    """Raise QuotaExceeded if folder_max_files or folder_monthly_tokens
+    would be violated. -1 in either feature means unlimited."""
+    max_files = _tier_manager.get_feature_value(tier, "folder_max_files")
+    if max_files != -1 and estimated_files > max_files:
+        raise QuotaExceeded(
+            "max_files",
+            f"Folder has {estimated_files} files; tier '{tier}' allows max {max_files}.",
+        )
+
+    cap = _tier_manager.get_feature_value(tier, "folder_monthly_tokens")
+    if cap == -1:
+        return
+    ym = _current_year_month()
+    row = (
+        await db.execute(
+            select(MonthlyTokenUsage).where(
+                MonthlyTokenUsage.user_id == user_id,
+                MonthlyTokenUsage.year_month == ym,
+                MonthlyTokenUsage.feature == "folder_index",
+            )
+        )
+    ).scalar_one_or_none()
+    used = row.tokens_used if row else 0
+    if used >= cap:
+        raise QuotaExceeded(
+            "monthly_tokens",
+            f"Monthly token budget exhausted ({used}/{cap}); resets next month.",
+        )
+
+
+async def add_token_usage(
+    *,
+    user_id: str,
+    feature: str,
+    tokens: int,
+    db: AsyncSession,
+    cap: int | None = None,
+) -> TokenUsageResult:
+    """Atomically add `tokens` to MonthlyTokenUsage row for (user, current month, feature).
+
+    Uses PostgreSQL ``INSERT … ON CONFLICT DO UPDATE`` when available; falls
+    back to a read-then-write on other engines (e.g. aiosqlite in tests).
+    Returns post-update total and whether cap is exhausted.
+    """
+    ym = _current_year_month()
+
+    # Detect dialect to choose between native upsert and portable fallback.
+    dialect_name: str = db.bind.dialect.name if db.bind is not None else ""  # type: ignore[union-attr]
+
+    if dialect_name == "postgresql":
+        # Native atomic upsert — production path.
+        stmt = (
+            pg_insert(MonthlyTokenUsage)
+            .values(
+                user_id=user_id,
+                year_month=ym,
+                feature=feature,
+                tokens_used=tokens,
+            )
+            .on_conflict_do_update(
+                index_elements=["user_id", "year_month", "feature"],
+                set_={"tokens_used": MonthlyTokenUsage.tokens_used + tokens},
+            )
+            .returning(MonthlyTokenUsage.tokens_used)
+        )
+        used: int = (await db.execute(stmt)).scalar_one()
+        await db.commit()
+    else:
+        # Portable fallback — used in tests (SQLite) and any non-PG engine.
+        row = (
+            await db.execute(
+                select(MonthlyTokenUsage).where(
+                    MonthlyTokenUsage.user_id == user_id,
+                    MonthlyTokenUsage.year_month == ym,
+                    MonthlyTokenUsage.feature == feature,
+                )
+            )
+        ).scalar_one_or_none()
+
+        if row is None:
+            row = MonthlyTokenUsage(
+                user_id=user_id,
+                year_month=ym,
+                feature=feature,
+                tokens_used=tokens,
+            )
+            db.add(row)
+        else:
+            row.tokens_used += tokens
+
+        await db.commit()
+        await db.refresh(row)
+        used = row.tokens_used
+
+    exhausted = cap is not None and cap != -1 and used >= cap
+    return TokenUsageResult(tokens_used=used, exhausted=exhausted)

app/billing/stripe_service.py

@@ -1,7 +1,7 @@
 """Stripe service: checkout sessions, webhook handling, subscription management.
 
-Adapted for the Billing microservice — uses shared.models and shared.db.
+Subscription records are persisted in the PostgreSQL ``subscriptions`` table.
-All Stripe calls are gracefully stubbed when STRIPE_SECRET_KEY is not
+All Stripe calls are gracefully stubbed when ``STRIPE_SECRET_KEY`` is not
 configured, enabling local development without live credentials.
 """
 
@@ -15,8 +15,7 @@ from fastapi import HTTPException, status
 from sqlalchemy import select
 from sqlalchemy.ext.asyncio import AsyncSession
 
-from shared.config import settings
+from app.config.settings import settings
-from shared.models import Subscription
 
 # Stripe price IDs per tier — replace with real IDs in production .env
 TIER_PRICE_IDS: dict[str, str] = {
@@ -44,10 +43,14 @@ class StripeService:
         self,
         user_id: str,
         tier: str,
-        success_url: str = "https://app.adiuva.app/billing/success?session_id={CHECKOUT_SESSION_ID}",
+        success_url: str = "https://app.adiuvai.app/billing/success?session_id={CHECKOUT_SESSION_ID}",
-        cancel_url: str = "https://app.adiuva.app/billing/cancel",
+        cancel_url: str = "https://app.adiuvai.app/billing/cancel",
     ) -> str:
-        """Create a Stripe checkout session and return the URL."""
+        """Create a Stripe checkout session and return the URL.
+
+        Returns a stub URL when Stripe is not configured.
+        Raises ``HTTP 400`` for the free tier or an unknown tier.
+        """
         if tier == "free":
             raise HTTPException(
                 status_code=status.HTTP_400_BAD_REQUEST,
@@ -84,6 +87,8 @@ class StripeService:
         """Process a Stripe webhook event.
 
         Verifies the signature, then dispatches on event type.
+        Raises ``HTTP 400`` on signature mismatch.
+        No-ops when Stripe is not configured.
         """
         if not self._configured():
             return
@@ -150,7 +155,9 @@ class StripeService:
     async def get_subscription(
         self, user_id: str, db: AsyncSession
     ) -> dict[str, Any] | None:
-        """Return the subscription record for user_id, or None."""
+        """Return the subscription record for ``user_id``, or ``None`` if absent."""
+        from app.models import Subscription  # noqa: PLC0415
+
         result = await db.execute(
             select(Subscription).where(Subscription.user_id == user_id)
         )
@@ -169,7 +176,12 @@ class StripeService:
         }
 
     async def cancel_subscription(self, user_id: str, db: AsyncSession) -> None:
-        """Cancel the user's Stripe subscription and downgrade to free."""
+        """Cancel the user's Stripe subscription and downgrade them to free.
+
+        Raises ``HTTP 404`` when no active subscription exists.
+        """
+        from app.models import Subscription  # noqa: PLC0415
+
         result = await db.execute(
             select(Subscription).where(Subscription.user_id == user_id)
         )
@@ -188,6 +200,45 @@ class StripeService:
         sub.status = "canceled"
         await db.commit()
 
+    async def list_invoices(
+        self, user_id: str, db: AsyncSession, limit: int = 24
+    ) -> list[dict[str, Any]]:
+        """Return recent invoices for the user from Stripe.
+
+        Returns an empty list when Stripe is not configured or the user has
+        no ``stripe_customer_id``.
+        """
+        if not self._configured():
+            return []
+
+        from app.models import User  # noqa: PLC0415
+
+        result = await db.execute(
+            select(User.stripe_customer_id).where(User.id == user_id)
+        )
+        customer_id = result.scalar_one_or_none()
+        if not customer_id:
+            return []
+
+        try:
+            s = self._client()
+            invoices = s.Invoice.list(customer=customer_id, limit=limit)
+            return [
+                {
+                    "id": inv.id,
+                    "amount_due": inv.amount_due,
+                    "amount_paid": inv.amount_paid,
+                    "currency": inv.currency,
+                    "status": inv.status,
+                    "created": inv.created * 1000,  # epoch ms
+                    "invoice_url": inv.hosted_invoice_url,
+                    "invoice_pdf": inv.invoice_pdf,
+                }
+                for inv in invoices.auto_paging_iter()
+            ]
+        except Exception:
+            return []
+
     # ── Private DB helpers ───────────────────────────────────────────────
 
     async def _upsert_subscription(
@@ -199,6 +250,8 @@ class StripeService:
         sub_status: str,
         current_period_end: datetime | None,
     ) -> None:
+        from app.models import Subscription  # noqa: PLC0415
+
         result = await db.execute(
             select(Subscription).where(Subscription.user_id == user_id)
         )
@@ -220,6 +273,8 @@ class StripeService:
         status: str | None = None,
         current_period_end: datetime | None = None,
     ) -> None:
+        from app.models import Subscription  # noqa: PLC0415
+
         result = await db.execute(
             select(Subscription).where(
                 Subscription.stripe_subscription_id == stripe_subscription_id
@@ -236,5 +291,5 @@ class StripeService:
             sub.current_period_end = current_period_end
 
 
-# Module-level singleton
+# Module-level singleton shared across the app.
 stripe_service = StripeService()

app/billing/tier_manager.py

@@ -0,0 +1,149 @@
+"""Tier manager: feature matrix and quota enforcement.
+
+``TierManager`` is the single source of truth for what each billing tier
+allows.  ``get_tier`` queries the ``subscriptions`` table for the live tier.
+Quota-enforcement helpers take ``tier`` directly — the caller already has it
+from ``current_user.tier`` (provided by ``get_current_user``).
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from fastapi import HTTPException, status
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.schemas import BillingTier
+
+# Feature matrix per tier.  -1 means unlimited; 0 means disabled.
+FEATURES: dict[str, dict[str, Any]] = {
+    "free": {
+        "agents": 3,
+        "batch_active": 2,
+        "batch_runs_per_day": 5,
+        "providers": 1,
+        "batch_builder": False,
+        "sso": False,
+        "real_embeddings": False,       # keyword fallback only
+        "realtime_extraction": False,   # batch queue (Phase 2)
+        "relational_memory": False,     # relational tier (Phase 3) — Pro+
+        "proactive_mining": False,      # Power+ only (Phase 5)
+        "folder_max_files": 200,
+        "folder_monthly_tokens": 100_000,
+    },
+    "pro": {
+        "agents": -1,           # unlimited
+        "batch_active": 10,
+        "batch_runs_per_day": 50,
+        "providers": -1,
+        "batch_builder": False,
+        "sso": False,
+        "real_embeddings": True,        # pgvector cosine search
+        "realtime_extraction": True,    # fire-and-forget asyncio.create_task
+        "relational_memory": True,      # person/project predicates
+        "proactive_mining": False,      # Power+ only (Phase 5)
+        "folder_max_files": 5000,
+        "folder_monthly_tokens": 2_000_000,
+    },
+    "power": {
+        "agents": -1,
+        "batch_active": -1,     # unlimited
+        "batch_runs_per_day": -1,  # unlimited
+        "providers": -1,
+        "batch_builder": True,
+        "sso": False,
+        "real_embeddings": True,
+        "realtime_extraction": True,
+        "relational_memory": True,      # all predicates incl. custom
+        "proactive_mining": True,       # scheduled pattern mining (Phase 5)
+        "folder_max_files": -1,         # unlimited
+        "folder_monthly_tokens": -1,    # unlimited
+    },
+    "team": {
+        "agents": -1,
+        "batch_active": -1,
+        "batch_runs_per_day": -1,  # unlimited
+        "providers": -1,
+        "batch_builder": True,
+        "sso": True,
+        "real_embeddings": True,
+        "realtime_extraction": True,
+        "relational_memory": True,      # all predicates incl. custom
+        "proactive_mining": True,       # scheduled pattern mining (Phase 5)
+        "folder_max_files": -1,         # unlimited
+        "folder_monthly_tokens": -1,    # unlimited
+    },
+}
+
+# Requests-per-minute limit per tier.
+RATE_LIMITS: dict[str, int] = {
+    "free": 20,
+    "pro": 60,
+    "power": 120,
+    "team": 200,
+}
+
+
+class TierManager:
+    """Centralises tier feature-gating, rate-limit lookups, and quota checks."""
+
+    # ── Tier lookup ─────────────────────────────────────────────────────
+
+    async def get_tier(self, user_id: str, db: AsyncSession) -> BillingTier:
+        """Return the current billing tier for ``user_id`` from the DB.
+
+        Falls back to ``'power'`` in dev (unlimited) or ``'free'`` in prod
+        when no subscription row exists.
+        """
+        from app.models import Subscription  # noqa: PLC0415
+        from app.config.settings import settings  # noqa: PLC0415
+
+        result = await db.execute(
+            select(Subscription.tier).where(Subscription.user_id == user_id)
+        )
+        tier: str | None = result.scalar_one_or_none()
+        if tier is None or tier not in FEATURES:
+            return "power" if settings.ENV == "dev" else "free"
+        return tier  # type: ignore[return-value]
+
+    # ── Feature access ───────────────────────────────────────────────────
+
+    def check_feature(self, tier: BillingTier, feature: str) -> bool:
+        """Return ``True`` if ``tier`` has ``feature`` enabled.
+
+        For numeric features, any value > 0 or -1 (unlimited) counts as enabled.
+        """
+        value = FEATURES.get(tier, FEATURES["free"]).get(feature)
+        if value is None:
+            return False
+        if isinstance(value, bool):
+            return value
+        return value != 0
+
+    def require_feature(self, tier: BillingTier, feature: str, tier_name: str = "") -> None:
+        """Raise ``HTTP 403`` if ``tier`` does not have ``feature``."""
+        if not self.check_feature(tier, feature):
+            detail = (
+                f"Feature '{feature}' requires {tier_name} tier or above."
+                if tier_name
+                else f"Feature '{feature}' is not available on your current tier."
+            )
+            raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail=detail)
+
+    def get_feature_value(self, tier: BillingTier, feature: str) -> int:
+        """Return integer feature value for tier. -1 means unlimited."""
+        value = FEATURES.get(tier, FEATURES["free"]).get(feature)
+        if not isinstance(value, int):
+            return 0
+        return value
+
+    # ── Rate limiting ────────────────────────────────────────────────────
+
+    def get_rate_limit(self, tier: BillingTier) -> int:
+        """Return the requests-per-minute limit for ``tier``."""
+        return RATE_LIMITS.get(tier, RATE_LIMITS["free"])
+
+
+# Module-level singleton shared across the app.
+tier_manager = TierManager()

app/config/settings.py

@@ -0,0 +1,95 @@
+from typing import Literal
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+
+class Settings(BaseSettings):
+    DATABASE_URL: str = "postgresql+asyncpg://postgres:postgres@localhost:5432/adiuvai"
+    JWT_SECRET: str = "change-me-in-production"
+    JWT_ALGORITHM: str = "HS256"
+    JWT_ACCESS_TOKEN_EXPIRE_MINUTES: int = 30
+    JWT_REFRESH_TOKEN_EXPIRE_DAYS: int = 30
+
+    STRIPE_SECRET_KEY: str = ""
+    STRIPE_WEBHOOK_SECRET: str = ""
+
+    OPENAI_API_KEY: str = ""
+    ANTHROPIC_API_KEY: str = ""
+    GOOGLE_API_KEY: str = ""
+    CEREBRAS_API_KEY: str = ""
+    GROQ_API_KEY: str = ""
+    DEEPSEEK_API_KEY: str = ""
+
+    LLM_MODEL: str = "gpt-4o"
+    LLM_EMBED_MODEL: str = "text-embedding-3-small"
+
+    # Per-agent model overrides. Leave empty to fall back to LLM_MODEL.
+    LLM_MODEL_CLASSIFIER: str = ""        # classifier (intent routing, future use)
+    LLM_MODEL_HOME_AGENT: str = ""        # home-agent (run_single_agent / stream)
+    LLM_MODEL_UNIFIED_PROCESSOR: str = "" # unified-processor (agent_runner)
+    LLM_MODEL_CLOUD_PROCESSOR: str = ""   # cloud-processor (agent_runner)
+    LLM_MODEL_BRIEF_AGENT: str = ""            # brief-agent (home + project text briefs)
+    LLM_MODEL_TASK_BRIEF_AGENT: str = ""      # task-brief-agent (per-task deep research)
+    LLM_MODEL_SETUP_AGENT: str = ""           # agent-setup journey
+    LLM_MODEL_MEMORY_EXTRACTOR: str = ""  # memory-extractor (Phase 2 extract/decide)
+    LLM_MODEL_MEMORY_MINER: str = ""      # memory-miner (Phase 5 proactive mining)
+    LLM_MODEL_MEMORY_AUDITOR: str = ""    # memory-auditor (Phase 7 weekly audit)
+
+    # GitHub Copilot OAuth token storage directory.
+    # Leave empty to use the LiteLLM default (~/.config/litellm/github_copilot).
+    # In Docker, set this to a path backed by a named volume so tokens survive restarts.
+    GITHUB_COPILOT_TOKEN_DIR: str = ""
+
+    # OAuth client credentials — used for Gmail and Microsoft (Outlook/Teams) flows.
+    GMAIL_CLIENT_ID: str = ""
+    GMAIL_CLIENT_SECRET: str = ""
+    MS_CLIENT_ID: str = ""
+    MS_CLIENT_SECRET: str = ""
+    # MS_TENANT_ID: set to 'common' to allow multi-tenant (personal + work accounts).
+    MS_TENANT_ID: str = "common"
+
+    # Google Login OAuth credentials — scope: openid email profile.
+    # Separate from GMAIL_CLIENT_ID/SECRET (which uses gmail.readonly scope).
+    GOOGLE_AUTH_CLIENT_ID: str = ""
+    GOOGLE_AUTH_CLIENT_SECRET: str = ""
+    # The redirect URI registered in Google Cloud Console.
+    # Google redirects here after consent; this backend route then bounces to
+    # the adiuvai:// deep link so the Electron app receives the code.
+    # Dev:  http://localhost:8000/api/v1/auth/oauth/google/web-callback
+    # Prod: https://api.adiuvai.com/api/v1/auth/oauth/google/web-callback
+    OAUTH_REDIRECT_URI: str = "http://localhost:8000/api/v1/auth/oauth/google/web-callback"
+
+    # Gmail Pub/Sub topic for push notifications.
+    # Full resource name, e.g. "projects/my-project/topics/gmail-push".
+    # Leave empty in dev — setup_watch will skip registration gracefully.
+    GMAIL_PUBSUB_TOPIC: str = ""
+    # OIDC token audience for Pub/Sub push subscription JWT verification.
+    # Set to the service account email or audience string configured in the
+    # Pub/Sub push subscription. Leave empty in dev to skip verification
+    # (a warning is logged — never silent in production).
+    GMAIL_PUBSUB_AUDIENCE: str = ""
+
+    # Fernet key (URL-safe base64, 32-byte key) for at-rest encryption of OAuth
+    # tokens stored in cloud_agent_configs.oauth_token_encrypted.
+    # Generate with: from cryptography.fernet import Fernet; Fernet.generate_key()
+    OAUTH_ENCRYPTION_KEY: str = ""
+
+    CORS_ORIGINS: list[str] = [
+        "app://.",
+        "http://localhost:3000",
+        "http://localhost:5173",
+        "http://localhost:4173",      # Vite preview (web SPA)
+        "https://app.adiuvai.com",    # Production web portal
+    ]
+
+    LANGFUSE_SECRET_KEY: str = ""
+    LANGFUSE_PUBLIC_KEY: str = ""
+    LANGFUSE_BASE_URL: str = "https://cloud.langfuse.com"
+
+    SCHEDULER_ENABLED: bool = True
+
+    ENV: Literal["dev", "prod"] = "dev"
+
+    model_config = SettingsConfigDict(env_file=".env", env_file_encoding="utf-8", extra="ignore")
+
+
+settings = Settings()

app/core/brief_agent.py

@@ -0,0 +1,228 @@
+"""Brief agent — produces plain-text home and project status briefs.
+
+Read-only tool subset only. Never calls _normalize_tagged_list_lines —
+the brief prompt forbids XML tags, so skipping post-processing is intentional.
+"""
+
+from __future__ import annotations
+
+from collections.abc import AsyncGenerator
+from datetime import date
+from typing import Any
+
+from app.agents.note_agent import NOTE_READ_TOOLS
+from app.agents.project_agent import PROJECT_READ_TOOLS
+from app.agents.task_agent import TASK_READ_TOOLS
+from app.agents.timeline_agent import TIMELINE_READ_TOOLS
+from app.core.deep_agent import (
+    _language_instruction,
+    _proactive_hints_injection,
+    _read_only_memory_tools,
+    _relational_memory_injection,
+    _run_single_agent_stream,
+    _trace_id_from_context,
+    build_brief_multi_project_manifest,
+)
+from app.core.langfuse_client import compile_prompt, get_prompt_or_fallback
+
+_LANGUAGE_NAMES: dict[str, str] = {
+    "en": "English", "it": "Italian", "es": "Spanish",
+    "fr": "French", "de": "German",
+    "english": "English", "italian": "Italian", "italiano": "Italian",
+    "spanish": "Spanish", "español": "Spanish",
+    "french": "French", "français": "French",
+    "german": "German", "deutsch": "German",
+}
+
+_HOME_BRIEF_FALLBACK = """\
+You are the user's personal assistant producing a short daily brief.
+
+ROLE
+Act like a calm, attentive secretary writing a stand-up note for your boss.
+Warm and human, never breezy. Never cheerful filler, never emojis, never
+"here is your brief" meta-text. The user is opening the app mid-workday and
+is probably stressed — your job is to lower cognitive load, not add noise.
+
+TOOLS — always call before writing
+Pull fresh data every run. Do not invent counts or titles. Use at minimum:
+- list_tasks_due_today — tasks the user owes today
+- list_timelines_today — events starting or ending today
+- list_all_projects — projects currently in progress or at risk
+- memory_list_blocks / memory_get — personal context about people, clients,
+  payment habits, working preferences
+If a tool returns nothing, simply omit that topic. Never report zeros.
+
+WHAT TO INCLUDE
+1. Tasks due today (title + priority; group the 1-2 most important).
+2. Timeline events starting or ending today (and anything that starts/ends
+   tomorrow if the user has a very light day).
+3. Active projects that need a nudge — stalled, blocked, or awaiting input.
+4. Memory-aware colour where it sharpens the brief. Examples:
+   - "Client Rossi tends to pay late — the Acme invoice is 6 days out."
+   - "You usually dislike meetings before 10:00 — the call at 09:30 is unusual."
+   Only add a memory line when it changes what the user does. Do not pad.
+
+WHAT TO OMIT
+- Zero-counts ("no overdue items", "0 meetings today").
+- Statistics ("2 active projects, 3 completed tasks").
+- Headers, titles, greetings, sign-offs, dates, emojis, slang.
+- Meta-phrases ("here is", "let me know if", "hope this helps").
+- XML/HTML tags of any kind. Plain prose only.
+
+LIGHT-DAY CLAUSE
+If tasks + events + active-project-nudges together produce fewer than two
+sentences of content, also list 1-2 projects in status on_hold or waiting
+and ask a single, specific question about them — e.g. "Is the Bianchi
+redesign still paused, or ready to pick back up?" One question max, grounded
+in a real project name.
+
+VOICE
+- Calm. Concise. Human. Short sentences.
+- Use **bold** sparingly for task titles, project names, and people's names.
+- No bullet lists. Flow as 2-4 sentences of prose.
+
+LENGTH
+2-4 sentences total. Hard cap 4. If the day is truly empty, one sentence.
+
+Respond in the user's language ({language}). Today is {today}.\
+"""
+
+_PROJECT_BRIEF_FALLBACK = """\
+You are the project assistant producing a short status brief for ONE project.
+
+ROLE
+A senior project manager summarising state-of-play for the owner. Factual,
+sharp, forward-looking. Never reassuring filler, never emojis.
+
+SCOPE
+Work only with project_id = {project_id}. Do not mention or pull data from
+other projects. Use tools to fetch fresh data:
+- get_project — current status, dates, description
+- list_tasks(project_id) — open work, split by status
+- list_timelines(project_id) — milestones hit, upcoming, overdue
+- list_notes(project_id) — any recent decisions or blockers
+- memory_get — relevant context about the client, collaborators, constraints
+
+STRUCTURE — follow exactly, one short paragraph per section, no headers
+1. **State.** One sentence: current phase, health (on track / at risk / blocked),
+   and why. Cite the concrete signal (overdue milestone, stalled tasks, recent
+   blocker note).
+2. **What's moving.** What was completed or progressed recently. Name specific
+   tasks or milestones.
+3. **Next steps.** The 1-3 most important things the user should do next, in
+   priority order. Be concrete — task name, who owns it, when due if known.
+   If waiting on someone else, name them and what the ask is.
+4. **Risks / memory-flagged items.** One line max. Only include when there is
+   a real risk or a relevant memory (e.g. late-paying client, tight deadline,
+   scope change). Omit the section entirely if nothing to say.
+
+WHAT TO OMIT
+- Zero-counts ("no overdue tasks").
+- Generic advice ("keep up the good work").
+- Greetings, headers, bullet lists, emojis, sign-offs, meta-phrases.
+- XML/HTML tags or bracketed id lists. Plain prose only.
+
+VOICE
+- Direct. Factual. No fluff.
+- Use **bold** sparingly for task titles, milestone names, and the owner's name.
+- Short sentences. Prefer verbs over nouns ("Client review is blocking release"
+  not "There is a blocker which is the client review").
+
+LENGTH
+4-8 sentences total across the 3-4 sections. Hard cap 8.
+
+Respond in the user's language ({language}). Today is {today}.\
+"""
+
+
+def _resolve_language(context: dict[str, Any]) -> str:
+    core = context.get("core_memory") or {}
+    raw = (core.get("language") or "en").strip().lower()
+    return _LANGUAGE_NAMES.get(raw, raw.title()) or "English"
+
+
+def _build_read_tools(user_id: str, trace_id: str | None) -> list[Any]:
+    return [
+        *TASK_READ_TOOLS,
+        *PROJECT_READ_TOOLS,
+        *TIMELINE_READ_TOOLS,
+        *NOTE_READ_TOOLS,
+        *_read_only_memory_tools(user_id, trace_id),
+    ]
+
+
+async def run_home_brief(
+    user_id: str,
+    context: dict[str, Any],
+) -> AsyncGenerator[tuple[str, Any], None]:
+    """Stream a plain-text daily home brief.
+
+    Yields (event_type, data) tuples identical to _run_single_agent_stream.
+    Do NOT post-process output through _normalize_tagged_list_lines.
+    """
+    from app.agents.folder_agent import FOLDER_TOOLS
+
+    trace_id = _trace_id_from_context(context)
+    today = date.today().isoformat()
+    language = _resolve_language(context)
+
+    raw_template, langfuse_prompt = get_prompt_or_fallback("home_brief", _HOME_BRIEF_FALLBACK)
+    system_prompt = compile_prompt(raw_template, langfuse_prompt, language=language, today=today)
+    system_prompt += _relational_memory_injection(context)
+    system_prompt += _proactive_hints_injection(context)
+    system_prompt += _language_instruction(context)
+    if today not in system_prompt:
+        system_prompt += f"\nToday is {today}."
+
+    brief_manifest = await build_brief_multi_project_manifest()
+    system_prompt = system_prompt + ("\n\n" + brief_manifest if brief_manifest else "")
+
+    tools = [*_build_read_tools(user_id, trace_id), *FOLDER_TOOLS]
+    async for event in _run_single_agent_stream(
+        user_id=user_id,
+        system_prompt=system_prompt,
+        message="Generate the daily brief.",
+        context=context,
+        langfuse_prompt=langfuse_prompt,
+        agent_name="brief-agent",
+        tools=tools,
+    ):
+        yield event
+
+
+async def run_project_brief(
+    user_id: str,
+    project_id: str,
+    context: dict[str, Any],
+) -> AsyncGenerator[tuple[str, Any], None]:
+    """Stream a plain-text project status brief for project_id.
+
+    Yields (event_type, data) tuples identical to _run_single_agent_stream.
+    Do NOT post-process output through _normalize_tagged_list_lines.
+    """
+    trace_id = _trace_id_from_context(context)
+    today = date.today().isoformat()
+    language = _resolve_language(context)
+
+    raw_template, langfuse_prompt = get_prompt_or_fallback("project_brief", _PROJECT_BRIEF_FALLBACK)
+    system_prompt = compile_prompt(
+        raw_template, langfuse_prompt,
+        language=language, today=today, project_id=project_id,
+    )
+    system_prompt += _relational_memory_injection(context)
+    system_prompt += _proactive_hints_injection(context)
+    system_prompt += _language_instruction(context)
+    if today not in system_prompt:
+        system_prompt += f"\nToday is {today}."
+
+    tools = _build_read_tools(user_id, trace_id)
+    async for event in _run_single_agent_stream(
+        user_id=user_id,
+        system_prompt=system_prompt,
+        message=f"Generate the project status brief for project {project_id}.",
+        context=context,
+        langfuse_prompt=langfuse_prompt,
+        agent_name="brief-agent",
+        tools=tools,
+    ):
+        yield event

app/core/device_manager.py

@@ -0,0 +1,151 @@
+"""Device connection manager.
+
+Maintains in-memory state for all active Electron → backend WebSocket
+connections.  One connection per user (latest replaces previous).
+
+The manager handles the **tool-call round-trip** pattern:
+  - Backend sends ``tool_call`` frame → Electron executes the action →
+    returns ``tool_result`` frame.
+  - ``create_pending_call`` registers a Future keyed by ``call_id``.
+  - ``resolve_pending_call`` fulfils the Future; callers awaiting it
+    receive the result dict from Electron.
+
+This pattern is used by all tools (CRUD, file-system, etc.) via
+``execute_on_client()`` in ``ws_context.py``.
+
+The ``device_manager`` module-level singleton is imported by both the
+device WS route and the agent runner.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+from dataclasses import dataclass, field
+
+from fastapi import WebSocket
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class DeviceConnection:
+    """State for a single connected Electron device."""
+
+    ws: WebSocket
+    device_id: str
+    # Futures indexed by tool_call id — resolved when tool_result arrives.
+    pending_calls: dict[str, asyncio.Future[dict]] = field(default_factory=dict)
+
+
+class DeviceConnectionManager:
+    """Singleton registry of active Electron WebSocket connections.
+
+    Thread/task safety note: asyncio is single-threaded by design.  All
+    mutations happen inside await-points on the main event loop, so no
+    locking is required for the in-memory dicts.
+    """
+
+    def __init__(self) -> None:
+        self._connections: dict[str, DeviceConnection] = {}
+
+    # ── Registration ──────────────────────────────────────────────────
+
+    def register(self, user_id: str, device_id: str, ws: WebSocket) -> None:
+        """Store the active connection for *user_id*, replacing any previous one."""
+        if user_id in self._connections:
+            old = self._connections[user_id]
+            logger.info(
+                "device_manager: replacing existing connection for user=%s device=%s",
+                user_id,
+                old.device_id,
+            )
+            # Cancel any futures that were waiting on the old connection.
+            for fut in old.pending_calls.values():
+                if not fut.done():
+                    fut.cancel()
+        self._connections[user_id] = DeviceConnection(ws=ws, device_id=device_id)
+        logger.info(
+            "device_manager: registered user=%s device=%s", user_id, device_id
+        )
+
+    def unregister(self, user_id: str) -> None:
+        """Remove the connection for *user_id* and cancel any pending futures."""
+        conn = self._connections.pop(user_id, None)
+        if conn is None:
+            return
+        for fut in conn.pending_calls.values():
+            if not fut.done():
+                fut.cancel()
+        logger.info("device_manager: unregistered user=%s", user_id)
+
+    # ── Presence queries ──────────────────────────────────────────────
+
+    def get_ws(self, user_id: str) -> WebSocket | None:
+        """Return the active WebSocket for *user_id*, or ``None`` if offline."""
+        conn = self._connections.get(user_id)
+        return conn.ws if conn else None
+
+    def is_online(self, user_id: str, device_id: str | None = None) -> bool:
+        """Return ``True`` if the user has an active connection.
+
+        If *device_id* is provided also checks that it matches the connected device.
+        """
+        conn = self._connections.get(user_id)
+        if conn is None:
+            return False
+        if device_id is not None:
+            return conn.device_id == device_id
+        return True
+
+    # ── Frame sending ─────────────────────────────────────────────────
+
+    async def send_frame(self, user_id: str, frame: dict) -> None:
+        """Send *frame* as a JSON text message to the device.
+
+        Raises ``RuntimeError`` if the user is not connected.
+        """
+        conn = self._connections.get(user_id)
+        if conn is None:
+            raise RuntimeError(
+                f"send_frame: user {user_id!r} is not connected"
+            )
+        await conn.ws.send_text(json.dumps(frame))
+
+    # ── Tool-call round-trip ──────────────────────────────────────────
+
+    def create_pending_call(
+        self, user_id: str, call_id: str
+    ) -> asyncio.Future[dict]:
+        """Register a Future that will be resolved when the tool_result arrives.
+
+        Raises ``RuntimeError`` if the user is not connected.
+        """
+        conn = self._connections.get(user_id)
+        if conn is None:
+            raise RuntimeError(
+                f"create_pending_call: user {user_id!r} is not connected"
+            )
+        loop = asyncio.get_event_loop()
+        fut: asyncio.Future[dict] = loop.create_future()
+        conn.pending_calls[call_id] = fut
+        return fut
+
+    def resolve_pending_call(
+        self, user_id: str, call_id: str, result: dict
+    ) -> None:
+        """Fulfil the Future registered under *call_id* with the Electron result.
+
+        No-ops if the call_id is unknown (already timed out or cancelled).
+        """
+        conn = self._connections.get(user_id)
+        if conn is None:
+            return
+        fut = conn.pending_calls.pop(call_id, None)
+        if fut is not None and not fut.done():
+            fut.set_result(result)
+
+
+# Module-level singleton — import this everywhere.
+device_manager = DeviceConnectionManager()

app/core/embeddings.py

@@ -0,0 +1,34 @@
+"""OpenAI embedding helper for associative memory tier.
+
+Single public function: ``embed_text(text) -> list[float] | None``.
+Returns None on any failure — callers must implement a keyword fallback.
+Never raises; all exceptions are logged as warnings.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from openai import AsyncOpenAI
+
+logger = logging.getLogger(__name__)
+
+_MAX_INPUT_CHARS = 8000
+_EMBEDDING_MODEL = "text-embedding-3-small"
+
+
+async def embed_text(text: str) -> list[float] | None:
+    """Call OpenAI text-embedding-3-small. Return None on failure (caller falls back to keyword)."""
+    try:
+        client = AsyncOpenAI()
+        truncated = text[:_MAX_INPUT_CHARS]
+        response = await client.embeddings.create(
+            input=truncated,
+            model=_EMBEDDING_MODEL,
+        )
+        result: list[float] = response.data[0].embedding
+        logger.debug("embeddings: embed_text dims=%d", len(result))
+        return result
+    except Exception as exc:
+        logger.warning("embeddings: embed_text failed: %s", exc)
+        return None

app/core/folder_indexer.py

@@ -0,0 +1,183 @@
+"""Per-file summarisation for project folder integration."""
+from __future__ import annotations
+
+import base64
+import io
+from dataclasses import dataclass
+
+from langchain_core.messages import HumanMessage, SystemMessage
+from pypdf import PdfReader
+from docx import Document as DocxDocument
+
+from app.core.langfuse_client import (
+    compile_prompt,
+    extract_usage,
+    get_langfuse,
+    get_prompt_or_fallback,
+)
+from app.core.llm import get_llm
+
+_TEXT_FALLBACK = (
+    "You are summarising a file for an AI assistant that helps the user manage a project.\n"
+    "Produce a single sentence (<=30 words, <=200 chars) that captures the file's purpose "
+    "and most important detail.\nFile extension: {ext}\nFile name: {name}\nContent (truncated if long):\n{content}"
+)
+_IMAGE_FALLBACK = (
+    "You are summarising an image attached to a project folder.\n"
+    "Produce a single sentence (<=30 words, <=200 chars) describing what the image shows "
+    "and any obvious purpose (logo, screenshot, diagram, photo of a whiteboard, etc.)."
+)
+_MAX_INPUT_CHARS = 6000
+
+
+@dataclass
+class IndexResult:
+    summary: str
+    tokens_used: int
+
+
+async def _llm_text(messages: list) -> object:
+    """Make the LLM call for text summarisation.
+
+    Defined as a standalone async function so tests can patch it cleanly
+    without needing to mock the LLM object itself.
+    """
+    llm = get_llm(model="gpt-4o-mini", temperature=0.2)
+    return await llm.ainvoke(messages)
+
+
+async def _llm_vision(messages: list) -> object:
+    """Make the LLM call for vision (image) summarisation.
+
+    Accepts the message list and returns the response directly, mirroring
+    the ``_llm_text`` caller pattern so tests can patch it at the module level.
+    """
+    llm = get_llm(model="gpt-4o-mini", temperature=0.2)
+    return await llm.ainvoke(messages)
+
+
+async def summarize_image(*, image_b64: str, mime: str, file_name: str | None = None) -> IndexResult:
+    """Return a compact summary of an image file using vision.
+
+    Parameters
+    ----------
+    image_b64:
+        Base64-encoded image bytes.
+    mime:
+        MIME type of the image, e.g. ``"image/png"``.
+    file_name:
+        Optional file name, attached to the Langfuse trace as input metadata.
+    """
+    template, prompt_obj = get_prompt_or_fallback("folder_file_summary_image", _IMAGE_FALLBACK)
+    messages = [
+        SystemMessage(content=template),
+        HumanMessage(content=[
+            {"type": "text", "text": "Summarise this image."},
+            {"type": "image_url", "image_url": {"url": f"data:{mime};base64,{image_b64}"}},
+        ]),
+    ]
+    lf = get_langfuse()
+    if lf is not None:
+        with lf.start_as_current_observation(
+            as_type="generation",
+            name="folder-summarize-image",
+            model="gpt-4o-mini",
+            prompt=prompt_obj,
+            input={"file_name": file_name, "mime": mime},
+        ) as gen:
+            response = await _llm_vision(messages)
+            usage = extract_usage(response)
+            gen.update(output=response.content, usage_details=usage)
+    else:
+        response = await _llm_vision(messages)
+        usage = extract_usage(response)
+    summary = (response.content or "").strip()[:500]
+    return IndexResult(summary=summary, tokens_used=usage.get("total", 0))
+
+
+async def summarize_text(*, content: str, ext: str, name: str) -> IndexResult:
+    """Return a compact summary of a text file.
+
+    Parameters
+    ----------
+    content:
+        Raw text content of the file (will be truncated to _MAX_INPUT_CHARS).
+    ext:
+        File extension including the leading dot, e.g. ``".md"``.
+    name:
+        File name, e.g. ``"kickoff.md"``.
+    """
+    template, prompt_obj = get_prompt_or_fallback("folder_file_summary_text", _TEXT_FALLBACK)
+    truncated = content[:_MAX_INPUT_CHARS]
+    compiled = compile_prompt(template, prompt_obj, ext=ext, name=name, content=truncated)
+    messages = [
+        SystemMessage(content=compiled),
+        HumanMessage(content="Summarise this file."),
+    ]
+    lf = get_langfuse()
+    if lf is not None:
+        with lf.start_as_current_observation(
+            as_type="generation",
+            name="folder-summarize-text",
+            model="gpt-4o-mini",
+            prompt=prompt_obj,
+            input={"file_name": name, "ext": ext, "content_chars": len(truncated)},
+        ) as gen:
+            response = await _llm_text(messages)
+            usage = extract_usage(response)
+            gen.update(output=response.content, usage_details=usage)
+    else:
+        response = await _llm_text(messages)
+        usage = extract_usage(response)
+    summary = (response.content or "").strip()[:500]
+    return IndexResult(summary=summary, tokens_used=usage.get("total", 0))
+
+
+def _extract_pdf_text(pdf_b64: str) -> str:
+    buf = io.BytesIO(base64.b64decode(pdf_b64))
+    reader = PdfReader(buf)
+    parts: list[str] = []
+    for page in reader.pages:
+        try:
+            parts.append(page.extract_text() or "")
+        except Exception:
+            continue
+    return "\n".join(parts).strip()
+
+
+def _extract_docx_text(docx_b64: str) -> str:
+    buf = io.BytesIO(base64.b64decode(docx_b64))
+    doc = DocxDocument(buf)
+    return "\n".join(p.text for p in doc.paragraphs if p.text).strip()
+
+
+async def summarize_pdf(*, pdf_b64: str, name: str) -> IndexResult:
+    """Return a compact summary of a PDF file.
+
+    Parameters
+    ----------
+    pdf_b64:
+        Base64-encoded PDF bytes.
+    name:
+        File name, e.g. ``"report.pdf"``.
+    """
+    text = _extract_pdf_text(pdf_b64)
+    if not text:
+        return IndexResult(summary="Could not extract text", tokens_used=0)
+    return await summarize_text(content=text, ext=".pdf", name=name)
+
+
+async def summarize_docx(*, docx_b64: str, name: str) -> IndexResult:
+    """Return a compact summary of a DOCX file.
+
+    Parameters
+    ----------
+    docx_b64:
+        Base64-encoded DOCX bytes.
+    name:
+        File name, e.g. ``"spec.docx"``.
+    """
+    text = _extract_docx_text(docx_b64)
+    if not text:
+        return IndexResult(summary="Could not extract text", tokens_used=0)
+    return await summarize_text(content=text, ext=".docx", name=name)

app/core/langfuse_client.py

@@ -0,0 +1,190 @@
+"""Langfuse observability — singleton client and prompt helpers.
+
+If LANGFUSE_SECRET_KEY / LANGFUSE_PUBLIC_KEY are not set,
+all helpers are no-ops so the app works without Langfuse configured.
+
+Usage
+-----
+Tracing::
+
+    from app.core.langfuse_client import get_langfuse
+
+    lf = get_langfuse()
+    if lf:
+        with lf.start_as_current_observation(as_type="span", name="my-agent") as span:
+            span.update(input=user_message)
+            # ... do work ...
+            span.update(output=result)
+        lf.flush()
+
+Prompt management::
+
+    from app.core.langfuse_client import get_prompt_or_fallback
+
+    text, prompt_obj = get_prompt_or_fallback("home_system", FALLBACK_PROMPT)
+    # Use text as the system prompt; pass prompt_obj to generations for linking.
+
+Linking a prompt to a generation::
+
+    with lf.start_as_current_observation(
+        as_type="generation",
+        name="llm-call",
+        model="gpt-4o",
+        prompt=prompt_obj,   # links generation → prompt version in the UI
+        input=messages,
+    ) as gen:
+        response = await llm.ainvoke(messages)
+        gen.update(output=response.content, usage=_usage(response))
+"""
+
+from __future__ import annotations
+
+import hashlib
+import logging
+from contextlib import contextmanager
+from typing import Any, Generator
+
+logger = logging.getLogger(__name__)
+
+_client: Any = None
+_initialized: bool = False
+
+
+def get_langfuse() -> Any | None:
+    """Return the Langfuse singleton, or ``None`` when not configured."""
+    global _client, _initialized
+    if _initialized:
+        return _client
+    _initialized = True
+
+    from app.config.settings import settings  # local import to avoid circular deps
+
+    if not settings.LANGFUSE_SECRET_KEY or not settings.LANGFUSE_PUBLIC_KEY:
+        logger.debug("langfuse: not configured — observability disabled")
+        return None
+
+    try:
+        from langfuse import Langfuse
+
+        _client = Langfuse(
+            secret_key=settings.LANGFUSE_SECRET_KEY,
+            public_key=settings.LANGFUSE_PUBLIC_KEY,
+            host=settings.LANGFUSE_BASE_URL,
+        )
+        logger.info("langfuse: client initialized host=%s", settings.LANGFUSE_BASE_URL)
+    except Exception as exc:
+        logger.warning("langfuse: failed to initialize: %s", exc)
+        _client = None
+
+    return _client
+
+
+def get_prompt_or_fallback(name: str, fallback: str) -> tuple[str, Any]:
+    """Fetch a text prompt from Langfuse; fall back to ``fallback`` on any error.
+
+    Returns ``(raw_template, prompt_obj_or_None)``.
+
+    * ``raw_template`` — the uncompiled template string.  Do NOT call ``.format()``
+      on it directly; use :func:`compile_prompt` instead so the correct variable
+      syntax is applied (``{{var}}`` for Langfuse, ``{var}`` for the fallback).
+    * ``prompt_obj`` — the Langfuse prompt object, or ``None`` when Langfuse is
+      unavailable / the fetch failed.  Pass this to generation observations so
+      Langfuse links the generation to the exact prompt version in the UI.
+    """
+    lf = get_langfuse()
+    if lf is None:
+        return fallback, None
+
+    try:
+        prompt = lf.get_prompt(name, label="production", fallback=fallback)
+        # For text-type prompts .prompt holds the raw template string.
+        raw = prompt.prompt if hasattr(prompt, "prompt") and isinstance(prompt.prompt, str) else fallback
+        return raw, prompt
+    except Exception as exc:
+        logger.warning("langfuse: get_prompt %r failed: %s — using fallback", name, exc)
+        return fallback, None
+
+
+def compile_prompt(template: str, prompt_obj: Any, **variables: Any) -> str:
+    """Compile *template* with *variables*, choosing the right syntax.
+
+    * When *prompt_obj* is a real Langfuse prompt object, calls
+      ``prompt_obj.compile(**variables)`` which handles ``{{variable}}``
+      substitution as defined in the Langfuse UI.
+    * When *prompt_obj* is ``None`` (Langfuse unavailable or fetch failed),
+      falls back to ``template.format(**variables)`` which handles the
+      ``{variable}`` syntax used in the hardcoded fallback strings.
+
+    This keeps callers oblivious to which syntax is in use.
+    """
+    if prompt_obj is not None:
+        try:
+            compiled = prompt_obj.compile(**variables)
+            # compile() returns a string for text prompts.
+            if isinstance(compiled, str):
+                return compiled
+            # Chat prompts return a list of dicts — join text parts.
+            if isinstance(compiled, list):
+                return "\n".join(
+                    m.get("content", "") for m in compiled if isinstance(m, dict)
+                )
+        except Exception as exc:
+            logger.warning(
+                "langfuse: compile failed for prompt %r: %s — falling back to .format()",
+                getattr(prompt_obj, "name", "?"),
+                exc,
+            )
+    return template.format(**variables)
+
+
+def extract_usage(response: Any) -> dict[str, int]:
+    """Extract token usage from a LangChain AI message into Langfuse format."""
+    meta = getattr(response, "usage_metadata", None)
+    if not meta:
+        return {}
+    return {
+        "input": int(meta.get("input_tokens", 0)),
+        "output": int(meta.get("output_tokens", 0)),
+        "total": int(meta.get("total_tokens", 0)),
+    }
+
+
+def hash_user_id(user_id: str) -> str:
+    """Return a SHA-256 hash of *user_id* for use as Langfuse ``user_id``.
+
+    This avoids sending raw database UUIDs to external observability services
+    while still providing a stable, deterministic identifier for per-user
+    metrics in the Langfuse dashboard.
+    """
+    return hashlib.sha256(user_id.encode()).hexdigest()
+
+
+@contextmanager
+def langfuse_context(
+    user_id: str | None = None,
+    session_id: str | None = None,
+) -> Generator[None, None, None]:
+    """Propagate ``user_id`` (hashed) and ``session_id`` to all Langfuse observations.
+
+    No-op when Langfuse is not configured or parameters are empty.
+    """
+    lf = get_langfuse()
+    if lf is None or (not user_id and not session_id):
+        yield
+        return
+
+    try:
+        from langfuse import propagate_attributes
+    except ImportError:
+        logger.debug("langfuse: propagate_attributes not available — skipping context")
+        yield
+        return
+
+    attrs: dict[str, str] = {}
+    if user_id:
+        attrs["user_id"] = hash_user_id(user_id)
+    if session_id:
+        attrs["session_id"] = session_id
+
+    with propagate_attributes(**attrs):
+        yield

app/core/llm.py

@@ -0,0 +1,156 @@
+"""LLM factory — centralised model instantiation via LiteLLM.
+
+Every agent and the orchestrator call ``get_llm()``
+instead of directly constructing a provider-specific class.  The model string
+follows the `LiteLLM model naming convention
+<https://docs.litellm.ai/docs/providers>`_:
+
+* OpenAI:     ``gpt-4o``, ``gpt-4o-mini``
+* Anthropic:  ``anthropic/claude-3.5-sonnet``
+* Google:     ``gemini/gemini-pro``
+* Ollama:     ``ollama/llama3``
+* Bedrock:    ``bedrock/anthropic.claude-v2``
+
+Switch providers by changing **LLM_MODEL** in ``.env``
+— no code changes required.
+"""
+
+from __future__ import annotations
+
+import os
+import warnings
+from collections.abc import Callable
+
+from openai import AsyncOpenAI
+import litellm
+
+from langchain_openai import ChatOpenAI
+from langchain_litellm import ChatLiteLLM
+from litellm import get_supported_openai_params  # noqa: F401 – validates install
+
+from app.config.settings import settings
+
+# Some models (e.g. gpt-5, o-series) reject unsupported params like temperature.
+# Drop them silently instead of raising UnsupportedParamsError.
+litellm.drop_params = True
+
+# Some provider responses include a plain dict in the `usage` field where a
+# richer Pydantic model is expected. This warning is noisy but non-fatal.
+warnings.filterwarnings(
+    "ignore",
+    message=r"PydanticSerializationUnexpectedValue\(Expected `ResponseAPIUsage`",
+    category=UserWarning,
+)
+
+
+def _api_key_for_model(model: str) -> str | None:
+    """Return the most appropriate API key for the given LiteLLM model string."""
+    if model.startswith("anthropic/"):
+        return settings.ANTHROPIC_API_KEY or None
+    if model.startswith("gemini/") or model.startswith("google/"):
+        return settings.GOOGLE_API_KEY or None
+    if model.startswith("cerebras/"):
+        return settings.CEREBRAS_API_KEY or None
+    if model.startswith("groq/"):
+        return settings.GROQ_API_KEY or None
+    if model.startswith("deepseek/"):
+        return settings.DEEPSEEK_API_KEY or None
+    if model.startswith("github_copilot/"):
+        # GitHub Copilot uses OAuth device-flow tokens managed by LiteLLM.
+        # No API key is required; returning None lets LiteLLM handle auth.
+        return None
+    # Default: OpenAI-compatible (covers plain model names like "gpt-4o")
+    return settings.OPENAI_API_KEY or None
+
+
+def get_llm(
+    *,
+    model: str | None = None,
+    temperature: float = 0,
+) -> ChatOpenAI | ChatLiteLLM:
+    """Return a LangChain chat model backed by LiteLLM.
+
+    LiteLLM exposes an OpenAI-compatible API, so we use ``ChatOpenAI`` pointed
+    at the LiteLLM proxy endpoint.  In practice, ``litellm`` patches the
+    ``openai`` client transparently when the model string contains a provider
+    prefix (``anthropic/…``, ``gemini/…``, etc.).
+
+    Parameters
+    ----------
+    model:
+        LiteLLM model identifier. Defaults to ``settings.LLM_MODEL``.
+    temperature:
+        Sampling temperature.  ``0`` = deterministic.
+    """
+    model = model or settings.LLM_MODEL
+
+    # Point LiteLLM to the custom token directory when configured.
+    if settings.GITHUB_COPILOT_TOKEN_DIR:
+        os.environ.setdefault("GITHUB_COPILOT_TOKEN_DIR", settings.GITHUB_COPILOT_TOKEN_DIR)
+
+    # Use ChatLiteLLM for provider-prefixed models (github_copilot/, anthropic/, etc.)
+    # so LiteLLM handles routing and auth. ChatOpenAI for plain OpenAI model names.
+    if "/" in model:
+        return ChatLiteLLM(model=model, temperature=temperature)
+
+    return ChatOpenAI(
+        model=model,
+        temperature=temperature,
+        api_key=_api_key_for_model(model),
+    )
+
+
+_AGENT_MODEL_SETTINGS: dict[str, Callable[[], str]] = {
+    "classifier":          lambda: settings.LLM_MODEL_CLASSIFIER or settings.LLM_MODEL,
+    "home-agent":          lambda: settings.LLM_MODEL_HOME_AGENT or settings.LLM_MODEL,
+    "unified-processor":   lambda: settings.LLM_MODEL_UNIFIED_PROCESSOR or settings.LLM_MODEL,
+    "cloud-processor":     lambda: settings.LLM_MODEL_CLOUD_PROCESSOR or settings.LLM_MODEL,
+    "brief-agent":         lambda: settings.LLM_MODEL_BRIEF_AGENT or settings.LLM_MODEL,
+    "task-brief-agent":    lambda: settings.LLM_MODEL_TASK_BRIEF_AGENT or settings.LLM_MODEL,
+    "setup":               lambda: settings.LLM_MODEL_SETUP_AGENT or settings.LLM_MODEL,
+    "memory-extractor":    lambda: settings.LLM_MODEL_MEMORY_EXTRACTOR or "gpt-4o-mini",
+    "memory-miner":        lambda: settings.LLM_MODEL_MEMORY_MINER or "gpt-4o-mini",
+    "memory-auditor":      lambda: settings.LLM_MODEL_MEMORY_AUDITOR or settings.LLM_MODEL,
+    "note-summarizer":     lambda: "gpt-4o-mini",
+}
+
+
+def model_for_agent(agent_name: str) -> str:
+    """Return the resolved model string for *agent_name* (for Langfuse tracking)."""
+    return _AGENT_MODEL_SETTINGS.get(agent_name, lambda: settings.LLM_MODEL)()
+
+
+def get_agent_llm(
+    agent_name: str,
+    *,
+    temperature: float = 0,
+) -> ChatOpenAI | ChatLiteLLM:
+    """Return an LLM configured for *agent_name*, respecting per-agent overrides.
+
+    Falls back to ``settings.LLM_MODEL`` for unknown agent names or when the
+    per-agent override is left empty in ``.env``.
+    """
+    model = model_for_agent(agent_name)
+    return get_llm(model=model, temperature=temperature)
+
+
+async def embed(text: str) -> list[float]:
+    """Return an embedding vector for *text*.
+
+    Uses ``settings.LLM_EMBED_MODEL`` so the same provider switch in ``.env``
+    (e.g. ``github_copilot/text-embedding-3-small``) applies here without any
+    code changes.  Falls back to the raw AsyncOpenAI client for plain OpenAI
+    model names to preserve existing behaviour.
+    """
+    model = settings.LLM_EMBED_MODEL
+
+    if model.startswith("github_copilot/") or "/" in model:
+        # Use LiteLLM for all provider-prefixed models (Copilot, Bedrock, etc.)
+        # so the provider's auth mechanism is applied correctly.
+        response = await litellm.aembedding(model=model, input=[text])
+        return response.data[0]["embedding"]
+
+    # Plain OpenAI model name — use the raw AsyncOpenAI client (existing path).
+    client = AsyncOpenAI(api_key=settings.OPENAI_API_KEY)
+    response = await client.embeddings.create(model=model, input=text)
+    return response.data[0].embedding

app/core/memory_extraction.py

@@ -0,0 +1,450 @@
+"""Mem0-style Extract/Update pipeline — Phase 2.
+
+Runs after every ``store_episode`` call to distil durable facts, preferences,
+routines, and relations from the latest conversation turn.
+
+Entry point: ``run_extraction(db, user_id, last_user_msg, last_assistant_msg, session_id)``
+
+Design notes
+------------
+- Two gpt-4o-mini calls per turn: extract candidates, then decide action per candidate.
+- Short-circuit: if no existing neighbours → ADD without a second LLM call (cost saving).
+- Zero-trust: never logs decrypted user content; relation subject/object labels are
+  treated as identifiers (safe to log per spec).
+- Must not raise into the request path — caller wraps in asyncio.create_task().
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.core.langfuse_client import get_langfuse, get_prompt_or_fallback, extract_usage, langfuse_context
+from app.core.llm import get_agent_llm, model_for_agent
+
+logger = logging.getLogger(__name__)
+
+# ── Fallback prompts (used when Langfuse unavailable) ─────────────────────────
+
+_EXTRACTION_FALLBACK = (
+    "You are a memory extractor for a personal AI secretary. Given the last conversation "
+    "turn, the user's core memory, and recent episode summaries, identify durable facts, "
+    "preferences, routines, and person/project relations worth remembering.\n\n"
+    "Output JSON matching this schema exactly:\n"
+    '{{"candidates": [{{"type": "<fact|preference|relation|routine>", '
+    '"content": "<short canonical statement>", '
+    '"target_tier": "<core|associative|relational|proactive>", '
+    '"subject": null, "predicate": null, "object": null, "confidence": 0.7}}]}}\n\n'
+    "Rules:\n"
+    "- Skip small talk, greetings, one-off questions.\n"
+    "- Max 5 candidates per call.\n"
+    "- Only extract durable information (still true next week).\n"
+    "- For type=relation: subject/predicate/object required.\n"
+    "- Default confidence=0.7.\n\n"
+    "## Last turn\n{last_turn}\n\n"
+    "## Core memory (current)\n{core_memory}\n\n"
+    "## Recent episodes\n{recent_episodes}"
+)
+
+_DECIDE_FALLBACK = (
+    "You are a memory update decision engine. Given a new memory candidate and a list of "
+    "existing memories from the same tier, decide what action to take.\n\n"
+    "Respond with exactly one word: ADD, UPDATE, DELETE, or NOOP.\n\n"
+    "- ADD: new information not in existing memories.\n"
+    "- UPDATE: contradicts or supersedes an existing memory.\n"
+    "- DELETE: states something is no longer true.\n"
+    "- NOOP: already captured accurately.\n\n"
+    "## New candidate\n{candidate}\n\n"
+    "## Existing memories (same tier, top neighbours)\n{existing_memories}"
+)
+
+
+# ── Pydantic schemas ───────────────────────────────────────────────────────────
+
+class MemoryCandidate(BaseModel):
+    type: Literal["fact", "preference", "relation", "routine"]
+    content: str
+    target_tier: Literal["core", "associative", "relational", "proactive"]
+    subject: str | None = None
+    predicate: str | None = None
+    object: str | None = None
+    confidence: float = Field(default=0.7, ge=0.0, le=1.0)
+
+
+class ExtractionResult(BaseModel):
+    candidates: list[MemoryCandidate] = Field(default_factory=list)
+
+
+# ── Task 2.1 — Extract candidates ─────────────────────────────────────────────
+
+async def extract_candidates(
+    last_turn: str,
+    core_memory: dict[str, str],
+    recent_episodes: list[str],
+) -> ExtractionResult:
+    """Call gpt-4o-mini to extract memory candidates from the latest turn.
+
+    Returns an ExtractionResult (may be empty on failure — never raises).
+    """
+    core_str = "\n".join(f"{k}: {v}" for k, v in core_memory.items()) or "(empty)"
+    episodes_str = "\n---\n".join(recent_episodes[-5:]) or "(none)"
+
+    template, prompt_obj = get_prompt_or_fallback("memory_extraction", _EXTRACTION_FALLBACK)
+
+    # Compile with Langfuse variable syntax ({{var}}) or fallback {var}
+    if prompt_obj is not None:
+        try:
+            system_text = prompt_obj.compile(
+                last_turn=last_turn,
+                core_memory=core_str,
+                recent_episodes=episodes_str,
+            )
+            if isinstance(system_text, list):
+                system_text = "\n".join(m.get("content", "") for m in system_text if isinstance(m, dict))
+        except Exception as exc:
+            logger.warning("memory_extraction: compile failed: %s", exc)
+            system_text = template.format(
+                last_turn=last_turn,
+                core_memory=core_str,
+                recent_episodes=episodes_str,
+            )
+    else:
+        system_text = template.format(
+            last_turn=last_turn,
+            core_memory=core_str,
+            recent_episodes=episodes_str,
+        )
+
+    llm = get_agent_llm("memory-extractor", temperature=0)
+    # Bind JSON mode so the model always returns parseable output.
+    llm_json = llm.bind(response_format={"type": "json_object"})  # type: ignore[attr-defined]
+
+    lf = get_langfuse()
+    try:
+        from langchain_core.messages import HumanMessage, SystemMessage  # noqa: PLC0415
+        messages = [
+            SystemMessage(content=system_text),
+            HumanMessage(content="Extract memory candidates as JSON."),
+        ]
+
+        if lf:
+            with lf.start_as_current_observation(
+                as_type="generation",
+                name="memory-extraction",
+                model=model_for_agent("memory-extractor"),
+                prompt=prompt_obj,
+                input=messages,
+            ) as gen:
+                response = await llm_json.ainvoke(messages)
+                gen.update(output=response.content, usage=extract_usage(response))
+        else:
+            response = await llm_json.ainvoke(messages)
+
+        raw = json.loads(response.content)
+        result = ExtractionResult.model_validate(raw)
+        logger.info("memory_extraction: extracted %d candidates", len(result.candidates))
+        return result
+
+    except Exception as exc:
+        logger.warning("memory_extraction: extract_candidates failed: %s", exc)
+        return ExtractionResult(candidates=[])
+
+
+# ── Task 2.2 — Decide action ──────────────────────────────────────────────────
+
+async def decide_action(
+    candidate: MemoryCandidate,
+    existing: list[str],
+) -> Literal["ADD", "UPDATE", "DELETE", "NOOP"]:
+    """Decide what to do with a candidate given existing memories in the same tier.
+
+    Short-circuits to ADD without an LLM call when existing is empty (cost saving).
+    Never raises.
+    """
+    if not existing:
+        return "ADD"
+
+    candidate_str = f"[{candidate.type}] {candidate.content}"
+    existing_str = "\n".join(f"- {m}" for m in existing)
+
+    template, prompt_obj = get_prompt_or_fallback("memory_decide_action", _DECIDE_FALLBACK)
+
+    if prompt_obj is not None:
+        try:
+            system_text = prompt_obj.compile(
+                candidate=candidate_str,
+                existing_memories=existing_str,
+            )
+            if isinstance(system_text, list):
+                system_text = "\n".join(m.get("content", "") for m in system_text if isinstance(m, dict))
+        except Exception as exc:
+            logger.warning("memory_extraction: decide compile failed: %s", exc)
+            system_text = template.format(candidate=candidate_str, existing_memories=existing_str)
+    else:
+        system_text = template.format(candidate=candidate_str, existing_memories=existing_str)
+
+    llm = get_agent_llm("memory-extractor", temperature=0)
+    lf = get_langfuse()
+
+    try:
+        from langchain_core.messages import HumanMessage, SystemMessage  # noqa: PLC0415
+        messages = [
+            SystemMessage(content=system_text),
+            HumanMessage(content="Decide action."),
+        ]
+
+        if lf:
+            with lf.start_as_current_observation(
+                as_type="generation",
+                name="memory-decide-action",
+                model=model_for_agent("memory-extractor"),
+                prompt=prompt_obj,
+                input=messages,
+            ) as gen:
+                response = await llm.ainvoke(messages)
+                gen.update(output=response.content, usage=extract_usage(response))
+        else:
+            response = await llm.ainvoke(messages)
+
+        verb = response.content.strip().upper()
+        if verb in ("ADD", "UPDATE", "DELETE", "NOOP"):
+            return verb  # type: ignore[return-value]
+        logger.warning("memory_extraction: unexpected decide verb=%r, defaulting ADD", verb)
+        return "ADD"
+
+    except Exception as exc:
+        logger.warning("memory_extraction: decide_action failed: %s", exc)
+        return "ADD"
+
+
+# ── Task 2.3 — Pipeline orchestrator ──────────────────────────────────────────
+
+async def run_extraction(
+    db: AsyncSession,
+    user_id: str,
+    last_user_msg: str,
+    last_assistant_msg: str,
+    session_id: str | None,
+) -> None:
+    """Full Mem0-style extract/update pipeline for one conversation turn.
+
+    Steps:
+    1. Load core memory + last 5 episodes.
+    2. extract_candidates() → up to 5 MemoryCandidate objects.
+    3. For each candidate: find top-3 neighbours → decide_action() → apply.
+    4. Trace via Langfuse.
+
+    Never raises — wraps everything in try/except.
+    """
+    try:
+        await _run_extraction_inner(db, user_id, last_user_msg, last_assistant_msg, session_id)
+    except Exception as exc:
+        logger.warning("memory_extraction: run_extraction failed user=%s: %s", user_id, exc)
+
+
+async def _run_extraction_inner(
+    db: AsyncSession,
+    user_id: str,
+    last_user_msg: str,
+    last_assistant_msg: str,
+    session_id: str | None,
+) -> None:
+    from app.core.memory_middleware import MemoryMiddleware  # noqa: PLC0415
+
+    middleware = MemoryMiddleware(db)
+    fernet = await middleware._get_fernet(user_id)
+    if fernet is None:
+        logger.warning("memory_extraction: no fernet for user=%s, skipping", user_id)
+        return
+
+    # 1. Load context
+    core: dict[str, str] = await middleware._load_core(user_id, fernet)
+    episodes: list[str] = await middleware._load_episodic(user_id, fernet, session_id=session_id)
+
+    last_turn = f"User: {last_user_msg}\nAssistant: {last_assistant_msg}"
+
+    lf = get_langfuse()
+
+    async def _run(trace_id: str | None) -> dict[str, Any]:
+        # 2. Extract candidates
+        result = await extract_candidates(last_turn, core, episodes)
+        if not result.candidates:
+            logger.info("memory_extraction: no candidates user=%s", user_id)
+            return {"candidates": 0, "applied": 0}
+
+        logger.info(
+            "memory_extraction: processing %d candidates user=%s trace=%s",
+            len(result.candidates),
+            user_id,
+            trace_id or "-",
+        )
+
+        # 3. Apply each candidate
+        applied = 0
+        actions: list[str] = []
+        for candidate in result.candidates:
+            try:
+                await _apply_candidate(middleware, db, user_id, fernet, candidate, trace_id)
+                applied += 1
+                actions.append(f"{candidate.type}:{candidate.target_tier}")
+            except Exception as exc:
+                logger.warning(
+                    "memory_extraction: apply failed candidate=%r user=%s: %s",
+                    candidate.content[:80],
+                    user_id,
+                    exc,
+                )
+
+        logger.info(
+            "memory_extraction: applied %d/%d candidates user=%s",
+            applied,
+            len(result.candidates),
+            user_id,
+        )
+        return {"candidates": len(result.candidates), "applied": applied, "actions": actions}
+
+    with langfuse_context(user_id=user_id, session_id=session_id):
+        if lf:
+            with lf.start_as_current_observation(
+                as_type="span",
+                name="memory-extraction-pipeline",
+                input={"last_turn_preview": last_turn[:200]},
+            ) as span:
+                summary = await _run(trace_id=span.id)
+                span.update(output=summary)
+            try:
+                lf.flush()
+            except Exception:
+                pass
+        else:
+            await _run(trace_id=None)
+
+
+async def _apply_candidate(
+    middleware: Any,
+    db: AsyncSession,
+    user_id: str,
+    fernet: Any,
+    candidate: MemoryCandidate,
+    trace_id: str | None,
+) -> None:
+    """Fetch neighbours, decide action, apply to the appropriate tier."""
+
+    neighbours: list[str] = []
+
+    if candidate.target_tier == "core":
+        # For core tier: neighbours are existing core block values for similar keys.
+        blocks = await middleware.list_core_blocks(user_id)
+        neighbours = [b["value"] for b in blocks[:3]]
+
+    elif candidate.target_tier == "associative":
+        neighbours = await middleware.search_archival(user_id, candidate.content, top_k=3)
+
+    elif candidate.target_tier == "relational":
+        # Relation candidates handled specially — passed to upsert_relation directly.
+        # Neighbours: search by subject label if available.
+        neighbours = []
+
+    elif candidate.target_tier == "proactive":
+        neighbours = await middleware.search_recall(user_id, candidate.content, top_k=3)
+
+    action = await decide_action(candidate, neighbours)
+    logger.info(
+        "memory_extraction: candidate type=%s tier=%s action=%s",
+        candidate.type,
+        candidate.target_tier,
+        action,
+    )
+
+    if action == "NOOP":
+        return
+
+    if candidate.target_tier == "relational":
+        # Always upsert relations — decide_action skipped (no neighbour search).
+        if candidate.subject and candidate.predicate and candidate.object:
+            await _upsert_relation(
+                middleware, db, user_id, candidate, trace_id
+            )
+        return
+
+    if action in ("ADD", "UPDATE"):
+        if candidate.target_tier == "core":
+            # Derive a short key from the content (first 40 chars, snake_cased).
+            key = _content_to_key(candidate.content)
+            await middleware.update_core(user_id, key, candidate.content, trace_id=trace_id)
+
+        elif candidate.target_tier == "associative":
+            await middleware.store_associative(user_id, candidate.content)
+
+        elif candidate.target_tier == "proactive":
+            await _store_proactive_stub(middleware, db, user_id, candidate, fernet)
+
+    elif action == "DELETE":
+        if candidate.target_tier == "core":
+            key = _content_to_key(candidate.content)
+            await middleware.delete_core(user_id, key)
+
+
+def _content_to_key(content: str) -> str:
+    """Derive a short snake_case key from a content string (first 40 chars)."""
+    import re  # noqa: PLC0415
+    slug = re.sub(r"[^a-z0-9]+", "_", content[:40].lower()).strip("_")
+    return slug or "memory"
+
+
+async def _upsert_relation(
+    middleware: Any,
+    db: AsyncSession,
+    user_id: str,
+    candidate: MemoryCandidate,
+    trace_id: str | None,
+) -> None:
+    """Upsert a relation row via MemoryMiddleware.upsert_relation (Phase 3)."""
+    await middleware.upsert_relation(
+        user_id=user_id,
+        subject=candidate.subject or "unknown",
+        subject_type="unknown",
+        predicate=candidate.predicate or "related_to",
+        object_=candidate.object or "unknown",
+        object_type="unknown",
+        confidence=candidate.confidence,
+    )
+    logger.info(
+        "memory_extraction: upserted relation subject=%s predicate=%s object=%s",
+        candidate.subject,
+        candidate.predicate,
+        candidate.object,
+    )
+
+
+async def _store_proactive_stub(
+    middleware: Any,
+    db: AsyncSession,
+    user_id: str,
+    candidate: MemoryCandidate,
+    fernet: Any,
+) -> None:
+    """Store a proactive pattern row directly (MemoryProactive model)."""
+    import uuid  # noqa: PLC0415
+    from app.models import MemoryProactive  # noqa: PLC0415
+    from app.core.memory_middleware import _encrypt  # noqa: PLC0415
+
+    encrypted = _encrypt(fernet, candidate.content)
+    row = MemoryProactive(
+        id=str(uuid.uuid4()),
+        user_id=user_id,
+        pattern_encrypted=encrypted,
+        confidence=candidate.confidence,
+        source="inferred",
+    )
+    db.add(row)
+    try:
+        await db.commit()
+        logger.info("memory_extraction: stored proactive pattern user=%s", user_id)
+    except Exception as exc:
+        logger.warning("memory_extraction: store proactive failed: %s", exc)
+        await db.rollback()

app/core/memory_maintenance.py

@@ -0,0 +1,581 @@
+"""Memory maintenance jobs — Phase 3/5.
+
+Three entrypoints called by the scheduler (APScheduler) registered in app/main.py:
+
+  drain_extraction_queue(db) — Free-tier batch extraction (Phase 2/5).
+  mine_proactive_patterns(db, user_id) — Power+ pattern mining (Phase 5).
+  decay_relations(db, user_id) — confidence decay + pruning for memory_relations (Phase 3).
+
+All are safe to call manually or from tests; they never raise.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import uuid
+from datetime import datetime, timedelta, timezone
+
+from cryptography.fernet import Fernet
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.core.langfuse_client import compile_prompt, extract_usage, get_langfuse, get_prompt_or_fallback
+from app.models import MemoryAssociative, MemoryEpisodic, MemoryProactive, MemoryRelation, User
+
+logger = logging.getLogger(__name__)
+
+# Decay parameters for relations
+_DECAY_FACTOR = 0.95
+_DECAY_PERIOD_DAYS = 30
+_PRUNE_THRESHOLD = 0.2
+
+# Proactive pattern decay: 10 % per 7 days since last sighting
+_PROACTIVE_DECAY_FACTOR = 0.9
+_PROACTIVE_DECAY_PERIOD_DAYS = 7
+_PROACTIVE_PRUNE_THRESHOLD = 0.2
+
+# Mining: require at least this many episodes to attempt pattern extraction
+_MIN_EPISODES_FOR_MINING = 3
+_MINING_LOOKBACK_DAYS = 30
+
+# Audit: caps to control token cost
+_AUDIT_MAX_FACTS = 50
+_AUDIT_MAX_LABELS = 100
+
+
+async def decay_relations(db: AsyncSession, user_id: str) -> None:
+    """Apply confidence decay to all relation rows for a user.
+
+    Decay rule: confidence *= 0.95 for every 30 days since last_confirmed_at.
+    Rows whose confidence falls below 0.2 are deleted.
+
+    Never raises — wraps in try/except.
+    """
+    try:
+        await _decay_relations_inner(db, user_id)
+    except Exception as exc:
+        logger.warning("memory_maintenance: decay_relations failed user=%s: %s", user_id, exc)
+
+
+async def _decay_relations_inner(db: AsyncSession, user_id: str) -> None:
+    result = await db.execute(
+        select(MemoryRelation).where(MemoryRelation.user_id == user_id)
+    )
+    rows = result.scalars().all()
+    now = datetime.now(timezone.utc)
+    deleted = 0
+    decayed = 0
+
+    for row in rows:
+        reference = row.last_confirmed_at or row.created_at
+        if reference is None:
+            continue
+        if reference.tzinfo is None:
+            reference = reference.replace(tzinfo=timezone.utc)
+
+        days_elapsed = (now - reference).days
+        if days_elapsed < _DECAY_PERIOD_DAYS:
+            continue
+
+        periods = days_elapsed // _DECAY_PERIOD_DAYS
+        new_confidence = row.confidence * (_DECAY_FACTOR ** periods)
+
+        if new_confidence < _PRUNE_THRESHOLD:
+            await db.delete(row)
+            deleted += 1
+            logger.info(
+                "memory_maintenance: pruned relation id=%s user=%s subject=%s predicate=%s "
+                "confidence=%.3f (below threshold)",
+                row.id, user_id, row.subject_label, row.predicate, new_confidence,
+            )
+        else:
+            row.confidence = new_confidence
+            decayed += 1
+
+    try:
+        await db.commit()
+        logger.info(
+            "memory_maintenance: decay_relations user=%s decayed=%d deleted=%d",
+            user_id, decayed, deleted,
+        )
+    except Exception as exc:
+        logger.warning("memory_maintenance: decay_relations commit failed user=%s: %s", user_id, exc)
+        await db.rollback()
+
+
+async def drain_extraction_queue(db: AsyncSession) -> None:
+    """Process pending ExtractionQueue rows for Free-tier users.
+
+    Each row corresponds to a stored episode that should be fed through the
+    Mem0-style extraction pipeline. Rows are deleted after successful processing.
+    Never raises — wraps in try/except.
+    """
+    try:
+        await _drain_extraction_queue_inner(db)
+    except Exception as exc:
+        logger.warning("memory_maintenance: drain_extraction_queue failed: %s", exc)
+
+
+async def _drain_extraction_queue_inner(db: AsyncSession) -> None:
+    from app.models import ExtractionQueue  # noqa: PLC0415
+
+    result = await db.execute(select(ExtractionQueue))
+    rows = result.scalars().all()
+
+    if not rows:
+        logger.debug("memory_maintenance: drain_extraction_queue nothing to drain")
+        return
+
+    logger.info("memory_maintenance: drain_extraction_queue pending=%d", len(rows))
+
+    from app.core.memory_extraction import run_extraction  # noqa: PLC0415
+
+    processed = 0
+    for row in rows:
+        try:
+            await run_extraction(
+                db=db,
+                user_id=row.user_id,
+                last_user_msg="",
+                last_assistant_msg="",
+                session_id=None,
+            )
+            await db.delete(row)
+            await db.commit()
+            processed += 1
+        except Exception as exc:
+            logger.warning(
+                "memory_maintenance: drain failed row=%s user=%s: %s",
+                row.id, row.user_id, exc,
+            )
+            await db.rollback()
+
+    logger.info("memory_maintenance: drain_extraction_queue processed=%d/%d", processed, len(rows))
+
+
+async def mine_proactive_patterns(db: AsyncSession, user_id: str) -> None:
+    """Mine recurring behavioral patterns from last 30 days of episodes (Power+ only).
+
+    Steps:
+    1. Gate on proactive_mining tier feature.
+    2. Load + decrypt last 30 days of episodic summaries.
+    3. Call gpt-4o-mini to identify recurring patterns.
+    4. Encrypt and store each pattern in memory_proactive.
+    5. Apply decay to existing proactive rows.
+
+    Never raises — wraps in try/except.
+    """
+    try:
+        await _mine_proactive_patterns_inner(db, user_id)
+    except Exception as exc:
+        logger.warning("memory_maintenance: mine_proactive_patterns failed user=%s: %s", user_id, exc)
+
+
+async def _mine_proactive_patterns_inner(db: AsyncSession, user_id: str) -> None:
+    from app.billing.tier_manager import tier_manager  # noqa: PLC0415
+
+    tier = await tier_manager.get_tier(user_id, db)
+    if not tier_manager.check_feature(tier, "proactive_mining"):
+        logger.debug("memory_maintenance: mine_proactive_patterns skipped (tier=%s)", tier)
+        return
+
+    # Load user Fernet key
+    result = await db.execute(select(User).where(User.id == user_id))
+    user = result.scalar_one_or_none()
+    if user is None or not user.encryption_key:
+        logger.warning("memory_maintenance: mine_proactive_patterns no encryption_key user=%s", user_id)
+        return
+
+    fernet = Fernet(user.encryption_key.encode())
+    cutoff = datetime.now(timezone.utc) - timedelta(days=_MINING_LOOKBACK_DAYS)
+
+    episodes_result = await db.execute(
+        select(MemoryEpisodic)
+        .where(
+            MemoryEpisodic.user_id == user_id,
+            MemoryEpisodic.created_at >= cutoff,
+        )
+        .order_by(MemoryEpisodic.created_at.asc())
+    )
+    episode_rows = episodes_result.scalars().all()
+
+    if len(episode_rows) < _MIN_EPISODES_FOR_MINING:
+        logger.info(
+            "memory_maintenance: mine_proactive_patterns skipped user=%s episodes=%d (< %d)",
+            user_id, len(episode_rows), _MIN_EPISODES_FOR_MINING,
+        )
+        return
+
+    summaries: list[str] = []
+    for ep in episode_rows:
+        try:
+            plaintext = fernet.decrypt(ep.summary_encrypted.encode()).decode()
+            summaries.append(plaintext)
+        except Exception:
+            pass
+
+    if not summaries:
+        return
+
+    patterns = await _extract_proactive_patterns(summaries)
+    if not patterns:
+        logger.info("memory_maintenance: mine_proactive_patterns user=%s no patterns extracted", user_id)
+        return
+
+    stored = 0
+    for pattern_text in patterns:
+        try:
+            encrypted = fernet.encrypt(pattern_text.encode()).decode()
+            row = MemoryProactive(
+                id=str(uuid.uuid4()),
+                user_id=user_id,
+                pattern_encrypted=encrypted,
+                confidence=0.7,
+                source="inferred",
+            )
+            db.add(row)
+            stored += 1
+        except Exception as exc:
+            logger.warning("memory_maintenance: failed to store pattern user=%s: %s", user_id, exc)
+
+    try:
+        await db.commit()
+        logger.info(
+            "memory_maintenance: mine_proactive_patterns user=%s stored=%d",
+            user_id, stored,
+        )
+    except Exception as exc:
+        logger.warning("memory_maintenance: mine_proactive_patterns commit failed user=%s: %s", user_id, exc)
+        await db.rollback()
+        return
+
+    await _decay_proactive_patterns(db, user_id, fernet)
+
+
+async def _extract_proactive_patterns(summaries: list[str]) -> list[str]:
+    """Call memory-miner LLM to identify recurring behavioral/temporal patterns."""
+    from app.core.llm import get_agent_llm  # noqa: PLC0415
+
+    llm = get_agent_llm("memory-miner", temperature=0)
+    combined = "\n---\n".join(summaries[-20:])  # cap at last 20 to control token usage
+    prompt = (
+        "You are analyzing conversation history for a personal AI secretary. "
+        "Identify 3-5 recurring temporal or behavioral patterns (e.g. 'always works late on Thursdays', "
+        "'prefers bullet-point summaries', 'frequently asks about Project Acme status'). "
+        "Return each pattern as a plain, short English sentence on its own line. "
+        "No numbering, no bullet points, no extra text.\n\n"
+        f"Conversation history:\n{combined}"
+    )
+    try:
+        response = await llm.ainvoke(prompt)
+        text = response.content if hasattr(response, "content") else str(response)
+        lines = [line.strip() for line in str(text).splitlines() if line.strip()]
+        return lines[:5]
+    except Exception as exc:
+        logger.warning("memory_maintenance: _extract_proactive_patterns LLM failed: %s", exc)
+        return []
+
+
+async def _decay_proactive_patterns(db: AsyncSession, user_id: str, fernet: Fernet) -> None:
+    """Decay confidence of existing proactive patterns; prune below threshold."""
+    result = await db.execute(
+        select(MemoryProactive).where(MemoryProactive.user_id == user_id)
+    )
+    rows = result.scalars().all()
+    now = datetime.now(timezone.utc)
+    deleted = 0
+    decayed = 0
+
+    for row in rows:
+        reference = row.created_at
+        if reference is None:
+            continue
+        if reference.tzinfo is None:
+            reference = reference.replace(tzinfo=timezone.utc)
+
+        days_elapsed = (now - reference).days
+        if days_elapsed < _PROACTIVE_DECAY_PERIOD_DAYS:
+            continue
+
+        periods = days_elapsed // _PROACTIVE_DECAY_PERIOD_DAYS
+        new_confidence = row.confidence * (_PROACTIVE_DECAY_FACTOR ** periods)
+
+        if new_confidence < _PROACTIVE_PRUNE_THRESHOLD:
+            await db.delete(row)
+            deleted += 1
+        else:
+            row.confidence = new_confidence
+            decayed += 1
+
+    try:
+        await db.commit()
+        logger.info(
+            "memory_maintenance: decay_proactive user=%s decayed=%d deleted=%d",
+            user_id, decayed, deleted,
+        )
+    except Exception as exc:
+        logger.warning("memory_maintenance: decay_proactive commit failed user=%s: %s", user_id, exc)
+        await db.rollback()
+
+
+# ── Phase 7: weekly memory audit ──────────────────────────────────────────────
+
+_AUDIT_CONTRADICTIONS_FALLBACK = (
+    "You are auditing a personal AI assistant's memory bank. "
+    "Each fact has an ID in brackets. "
+    "Find pairs that directly contradict each other "
+    "(e.g. 'prefers morning meetings' vs 'never schedules before noon'). "
+    "For each contradiction, pick the ID to DELETE (the older or less specific one). "
+    'Return ONLY a valid JSON array, no markdown fences: '
+    '[{{"delete": "<id>", "reason": "<one line>"}}]. '
+    "If no contradictions, return [].\n\n"
+    "Facts:\n{facts}"
+)
+
+_AUDIT_CANONICALIZE_FALLBACK = (
+    "You are auditing entity labels in a personal AI assistant's relational memory. "
+    "These are names of people, companies, projects, or topics. "
+    "Group labels that clearly refer to the same real-world entity "
+    "(e.g. 'giulia', 'Giulia', 'Giulia R.' → canonical 'Giulia'). "
+    "Return ONLY a valid JSON array, no markdown fences: "
+    '[{{"canonical": "<best label>", "variants": ["<v1>", "<v2>"]}}]. '
+    "Only include groups with at least one variant. Singletons: omit.\n\n"
+    "Labels:\n{labels}"
+)
+
+
+async def audit_memory(db: AsyncSession, user_id: str) -> None:
+    """Weekly audit: contradiction scan on associative facts + label canonicalization on relations.
+
+    Steps:
+    1. Decrypt up to _AUDIT_MAX_FACTS associative rows; send list to memory-auditor LLM.
+    2. LLM flags rows to delete (direct contradictions); hard-delete them.
+    3. Collect unique subject/object labels from memory_relations; ask LLM to group duplicates.
+    4. Rewrite variant labels to their canonical form in-place.
+
+    Never raises — wraps in try/except.
+    """
+    try:
+        await _audit_memory_inner(db, user_id)
+    except Exception as exc:
+        logger.warning("memory_maintenance: audit_memory failed user=%s: %s", user_id, exc)
+
+
+async def _audit_memory_inner(db: AsyncSession, user_id: str) -> None:
+    result = await db.execute(select(User).where(User.id == user_id))
+    user = result.scalar_one_or_none()
+    if user is None or not user.encryption_key:
+        logger.warning("memory_maintenance: audit_memory no encryption_key user=%s", user_id)
+        return
+
+    fernet = Fernet(user.encryption_key.encode())
+    await _scan_associative_contradictions(db, user_id, fernet)
+    await _canonicalize_relation_labels(db, user_id)
+
+
+async def _scan_associative_contradictions(
+    db: AsyncSession,
+    user_id: str,
+    fernet: Fernet,
+) -> None:
+    """Decrypt associative facts, ask LLM to flag contradictions, delete superseded rows."""
+    result = await db.execute(
+        select(MemoryAssociative)
+        .where(MemoryAssociative.user_id == user_id)
+        .order_by(MemoryAssociative.updated_at.desc())
+        .limit(_AUDIT_MAX_FACTS)
+    )
+    rows = result.scalars().all()
+    if len(rows) < 2:
+        return
+
+    id_to_text: dict[str, str] = {}
+    for row in rows:
+        try:
+            plaintext = fernet.decrypt(row.content_encrypted.encode()).decode()
+            id_to_text[row.id] = plaintext
+        except Exception:
+            pass
+
+    if len(id_to_text) < 2:
+        return
+
+    id_list = list(id_to_text.keys())
+    numbered = "\n".join(
+        f"{i + 1}. [{rid}] {id_to_text[rid]}" for i, rid in enumerate(id_list)
+    )
+
+    template, prompt_obj = get_prompt_or_fallback(
+        "memory_audit_contradictions", _AUDIT_CONTRADICTIONS_FALLBACK
+    )
+    system_text = compile_prompt(template, prompt_obj, facts=numbered)
+
+    from app.core.llm import get_agent_llm, model_for_agent  # noqa: PLC0415
+    from langchain_core.messages import HumanMessage, SystemMessage  # noqa: PLC0415
+
+    llm = get_agent_llm("memory-auditor", temperature=0)
+    lf = get_langfuse()
+    messages = [
+        SystemMessage(content=system_text),
+        HumanMessage(content="Audit facts for contradictions."),
+    ]
+    try:
+        if lf:
+            with lf.start_as_current_observation(
+                as_type="generation",
+                name="memory-audit-contradictions",
+                model=model_for_agent("memory-auditor"),
+                prompt=prompt_obj,
+                input=messages,
+            ) as gen:
+                response = await llm.ainvoke(messages)
+                gen.update(output=response.content, usage=extract_usage(response))
+        else:
+            response = await llm.ainvoke(messages)
+
+        text = response.content if hasattr(response, "content") else str(response)
+        deletions = json.loads(text.strip())
+        if not isinstance(deletions, list):
+            return
+    except Exception as exc:
+        logger.warning(
+            "memory_maintenance: _scan_associative_contradictions LLM/parse failed user=%s: %s",
+            user_id, exc,
+        )
+        return
+
+    deleted = 0
+    for item in deletions:
+        if not isinstance(item, dict):
+            continue
+        rid = item.get("delete")
+        if not rid or rid not in id_to_text:
+            continue
+        result2 = await db.execute(
+            select(MemoryAssociative).where(
+                MemoryAssociative.id == rid,
+                MemoryAssociative.user_id == user_id,
+            )
+        )
+        target = result2.scalar_one_or_none()
+        if target:
+            await db.delete(target)
+            deleted += 1
+            logger.info(
+                "memory_maintenance: audit deleted contradiction id=%s user=%s reason=%s",
+                rid, user_id, item.get("reason", ""),
+            )
+
+    if deleted:
+        try:
+            await db.commit()
+        except Exception as exc:
+            logger.warning(
+                "memory_maintenance: audit contradiction commit failed user=%s: %s", user_id, exc
+            )
+            await db.rollback()
+
+    logger.info(
+        "memory_maintenance: _scan_associative_contradictions user=%s deleted=%d", user_id, deleted
+    )
+
+
+async def _canonicalize_relation_labels(db: AsyncSession, user_id: str) -> None:
+    """Group near-duplicate entity labels in memory_relations and unify to canonical form."""
+    result = await db.execute(
+        select(MemoryRelation).where(MemoryRelation.user_id == user_id)
+    )
+    rows = result.scalars().all()
+    if not rows:
+        return
+
+    all_labels: set[str] = set()
+    for row in rows:
+        all_labels.add(row.subject_label)
+        all_labels.add(row.object_label)
+
+    labels_list = sorted(all_labels)[:_AUDIT_MAX_LABELS]
+    if len(labels_list) < 2:
+        return
+
+    labels_block = "\n".join(f"- {lbl}" for lbl in labels_list)
+    template, prompt_obj = get_prompt_or_fallback(
+        "memory_audit_canonicalize", _AUDIT_CANONICALIZE_FALLBACK
+    )
+    system_text = compile_prompt(template, prompt_obj, labels=labels_block)
+
+    from app.core.llm import get_agent_llm, model_for_agent  # noqa: PLC0415
+    from langchain_core.messages import HumanMessage, SystemMessage  # noqa: PLC0415
+
+    llm = get_agent_llm("memory-auditor", temperature=0)
+    lf = get_langfuse()
+    messages = [
+        SystemMessage(content=system_text),
+        HumanMessage(content="Canonicalize entity labels."),
+    ]
+    try:
+        if lf:
+            with lf.start_as_current_observation(
+                as_type="generation",
+                name="memory-audit-canonicalize",
+                model=model_for_agent("memory-auditor"),
+                prompt=prompt_obj,
+                input=messages,
+            ) as gen:
+                response = await llm.ainvoke(messages)
+                gen.update(output=response.content, usage=extract_usage(response))
+        else:
+            response = await llm.ainvoke(messages)
+
+        text = response.content if hasattr(response, "content") else str(response)
+        groups = json.loads(text.strip())
+        if not isinstance(groups, list):
+            return
+    except Exception as exc:
+        logger.warning(
+            "memory_maintenance: _canonicalize_relation_labels LLM/parse failed user=%s: %s",
+            user_id, exc,
+        )
+        return
+
+    # Build variant → canonical map
+    remap: dict[str, str] = {}
+    for group in groups:
+        if not isinstance(group, dict):
+            continue
+        canonical = group.get("canonical", "")
+        variants = group.get("variants") or []
+        if not canonical:
+            continue
+        for v in variants:
+            if isinstance(v, str) and v != canonical:
+                remap[v] = canonical
+
+    if not remap:
+        return
+
+    updated = 0
+    for row in rows:
+        changed = False
+        if row.subject_label in remap:
+            row.subject_label = remap[row.subject_label]
+            changed = True
+        if row.object_label in remap:
+            row.object_label = remap[row.object_label]
+            changed = True
+        if changed:
+            updated += 1
+
+    if updated:
+        try:
+            await db.commit()
+            logger.info(
+                "memory_maintenance: _canonicalize_relation_labels user=%s updated=%d",
+                user_id, updated,
+            )
+        except Exception as exc:
+            logger.warning(
+                "memory_maintenance: canonicalize commit failed user=%s: %s", user_id, exc
+            )
+            await db.rollback()

app/core/memory_middleware.py

@@ -0,0 +1,733 @@
+"""Memory Middleware — enrich requests with memory context and store interactions.
+
+Four-tier memory model (MemGPT-style):
+  core         — persistent key/value user preferences, always injected
+  associative  — semantic similarity search via pgvector (top-k)
+  episodic     — recent session summaries (last N)
+  proactive    — behavioral patterns above confidence threshold
+
+All memory content is encrypted at rest using the per-user Fernet key
+stored in User.encryption_key. Decryption happens in-memory only.
+
+Usage:
+    memory = MemoryMiddleware(db_session)
+    context = await memory.enrich_context(user_id, message)
+    # ... run agent ...
+    await memory.store_episode(user_id, session_id, message, response)
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import uuid
+from datetime import datetime, timezone
+from typing import Any
+
+from cryptography.fernet import Fernet, InvalidToken
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models import (
+    ExtractionQueue,
+    MemoryAssociative,
+    MemoryCore,
+    MemoryEpisodic,
+    MemoryProactive,
+    MemoryRelation,
+    User,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def _now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+# Tuning constants
+_ASSOCIATIVE_TOP_K = 5
+_EPISODIC_RECENT_N = 10
+_PROACTIVE_CONFIDENCE_THRESHOLD = 0.6
+
+
+class MemoryMiddleware:
+    """Enrich orchestrator context with memory and persist interactions after."""
+
+    def __init__(self, db: AsyncSession) -> None:
+        self._db = db
+
+    # ── Public API ────────────────────────────────────────────────────────────
+
+    async def enrich_context(
+        self,
+        user_id: str,
+        message: str,
+        trace_id: str | None = None,
+        session_id: str | None = None,
+    ) -> dict[str, Any]:
+        """Build memory context dict to inject into the orchestrator before LLM call.
+
+        Returns a dict with keys:
+          core_memory        — {key: plaintext_value, ...}
+          associative_memory — [plaintext_content, ...]  (top-k by keyword match)
+          episodic_memory    — [plaintext_summary, ...]  (most recent N)
+          proactive_hints    — [plaintext_pattern, ...]  (above threshold)
+          relational_memory  — ["subject --predicate--> object", ...] (top 10, Pro+)
+        """
+        fernet = await self._get_fernet(user_id)
+        if fernet is None:
+            return {}
+
+        user_dbg = await self._get_user_debug(user_id)
+        user_tier: str = user_dbg.get("tier") or "free"
+
+        core = await self._load_core(user_id, fernet)
+        associative = await self._load_associative(user_id, message, fernet, user_tier=user_tier)
+        episodic = await self._load_episodic(user_id, fernet, session_id=session_id)
+        proactive = await self._load_proactive(user_id, fernet)
+        relational = await self._load_relational(user_id, user_tier=user_tier)
+
+        logger.info(
+            "memory: enrich_context trace=%s user=%s tier=%s core=%d associative=%d episodic=%d proactive=%d relational=%d",
+            trace_id or "-",
+            user_id,
+            user_tier,
+            len(core),
+            len(associative),
+            len(episodic),
+            len(proactive),
+            len(relational),
+        )
+
+        return {
+            "core_memory": core,
+            "associative_memory": associative,
+            "episodic_memory": episodic,
+            "proactive_hints": proactive,
+            "relational_memory": relational,
+        }
+
+    async def store_episode(
+        self,
+        user_id: str,
+        session_id: str,
+        message: str,
+        response: str,
+        trace_id: str | None = None,
+    ) -> None:
+        """Summarise and store a completed interaction in episodic memory.
+
+        The summary is a simple heuristic concatenation (no LLM call) to keep
+        latency low. After committing the episode row, dispatches the Mem0-style
+        extraction pipeline:
+          - Pro/Power/Team → asyncio.create_task (fire-and-forget, realtime).
+          - Free → enqueue an ExtractionQueue row for the daily cron.
+        """
+        fernet = await self._get_fernet(user_id)
+        if fernet is None:
+            return
+
+        summary = f"User: {message[:200]}\nAssistant: {response[:200]}"
+        encrypted = _encrypt(fernet, summary)
+
+        episode = MemoryEpisodic(
+            id=str(uuid.uuid4()),
+            user_id=user_id,
+            summary_encrypted=encrypted,
+            session_id=session_id,
+        )
+        self._db.add(episode)
+        episode_id: str = episode.id
+        try:
+            await self._db.commit()
+            user_dbg = await self._get_user_debug(user_id)
+            tier = user_dbg.get("tier") or "free"
+            logger.info(
+                "memory: store_episode trace=%s user=%s tier=%s session=%s",
+                trace_id or "-",
+                user_id,
+                tier,
+                session_id,
+            )
+        except Exception as exc:
+            logger.error("memory: store_episode failed user=%s: %s", user_id, exc)
+            await self._db.rollback()
+            return
+
+        # ── Dispatch extraction pipeline (Phase 2) ────────────────────────────
+        await self._dispatch_extraction(
+            user_id=user_id,
+            episode_id=episode_id,
+            last_user_msg=message,
+            last_assistant_msg=response,
+            session_id=session_id,
+        )
+
+    async def _dispatch_extraction(
+        self,
+        user_id: str,
+        episode_id: str,
+        last_user_msg: str,
+        last_assistant_msg: str,
+        session_id: str | None,
+    ) -> None:
+        """Route extraction to realtime task or batch queue based on user tier."""
+        from app.billing.tier_manager import tier_manager  # noqa: PLC0415
+
+        tier = await tier_manager.get_tier(user_id, self._db)
+
+        if tier_manager.check_feature(tier, "realtime_extraction"):
+            # Pro/Power/Team: fire-and-forget in the background.
+            # Must open a fresh session — request session closes after handler returns.
+            from app.core.memory_extraction import run_extraction  # noqa: PLC0415
+            from app.db import async_session  # noqa: PLC0415
+
+            async def _task() -> None:
+                try:
+                    async with async_session() as fresh_db:
+                        await run_extraction(
+                            db=fresh_db,
+                            user_id=user_id,
+                            last_user_msg=last_user_msg,
+                            last_assistant_msg=last_assistant_msg,
+                            session_id=session_id,
+                        )
+                except Exception as exc:
+                    logger.warning(
+                        "memory: extraction task failed user=%s: %s", user_id, exc
+                    )
+
+            asyncio.create_task(_task())
+            logger.info("memory: realtime extraction dispatched user=%s", user_id)
+        else:
+            # Free tier: enqueue for daily batch cron.
+            queue_row = ExtractionQueue(
+                id=str(uuid.uuid4()),
+                user_id=user_id,
+                episode_id=episode_id,
+            )
+            self._db.add(queue_row)
+            try:
+                await self._db.commit()
+                logger.info(
+                    "memory: extraction enqueued (batch) user=%s episode=%s",
+                    user_id,
+                    episode_id,
+                )
+            except Exception as exc:
+                logger.warning(
+                    "memory: extraction queue insert failed user=%s: %s", user_id, exc
+                )
+                await self._db.rollback()
+
+    async def update_core(self, user_id: str, key: str, value: str, trace_id: str | None = None) -> None:
+        """Upsert a core memory key/value for a user."""
+        fernet = await self._get_fernet(user_id)
+        if fernet is None:
+            return
+
+        encrypted = _encrypt(fernet, value)
+
+        result = await self._db.execute(
+            select(MemoryCore).where(
+                MemoryCore.user_id == user_id,
+                MemoryCore.key == key,
+            )
+        )
+        existing = result.scalar_one_or_none()
+        if existing is not None:
+            existing.value_encrypted = encrypted
+        else:
+            self._db.add(MemoryCore(
+                id=str(uuid.uuid4()),
+                user_id=user_id,
+                key=key,
+                value_encrypted=encrypted,
+            ))
+        try:
+            await self._db.commit()
+            user_dbg = await self._get_user_debug(user_id)
+            logger.info(
+                "memory: update_core trace=%s user=%s tier=%s key=%s",
+                trace_id or "-",
+                user_id,
+                user_dbg.get("tier") or "-",
+                key,
+            )
+        except Exception as exc:
+            logger.error("memory: update_core failed user=%s key=%s: %s", user_id, key, exc)
+            await self._db.rollback()
+
+    async def list_core_blocks(self, user_id: str) -> list[dict[str, str]]:
+        """Return core memory as editable blocks (label/value)."""
+        fernet = await self._get_fernet(user_id)
+        if fernet is None:
+            return []
+
+        result = await self._db.execute(
+            select(MemoryCore)
+            .where(MemoryCore.user_id == user_id)
+            .order_by(MemoryCore.key.asc())
+        )
+        rows = result.scalars().all()
+        out: list[dict[str, str]] = []
+        for row in rows:
+            plaintext = _safe_decrypt(fernet, row.value_encrypted)
+            if plaintext is not None:
+                out.append({"label": row.key, "value": plaintext})
+        logger.debug("memory: list_core_blocks user=%s count=%d", user_id, len(out))
+        return out
+
+    async def get_core_block(self, user_id: str, label: str) -> str | None:
+        """Return a single core memory block value by label."""
+        fernet = await self._get_fernet(user_id)
+        if fernet is None:
+            return None
+
+        result = await self._db.execute(
+            select(MemoryCore).where(
+                MemoryCore.user_id == user_id,
+                MemoryCore.key == label,
+            )
+        )
+        row = result.scalar_one_or_none()
+        if row is None:
+            logger.debug("memory: get_core_block user=%s label=%s found=0", user_id, label)
+            return None
+        value = _safe_decrypt(fernet, row.value_encrypted)
+        logger.debug("memory: get_core_block user=%s label=%s found=%d", user_id, label, 1 if value is not None else 0)
+        return value
+
+    async def delete_core(self, user_id: str, label: str) -> bool:
+        """Delete a core memory block by label. Returns True if deleted."""
+        result = await self._db.execute(
+            select(MemoryCore).where(
+                MemoryCore.user_id == user_id,
+                MemoryCore.key == label,
+            )
+        )
+        row = result.scalar_one_or_none()
+        if row is None:
+            logger.debug("memory: delete_core user=%s label=%s found=0", user_id, label)
+            return False
+
+        await self._db.delete(row)
+        try:
+            await self._db.commit()
+            logger.info("memory: delete_core user=%s label=%s", user_id, label)
+            return True
+        except Exception as exc:
+            logger.error("memory: delete_core failed user=%s label=%s: %s", user_id, label, exc)
+            await self._db.rollback()
+            return False
+
+    async def append_core(self, user_id: str, label: str, content: str) -> None:
+        """Append content to a core block, creating it if missing."""
+        current = await self.get_core_block(user_id, label)
+        if current is None:
+            await self.update_core(user_id, label, content)
+            logger.info("memory: append_core user=%s label=%s created=1", user_id, label)
+            return
+        await self.update_core(user_id, label, f"{current}\n{content}")
+        logger.info("memory: append_core user=%s label=%s created=0", user_id, label)
+
+    async def replace_core(self, user_id: str, label: str, old: str, new: str) -> bool:
+        """Replace one exact string inside a core block. Returns False if not found."""
+        current = await self.get_core_block(user_id, label)
+        if current is None or old not in current:
+            logger.debug("memory: replace_core user=%s label=%s changed=0", user_id, label)
+            return False
+        await self.update_core(user_id, label, current.replace(old, new, 1))
+        logger.info("memory: replace_core user=%s label=%s changed=1", user_id, label)
+        return True
+
+    async def store_associative(
+        self,
+        user_id: str,
+        content: str,
+        entity_type: str | None = None,
+        entity_id: str | None = None,
+    ) -> None:
+        """Store associative memory; embed if user tier has real_embeddings."""
+        from app.billing.tier_manager import tier_manager  # noqa: PLC0415
+        from app.core.embeddings import embed_text  # noqa: PLC0415
+
+        fernet = await self._get_fernet(user_id)
+        if fernet is None:
+            return
+
+        encrypted = _encrypt(fernet, content)
+
+        user_dbg = await self._get_user_debug(user_id)
+        user_tier = user_dbg.get("tier") or "free"
+
+        embedding: list[float] | None = None
+        if tier_manager.check_feature(user_tier, "real_embeddings"):
+            embedding = await embed_text(content)
+
+        row = MemoryAssociative(
+            id=str(uuid.uuid4()),
+            user_id=user_id,
+            content_encrypted=encrypted,
+            embedding=embedding,
+            entity_type=entity_type,
+            entity_id=entity_id,
+        )
+        self._db.add(row)
+        try:
+            await self._db.commit()
+            logger.info(
+                "memory: store_associative user=%s embedded=%s",
+                user_id,
+                embedding is not None,
+            )
+        except Exception as exc:
+            logger.error("memory: store_associative failed user=%s: %s", user_id, exc)
+            await self._db.rollback()
+
+    async def upsert_relation(
+        self,
+        user_id: str,
+        subject: str,
+        subject_type: str,
+        predicate: str,
+        object_: str,
+        object_type: str,
+        *,
+        confidence: float = 0.7,
+        source_episode_id: str | None = None,
+        notes: str | None = None,
+    ) -> None:
+        """Insert or update a relation row.  Matches on (user_id, subject_label, predicate, object_label).
+
+        subject_label / object_label are plaintext entity identifiers — not encrypted.
+        notes is optional; encrypted with user Fernet if provided.
+        """
+        from app.billing.tier_manager import tier_manager  # noqa: PLC0415
+
+        user_dbg = await self._get_user_debug(user_id)
+        user_tier = user_dbg.get("tier") or "free"
+        if not tier_manager.check_feature(user_tier, "relational_memory"):
+            logger.debug("memory: upsert_relation skipped (tier=%s no relational_memory)", user_tier)
+            return
+
+        notes_encrypted: bytes | None = None
+        if notes:
+            fernet = await self._get_fernet(user_id)
+            if fernet:
+                notes_encrypted = fernet.encrypt(notes.encode())
+
+        result = await self._db.execute(
+            select(MemoryRelation).where(
+                MemoryRelation.user_id == user_id,
+                MemoryRelation.subject_label == subject,
+                MemoryRelation.predicate == predicate,
+                MemoryRelation.object_label == object_,
+            )
+        )
+        existing = result.scalar_one_or_none()
+
+        if existing is not None:
+            existing.subject_type = subject_type
+            existing.object_type = object_type
+            existing.confidence = confidence
+            existing.last_confirmed_at = _now()
+            if notes_encrypted is not None:
+                existing.notes_encrypted = notes_encrypted
+        else:
+            self._db.add(MemoryRelation(
+                id=str(uuid.uuid4()),
+                user_id=user_id,
+                subject_label=subject,
+                subject_type=subject_type,
+                predicate=predicate,
+                object_label=object_,
+                object_type=object_type,
+                confidence=confidence,
+                source_episode_id=source_episode_id,
+                notes_encrypted=notes_encrypted,
+            ))
+
+        try:
+            await self._db.commit()
+            logger.info(
+                "memory: upsert_relation user=%s subject=%s predicate=%s object=%s",
+                user_id, subject, predicate, object_,
+            )
+        except Exception as exc:
+            logger.error("memory: upsert_relation failed user=%s: %s", user_id, exc)
+            await self._db.rollback()
+
+    async def query_relations(
+        self,
+        user_id: str,
+        subject: str | None = None,
+        predicate: str | None = None,
+        object_: str | None = None,
+        limit: int = 20,
+    ) -> list[MemoryRelation]:
+        """Query relation rows for a user with optional filters."""
+        q = select(MemoryRelation).where(MemoryRelation.user_id == user_id)
+        if subject is not None:
+            q = q.where(MemoryRelation.subject_label == subject)
+        if predicate is not None:
+            q = q.where(MemoryRelation.predicate == predicate)
+        if object_ is not None:
+            q = q.where(MemoryRelation.object_label == object_)
+        q = q.order_by(MemoryRelation.confidence.desc()).limit(limit)
+        result = await self._db.execute(q)
+        return list(result.scalars().all())
+
+    async def insert_archival(self, user_id: str, content: str, source: str = "manual") -> None:
+        """Insert a long-term archival memory entry."""
+        fernet = await self._get_fernet(user_id)
+        if fernet is None:
+            return
+
+        encrypted = _encrypt(fernet, content)
+        row = MemoryAssociative(
+            id=str(uuid.uuid4()),
+            user_id=user_id,
+            content_encrypted=encrypted,
+            embedding=None,
+            entity_type=source,
+            entity_id=None,
+        )
+        self._db.add(row)
+        try:
+            await self._db.commit()
+            logger.info("memory: insert_archival user=%s source=%s", user_id, source)
+        except Exception as exc:
+            logger.error("memory: insert_archival failed user=%s: %s", user_id, exc)
+            await self._db.rollback()
+
+    async def search_archival(self, user_id: str, query: str, top_k: int = 5) -> list[str]:
+        """Search archival memory (keyword fallback; semantic ranking can replace this)."""
+        fernet = await self._get_fernet(user_id)
+        if fernet is None:
+            return []
+
+        result = await self._db.execute(
+            select(MemoryAssociative)
+            .where(MemoryAssociative.user_id == user_id)
+            .order_by(MemoryAssociative.updated_at.desc())
+            .limit(100)
+        )
+        rows = result.scalars().all()
+        needle = query.strip().lower()
+        out: list[str] = []
+        for row in rows:
+            plaintext = _safe_decrypt(fernet, row.content_encrypted)
+            if plaintext is None:
+                continue
+            if not needle or needle in plaintext.lower():
+                out.append(plaintext)
+            if len(out) >= max(top_k, 1):
+                break
+        logger.info("memory: search_archival user=%s query=%s hits=%d", user_id, query[:80], len(out))
+        return out
+
+    async def search_recall(self, user_id: str, query: str, top_k: int = 5) -> list[str]:
+        """Search recall memory (episodic summaries) by keyword."""
+        fernet = await self._get_fernet(user_id)
+        if fernet is None:
+            return []
+
+        result = await self._db.execute(
+            select(MemoryEpisodic)
+            .where(MemoryEpisodic.user_id == user_id)
+            .order_by(MemoryEpisodic.created_at.desc())
+            .limit(100)
+        )
+        rows = result.scalars().all()
+        needle = query.strip().lower()
+        out: list[str] = []
+        for row in rows:
+            plaintext = _safe_decrypt(fernet, row.summary_encrypted)
+            if plaintext is None:
+                continue
+            if not needle or needle in plaintext.lower():
+                out.append(plaintext)
+            if len(out) >= max(top_k, 1):
+                break
+        logger.info("memory: search_recall user=%s query=%s hits=%d", user_id, query[:80], len(out))
+        return out
+
+    # ── Private helpers ───────────────────────────────────────────────────────
+
+    async def _get_fernet(self, user_id: str) -> Fernet | None:
+        """Load the user's Fernet key from DB. Returns None if missing."""
+        result = await self._db.execute(select(User).where(User.id == user_id))
+        user = result.scalar_one_or_none()
+        if user is None or not user.encryption_key:
+            logger.warning("memory: no encryption_key for user=%s", user_id)
+            return None
+        return Fernet(user.encryption_key.encode())
+
+    async def _get_user_debug(self, user_id: str) -> dict[str, str | None]:
+        """Load lightweight user debug fields for trace logs."""
+        from app.config.settings import settings  # noqa: PLC0415
+        from app.models import Subscription  # noqa: PLC0415
+
+        result = await self._db.execute(select(User).where(User.id == user_id))
+        user = result.scalar_one_or_none()
+        if user is None:
+            return {"tier": None}
+
+        sub_result = await self._db.execute(
+            select(Subscription.tier).where(Subscription.user_id == user_id)
+        )
+        sub_tier: str | None = sub_result.scalar_one_or_none()
+        if sub_tier:
+            tier = sub_tier
+        elif settings.ENV == "dev":
+            tier = "power"
+        else:
+            tier = user.tier or "free"
+
+        return {"tier": tier}
+
+    async def _load_core(self, user_id: str, fernet: Fernet) -> dict[str, str]:
+        result = await self._db.execute(
+            select(MemoryCore).where(MemoryCore.user_id == user_id)
+        )
+        rows = result.scalars().all()
+        out: dict[str, str] = {}
+        for row in rows:
+            plaintext = _safe_decrypt(fernet, row.value_encrypted)
+            if plaintext is not None:
+                out[row.key] = plaintext
+        return out
+
+    async def _load_associative(
+        self, user_id: str, message: str, fernet: Fernet, *, user_tier: str = "free"
+    ) -> list[str]:
+        """Load top-k associative memories.
+
+        Pro+: pgvector cosine similarity on the message embedding (real_embeddings feature).
+        Free / embedding failure: keyword-ordered fallback (most recent rows).
+        """
+        from app.billing.tier_manager import tier_manager  # noqa: PLC0415
+        from app.core.embeddings import embed_text  # noqa: PLC0415
+
+        if tier_manager.check_feature(user_tier, "real_embeddings"):
+            vec = await embed_text(message)
+            if vec is not None:
+                try:
+                    result = await self._db.execute(
+                        select(MemoryAssociative)
+                        .where(
+                            MemoryAssociative.user_id == user_id,
+                            MemoryAssociative.embedding.isnot(None),
+                        )
+                        .order_by(MemoryAssociative.embedding.cosine_distance(vec))
+                        .limit(_ASSOCIATIVE_TOP_K)
+                    )
+                    rows = result.scalars().all()
+                    out: list[str] = []
+                    for row in rows:
+                        plaintext = _safe_decrypt(fernet, row.content_encrypted)
+                        if plaintext is not None:
+                            out.append(plaintext)
+                    logger.info(
+                        "memory: _load_associative user=%s mode=vector hits=%d",
+                        user_id,
+                        len(out),
+                    )
+                    return out
+                except Exception as exc:
+                    logger.warning(
+                        "memory: vector search failed user=%s, falling back to keyword: %s",
+                        user_id,
+                        exc,
+                    )
+
+        # Keyword fallback: most recent rows
+        result = await self._db.execute(
+            select(MemoryAssociative)
+            .where(MemoryAssociative.user_id == user_id)
+            .order_by(MemoryAssociative.updated_at.desc())
+            .limit(_ASSOCIATIVE_TOP_K)
+        )
+        rows = result.scalars().all()
+        out = []
+        for row in rows:
+            plaintext = _safe_decrypt(fernet, row.content_encrypted)
+            if plaintext is not None:
+                out.append(plaintext)
+        return out
+
+    async def _load_episodic(
+        self,
+        user_id: str,
+        fernet: Fernet,
+        session_id: str | None = None,
+    ) -> list[str]:
+        query = select(MemoryEpisodic).where(MemoryEpisodic.user_id == user_id)
+        if session_id:
+            query = query.where(MemoryEpisodic.session_id == session_id)
+        result = await self._db.execute(
+            query
+            .order_by(MemoryEpisodic.created_at.desc())
+            .limit(_EPISODIC_RECENT_N)
+        )
+        rows = result.scalars().all()
+        out: list[str] = []
+        for row in rows:
+            plaintext = _safe_decrypt(fernet, row.summary_encrypted)
+            if plaintext is not None:
+                out.append(plaintext)
+        return out
+
+    async def _load_relational(self, user_id: str, *, user_tier: str = "free") -> list[str]:
+        """Return top-10 relation strings for Pro+ users; empty list for Free."""
+        from app.billing.tier_manager import tier_manager  # noqa: PLC0415
+
+        if not tier_manager.check_feature(user_tier, "relational_memory"):
+            return []
+
+        result = await self._db.execute(
+            select(MemoryRelation)
+            .where(MemoryRelation.user_id == user_id)
+            .order_by(MemoryRelation.confidence.desc())
+            .limit(10)
+        )
+        rows = result.scalars().all()
+        out = [
+            f"{r.subject_label} --{r.predicate}--> {r.object_label}"
+            for r in rows
+        ]
+        return out
+
+    async def _load_proactive(self, user_id: str, fernet: Fernet) -> list[str]:
+        result = await self._db.execute(
+            select(MemoryProactive)
+            .where(
+                MemoryProactive.user_id == user_id,
+                MemoryProactive.confidence >= _PROACTIVE_CONFIDENCE_THRESHOLD,
+            )
+            .order_by(MemoryProactive.confidence.desc())
+        )
+        rows = result.scalars().all()
+        out: list[str] = []
+        for row in rows:
+            plaintext = _safe_decrypt(fernet, row.pattern_encrypted)
+            if plaintext is not None:
+                out.append(plaintext)
+        return out
+
+
+# ── Encryption helpers ────────────────────────────────────────────────────────
+
+def _encrypt(fernet: Fernet, plaintext: str) -> str:
+    return fernet.encrypt(plaintext.encode()).decode()
+
+
+def _safe_decrypt(fernet: Fernet, ciphertext: str) -> str | None:
+    """Decrypt and return plaintext, or None on error (corrupted/wrong key)."""
+    try:
+        return fernet.decrypt(ciphertext.encode()).decode()
+    except (InvalidToken, Exception) as exc:
+        logger.warning("memory: decrypt failed: %s", exc)
+        return None

app/core/note_summarizer.py

@@ -0,0 +1,51 @@
+"""Note summarizer — generates a compact AI summary for a note.
+
+Called fire-and-forget from create_note / update_note tools so the
+``notes.ai_summary`` column stays current without blocking the agent loop.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from langchain_core.messages import HumanMessage, SystemMessage
+
+from app.core.langfuse_client import get_prompt_or_fallback
+from app.core.llm import get_agent_llm
+
+logger = logging.getLogger(__name__)
+
+_FALLBACK_PROMPT = """\
+Summarize this note in <=250 characters. Be terse and dense.
+Keep proper nouns, dates, decisions, and action items.
+Do not start with "This note".
+Respond with the summary text only — no intro, no labels.
+
+Title: {title}
+Content: {content}"""
+
+_MAX_CONTENT_CHARS = 4000
+
+
+async def generate_note_summary(title: str, content: str) -> str:
+    """Return a <=250-char summary of *title* + *content*.
+
+    Uses the Langfuse ``note_summary`` prompt (hot-swappable) with a local
+    fallback.  Truncates *content* to 4000 chars before sending to avoid
+    token waste on large notes.
+    """
+    template, _ = get_prompt_or_fallback("note_summary", _FALLBACK_PROMPT)
+    trimmed = content[:_MAX_CONTENT_CHARS]
+    system_prompt = template.format(title=title, content=trimmed)
+
+    try:
+        llm = get_agent_llm("note-summarizer")
+        response = await llm.ainvoke([
+            SystemMessage(content=system_prompt),
+            HumanMessage(content="Generate the summary."),
+        ])
+        text = response.content if isinstance(response.content, str) else ""
+        return text.strip()[:250]
+    except Exception as exc:
+        logger.warning("note_summarizer: failed to generate summary: %s", exc)
+        return ""

app/core/output_formatter.py

@@ -0,0 +1,63 @@
+"""Output formatter for deep-agent stream events."""
+
+from __future__ import annotations
+
+import re
+from collections.abc import AsyncGenerator
+from typing import Any
+
+from app.schemas import WsStreamEnd, WsStreamStart, WsStreamText
+
+# Matches <canvas kind="...">...</canvas> blocks (single-line or multiline).
+_CANVAS_BLOCK_RE = re.compile(
+    r'<canvas\s+kind=["\']([^"\']+)["\']>(.*?)</canvas>',
+    re.DOTALL | re.IGNORECASE,
+)
+
+
+def extract_canvas_block(text: str) -> tuple[str, str | None, str | None]:
+    """Strip the first <canvas kind="...">...</canvas> block from *text*.
+
+    Returns ``(visible_text, canvas_content, canvas_kind)``.
+    ``canvas_content`` and ``canvas_kind`` are ``None`` when no block is found.
+    """
+    match = _CANVAS_BLOCK_RE.search(text)
+    if not match:
+        return text, None, None
+
+    canvas_kind = match.group(1).strip()
+    canvas_content = match.group(2).strip()
+    visible = text[: match.start()] + text[match.end() :]
+    visible = visible.strip()
+    return visible, canvas_content, canvas_kind
+
+WsFrame = WsStreamStart | WsStreamText | WsStreamEnd
+
+
+class StreamFormatter:
+    """Convert `(event_type, data)` stream events into websocket frame models."""
+
+    def __init__(self, request_id: str) -> None:
+        self.request_id = request_id
+
+    async def format(
+        self,
+        event_stream: AsyncGenerator[tuple[str, Any], None],
+    ) -> AsyncGenerator[WsFrame, None]:
+        started = False
+
+        async for event_type, data in event_stream:
+            if event_type != "token":
+                continue
+
+            if not started:
+                yield WsStreamStart(request_id=self.request_id)
+                started = True
+
+            text = str(data or "")
+            if text:
+                yield WsStreamText(request_id=self.request_id, chunk=text)
+
+        if not started:
+            yield WsStreamStart(request_id=self.request_id)
+        yield WsStreamEnd(request_id=self.request_id)

app/core/preprocessors/__init__.py

@@ -0,0 +1,104 @@
+"""Preprocessor registry: detect content type and dispatch to handlers.
+
+Public API
+----------
+detect_content_type(filename, raw_content) -> str
+    Heuristic detection based on file extension and content patterns.
+
+preprocess(content_type, raw_content) -> PreprocessResult
+    Dispatch to the appropriate handler.
+"""
+
+from __future__ import annotations
+
+import re
+
+from app.core.preprocessors.base import PreprocessResult
+
+# ── Heuristics ────────────────────────────────────────────────────────
+
+# Patterns that strongly suggest an email HTML file
+_EMAIL_SIGNALS = re.compile(
+    r"(Subject:|From:|To:|Date:|Sent:|MIME-Version:|Content-Type:\s*text/html)",
+    re.IGNORECASE,
+)
+
+# Patterns that suggest a generic HTML page (not an email)
+_GENERIC_HTML_SIGNALS = re.compile(
+    r"<(nav|main|header|footer|article|section)\b",
+    re.IGNORECASE,
+)
+
+
+def detect_content_type(filename: str, raw_content: str) -> str:
+    """Return a content-type string for the given file.
+
+    Supported types: ``"email_html"``, ``"generic_html"``,
+    ``"plain_text"``, ``"unknown"``.
+    """
+    ext = filename.rsplit(".", 1)[-1].lower() if "." in filename else ""
+
+    if ext == "txt":
+        return "plain_text"
+
+    if ext in ("html", "htm", "eml", "mhtml", "mht"):
+        # Prefer email detection over generic HTML
+        if _EMAIL_SIGNALS.search(raw_content[:4096]):
+            return "email_html"
+        if _GENERIC_HTML_SIGNALS.search(raw_content[:4096]) or "<html" in raw_content[:200].lower():
+            return "generic_html"
+        # .html without clear signals — check for any email header
+        if re.search(r"^(From|To|Subject|Date):", raw_content[:2048], re.MULTILINE | re.IGNORECASE):
+            return "email_html"
+        return "generic_html"
+
+    # Plain text files with email headers
+    if ext in ("", "txt") or not ext:
+        if _EMAIL_SIGNALS.search(raw_content[:4096]):
+            return "email_html"
+
+    # Detect binary content
+    try:
+        raw_content.encode("utf-8")
+    except (UnicodeEncodeError, AttributeError):
+        return "unknown"
+
+    # Non-text bytes heuristic: high ratio of non-printable chars
+    sample = raw_content[:512]
+    non_printable = sum(1 for c in sample if ord(c) < 32 and c not in "\r\n\t")
+    if len(sample) > 0 and non_printable / len(sample) > 0.1:
+        return "unknown"
+
+    return "unknown"
+
+
+# ── Generic fallback handler ──────────────────────────────────────────
+
+def _preprocess_generic(raw_content: str, content_type: str) -> PreprocessResult:
+    """Strip HTML tags if present, return text as-is."""
+    try:
+        from bs4 import BeautifulSoup
+        text = BeautifulSoup(raw_content, "html.parser").get_text(separator="\n")
+    except ImportError:
+        # No BeautifulSoup — strip tags with a simple regex
+        text = re.sub(r"<[^>]+>", "", raw_content)
+
+    text = re.sub(r"\n{3,}", "\n\n", text).strip()
+    return PreprocessResult(content_type=content_type, clean_text=text, metadata={})
+
+
+# ── Dispatch ──────────────────────────────────────────────────────────
+
+def preprocess(content_type: str, raw_content: str) -> PreprocessResult:
+    """Dispatch *raw_content* to the handler registered for *content_type*.
+
+    Falls back to the generic handler for unknown types.
+    """
+    if content_type == "email_html":
+        from app.core.preprocessors.email_html import preprocess_email_html
+        return preprocess_email_html(raw_content)
+
+    return _preprocess_generic(raw_content, content_type)
+
+
+__all__ = ["detect_content_type", "preprocess", "PreprocessResult"]

app/core/preprocessors/base.py

@@ -0,0 +1,25 @@
+"""Base types for the preprocessor system."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+
+
+@dataclass
+class PreprocessResult:
+    """Output of a preprocessor handler.
+
+    Attributes
+    ----------
+    content_type:
+        The detected content type (e.g. ``"email_html"``, ``"plain_text"``).
+    clean_text:
+        Human-readable text stripped of markup/binary noise.
+    metadata:
+        Dict of extracted metadata (keys vary by handler).
+        Common keys: ``subject``, ``from``, ``to``, ``date``, ``filename``.
+    """
+
+    content_type: str
+    clean_text: str
+    metadata: dict = field(default_factory=dict)

app/core/preprocessors/email_html.py

@@ -0,0 +1,111 @@
+"""Preprocessor for email HTML files.
+
+Handles:
+- HTML stripping via BeautifulSoup
+- Metadata extraction (Subject, From, To, Date)
+- Thread splitting — isolates the latest reply
+"""
+
+from __future__ import annotations
+
+import re
+from typing import TYPE_CHECKING
+
+from app.core.preprocessors.base import PreprocessResult
+
+if TYPE_CHECKING:
+    pass
+
+# ── Thread split markers ──────────────────────────────────────────────
+
+# Matches patterns like:
+#   "On Mon, Apr 7, 2026 at 10:00 AM, Alice <alice@co.com> wrote:"
+#   "-----Original Message-----"
+#   "> " (plain-text quote prefix)
+_THREAD_PATTERNS = [
+    re.compile(r"^On\s+.+wrote\s*:", re.IGNORECASE | re.MULTILINE),
+    re.compile(r"^-{3,}\s*(original message|forwarded message)\s*-{3,}", re.IGNORECASE | re.MULTILINE),
+    re.compile(r"^>{1,}\s+\S", re.MULTILINE),
+    re.compile(r"^From:\s+.+\nSent:\s+", re.IGNORECASE | re.MULTILINE),
+]
+
+# ── Metadata patterns (applied on raw HTML / plain fallback) ──────────
+
+_META_PATTERNS: dict[str, list[re.Pattern]] = {
+    "subject": [
+        re.compile(r"<title>(.+?)</title>", re.IGNORECASE | re.DOTALL),
+        re.compile(r"Subject:\s*(.+)", re.IGNORECASE),
+    ],
+    "from": [
+        re.compile(r'<meta[^>]+name=["\']?from["\']?[^>]+content=["\']([^"\']+)["\']', re.IGNORECASE),
+        re.compile(r"From:\s*(.+)", re.IGNORECASE),
+    ],
+    "to": [
+        re.compile(r'<meta[^>]+name=["\']?to["\']?[^>]+content=["\']([^"\']+)["\']', re.IGNORECASE),
+        re.compile(r"To:\s*(.+)", re.IGNORECASE),
+    ],
+    "date": [
+        re.compile(r'<meta[^>]+name=["\']?date["\']?[^>]+content=["\']([^"\']+)["\']', re.IGNORECASE),
+        re.compile(r"Date:\s*(.+)", re.IGNORECASE),
+        re.compile(r"Sent:\s*(.+)", re.IGNORECASE),
+    ],
+}
+
+
+def _extract_metadata(raw_html: str, text: str) -> dict:
+    """Extract Subject/From/To/Date from raw HTML or plain text."""
+    metadata: dict[str, str] = {}
+    for field, patterns in _META_PATTERNS.items():
+        for pat in patterns:
+            m = pat.search(raw_html) or pat.search(text)
+            if m:
+                metadata[field] = m.group(1).strip()
+                break
+    return metadata
+
+
+def _split_thread(text: str) -> str:
+    """Return only the latest message in a threaded email."""
+    earliest_pos: int | None = None
+    for pat in _THREAD_PATTERNS:
+        m = pat.search(text)
+        if m and (earliest_pos is None or m.start() < earliest_pos):
+            earliest_pos = m.start()
+
+    if earliest_pos is not None and earliest_pos > 0:
+        return text[:earliest_pos].strip()
+    return text.strip()
+
+
+def preprocess_email_html(raw_content: str) -> PreprocessResult:
+    """Strip HTML, extract metadata, split thread from an email HTML file."""
+    try:
+        from bs4 import BeautifulSoup  # lazy import — optional dep
+    except ImportError as exc:
+        raise ImportError(
+            "beautifulsoup4 is required for email_html preprocessing. "
+            "Install it with: pip install beautifulsoup4"
+        ) from exc
+
+    # Parse with lxml if available, fall back to html.parser
+    try:
+        soup = BeautifulSoup(raw_content, "lxml")
+    except Exception:
+        soup = BeautifulSoup(raw_content, "html.parser")
+
+    # Remove noise tags
+    for tag in soup(["style", "script", "head", "noscript"]):
+        tag.decompose()
+
+    clean_text = soup.get_text(separator="\n")
+    # Collapse excessive blank lines
+    clean_text = re.sub(r"\n{3,}", "\n\n", clean_text).strip()
+
+    metadata = _extract_metadata(raw_content, clean_text)
+    latest_message = _split_thread(clean_text)
+
+    return PreprocessResult(
+        content_type="email_html",
+        clean_text=latest_message,
+        metadata=metadata,
+    )

app/core/scout_registry.py

@@ -0,0 +1,30 @@
+"""Minimal agent base types retained for compatibility with batch runners."""
+
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from typing import Any
+
+
+class BaseAgent(ABC):
+    """Common base for non-chat agents still using the old base contract."""
+
+    def __init__(
+        self,
+        user_id: str = "",
+        shared_memory: dict[str, Any] | None = None,
+        vector_store_context: list[str] | None = None,
+    ) -> None:
+        self.user_id = user_id
+        self.shared_memory: dict[str, Any] = shared_memory or {}
+        self.vector_store_context: list[str] = vector_store_context or []
+
+    @abstractmethod
+    def get_name(self) -> str: ...
+
+    @abstractmethod
+    def get_description(self) -> str: ...
+
+    @property
+    def skills(self) -> list[str]:
+        return []

app/core/scout_session_buffer.py

@@ -0,0 +1,96 @@
+"""In-process TTL buffer for per-session LangChain message history.
+
+Stores the full message list (including AIMessage with tool_calls and ToolMessage)
+keyed by (user_id, session_id), so agents can reconstruct tool-call context across
+conversation turns without it being lossy through the wire.
+
+Single-process only. For multi-worker deployments, replace the _SessionBuffer
+implementation with one backed by Redis (serialize LangChain messages to dicts via
+message_to_dict / messages_from_dict from langchain_core.messages).
+"""
+from __future__ import annotations
+
+import time
+from threading import Lock
+
+from langchain_core.messages import BaseMessage
+
+SESSION_TTL_SECONDS = 1800  # 30-minute idle expiry
+MAX_MESSAGES_PER_SESSION = 80  # cap to avoid unbounded memory growth
+
+
+class _SessionBuffer:
+    def __init__(self) -> None:
+        self._store: dict[tuple[str, str], tuple[float, list[BaseMessage]]] = {}
+        self._lock = Lock()
+
+    def _evict_stale(self) -> None:
+        now = time.monotonic()
+        stale = [k for k, (ts, _) in self._store.items() if now - ts > SESSION_TTL_SECONDS]
+        for k in stale:
+            del self._store[k]
+
+    def get(self, user_id: str, session_id: str) -> list[BaseMessage] | None:
+        key = (user_id, session_id)
+        with self._lock:
+            entry = self._store.get(key)
+            if entry is None:
+                return None
+            ts, msgs = entry
+            if time.monotonic() - ts > SESSION_TTL_SECONDS:
+                del self._store[key]
+                return None
+            self._store[key] = (time.monotonic(), msgs)
+            return list(msgs)
+
+    def set(self, user_id: str, session_id: str, messages: list[BaseMessage]) -> None:
+        key = (user_id, session_id)
+        capped = messages[-MAX_MESSAGES_PER_SESSION:]
+        with self._lock:
+            self._evict_stale()
+            self._store[key] = (time.monotonic(), capped)
+
+    def clear(self, user_id: str, session_id: str) -> None:
+        with self._lock:
+            self._store.pop((user_id, session_id), None)
+
+    def append_system_message(self, user_id: str, session_id: str, text: str) -> None:
+        """Append a synthetic system message to the buffer for the given session.
+
+        Creates the session slot if it does not yet exist.  Used by the
+        contextual_scope_update handler to inject navigation events without
+        making an LLM call.
+        """
+        from langchain_core.messages import SystemMessage  # noqa: PLC0415
+
+        key = (user_id, session_id)
+        with self._lock:
+            entry = self._store.get(key)
+            if entry is None:
+                msgs: list[BaseMessage] = [SystemMessage(content=text)]
+            else:
+                _, existing = entry
+                msgs = list(existing) + [SystemMessage(content=text)]
+            capped = msgs[-MAX_MESSAGES_PER_SESSION:]
+            self._store[key] = (time.monotonic(), capped)
+
+
+class ContextualBufferProxy:
+    """Thin wrapper around _SessionBuffer that closes over user_id + session_id.
+
+    Returned by get_session_buffer() so callers can call
+    ``proxy.append_system_message(text)`` without threading user_id/session_id
+    through every call site.
+    """
+
+    def __init__(self, buf: "_SessionBuffer", user_id: str, session_id: str) -> None:
+        self._buf = buf
+        self._user_id = user_id
+        self._session_id = session_id
+
+    def append_system_message(self, text: str) -> None:
+        self._buf.append_system_message(self._user_id, self._session_id, text)
+
+
+# Module-level singleton — same pattern as _pending_states in api/app/api/routes/auth.py
+session_buffer = _SessionBuffer()

app/core/ws_context.py

@@ -0,0 +1,115 @@
+"""WebSocket client executor context.
+
+Holds a per-request async callback that tools call to execute CRUD
+operations on the Electron client's local SQLite / LanceDB databases.
+The callback sends a `tool_call` WS frame and awaits the `tool_result`.
+"""
+
+from __future__ import annotations
+
+import re
+from contextvars import ContextVar
+from typing import Any, Callable, Coroutine
+from uuid import uuid4
+
+_SNAKE_TO_CAMEL_RE = re.compile(r"_([a-z])")
+
+
+def _key_to_camel(key: str) -> str:
+    return _SNAKE_TO_CAMEL_RE.sub(lambda m: m.group(1).upper(), key)
+
+
+def _keys_to_camel(obj: Any) -> Any:
+    """Recursively convert dict keys from snake_case to camelCase.
+
+    Mirrors the JS-side ``toCamelCase`` applied to incoming WS frames in
+    ``adiuvAI/src/main/api/backend-client.ts``. The Electron executor wraps
+    tool_result payloads in ``toSnakeCase`` before sending; this restores the
+    camelCase schema property names that the tool code expects to read.
+    """
+    if isinstance(obj, dict):
+        return {_key_to_camel(k): _keys_to_camel(v) for k, v in obj.items()}
+    if isinstance(obj, list):
+        return [_keys_to_camel(v) for v in obj]
+    return obj
+
+# Holds the execute callback for the current WS session.
+# Set by the chat WS handler before the orchestrator runs; cleared after.
+_client_executor: ContextVar[Callable[[dict], Coroutine[Any, Any, dict]]] = ContextVar(
+    "_client_executor"
+)
+
+# Optional collector that captures raw execute_on_client results.
+# Set by _tool_loop / _tool_loop_stream to populate ChatAgent.tool_results.
+_tool_result_collector: ContextVar[list[dict] | None] = ContextVar(
+    "_tool_result_collector", default=None
+)
+
+
+def set_tool_result_collector(lst: list[dict]) -> None:
+    """Register *lst* as the collector for this async context."""
+    _tool_result_collector.set(lst)
+
+
+def clear_tool_result_collector() -> None:
+    """Clear the collector (best-effort)."""
+    _tool_result_collector.set(None)
+
+
+def set_client_executor(fn: Callable[[dict], Coroutine[Any, Any, dict]]) -> None:
+    """Bind *fn* as the executor for the current async context (task/coroutine)."""
+    _client_executor.set(fn)
+
+
+def clear_client_executor() -> None:
+    """Remove the executor binding (best-effort; ContextVar resets on task exit)."""
+    try:
+        _client_executor.set(None)  # type: ignore[arg-type]
+    except Exception:
+        pass
+
+
+async def execute_on_client(
+    action: str,
+    table: str | None = None,
+    data: dict[str, Any] | None = None,
+    filters: dict[str, Any] | None = None,
+    vector: list[float] | None = None,
+    limit: int | None = None,
+) -> dict[str, Any]:
+    """Send a CRUD/vector operation to the Electron client and return the result.
+
+    Builds a ``tool_call`` payload, invokes the per-session WS callback,
+    and returns the ``tool_result`` dict from Electron.
+
+    Raises ``RuntimeError`` if no executor is set (i.e. called outside a WS session).
+    """
+    callback = _client_executor.get(None)
+    if callback is None:
+        raise RuntimeError(
+            "execute_on_client() called outside a WebSocket session — "
+            "no client executor is set."
+        )
+
+    payload: dict[str, Any] = {"id": str(uuid4()), "action": action}
+    if table is not None:
+        payload["table"] = table
+    if data is not None:
+        payload["data"] = data
+    if filters is not None:
+        payload["filters"] = {k: v for k, v in filters.items() if v is not None}
+    if vector is not None:
+        payload["vector"] = vector
+    if limit is not None:
+        payload["limit"] = limit
+
+    result = await callback(payload)
+    result = _keys_to_camel(result)
+    collector = _tool_result_collector.get(None)
+    if collector is not None:
+        collector.append({
+            "action": action,
+            "table": table,
+            "data": result,
+        })
+    return result

app/db.py

@@ -1,7 +1,15 @@
-"""Database engine, session factory, and declarative base.
+"""Database engine, session factory, and base model.
 
-All services use the async SQLAlchemy API via ``get_session()``.
+All app code uses the async SQLAlchemy API.  Alembic migrations use the
-Alembic migrations use the synchronous psycopg2 URL (see alembic/env.py).
+synchronous psycopg2 URL for the CLI (see alembic/env.py).
+
+Usage in routes:
+    from app.db import get_session
+    from sqlalchemy.ext.asyncio import AsyncSession
+
+    async def my_route(db: AsyncSession = Depends(get_session)):
+        result = await db.execute(select(User).where(User.email == email))
+        user = result.scalar_one_or_none()
 """
 
 from __future__ import annotations
@@ -11,7 +19,7 @@ from collections.abc import AsyncGenerator
 from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
 from sqlalchemy.orm import DeclarativeBase
 
-from shared.config import settings
+from app.config.settings import settings
 
 engine = create_async_engine(
     settings.DATABASE_URL,

app/integrations/__init__.py

@@ -1,11 +1,20 @@
 """Cloud provider integration utilities.
 
-Adapted for Batch Agent Service: import from shared.config instead of app.config.
-
 Provides:
-  * Shared message dataclasses (EmailMessage, ChatMessage)
+  * Shared message dataclasses (``EmailMessage``, ``ChatMessage``) used by
-  * get_provider() — factory for Gmail/MS Graph clients
+    both the Gmail and MS Graph clients and consumed by ``agent_runner``.
-  * encrypt_token() / decrypt_token() — Fernet-based OAuth token encryption
+  * ``get_provider()`` — factory that returns the correct client given a
+    provider name and decrypted OAuth credentials dict.
+  * ``encrypt_token()`` / ``decrypt_token()`` — Fernet-based at-rest
+    encryption for OAuth tokens stored in ``cloud_agent_configs``.
+
+Encryption rationale
+--------------------
+Unlike user content (which is E2E-encrypted client-side and **never**
+decrypted server-side), OAuth tokens *must* be decrypted server-side
+because the backend makes provider API calls on behalf of the user.
+The Fernet key lives solely in ``OAUTH_ENCRYPTION_KEY`` env var — it
+is never returned to clients.
 """
 
 from __future__ import annotations
@@ -18,7 +27,7 @@ from typing import TYPE_CHECKING
 
 from cryptography.fernet import Fernet, InvalidToken
 
-from shared.config import settings
+from app.config.settings import settings
 
 if TYPE_CHECKING:
     from app.integrations.gmail import GmailClient
@@ -26,9 +35,13 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
+# ── Shared message types ──────────────────────────────────────────────────
+
 
 @dataclass
 class EmailMessage:
+    """A single email message fetched from Gmail or Outlook."""
+
     id: str
     subject: str
     sender: str
@@ -38,6 +51,7 @@ class EmailMessage:
 
     @property
     def as_text(self) -> str:
+        """Return a human-readable text representation for LLM extraction."""
         date_str = self.date.strftime("%Y-%m-%d %H:%M")
         labels_str = f" [{', '.join(self.labels)}]" if self.labels else ""
         return (
@@ -50,6 +64,8 @@ class EmailMessage:
 
 @dataclass
 class ChatMessage:
+    """A single Teams chat or channel message fetched from MS Graph."""
+
     id: str
     content: str
     sender: str
@@ -58,6 +74,7 @@ class ChatMessage:
 
     @property
     def as_text(self) -> str:
+        """Return a human-readable text representation for LLM extraction."""
         date_str = self.date.strftime("%Y-%m-%d %H:%M")
         channel_str = f" [channel: {self.channel}]" if self.channel else ""
         return (
@@ -67,7 +84,15 @@ class ChatMessage:
         )
 
 
+# ── Fernet helpers ────────────────────────────────────────────────────────
+
+
 def _get_fernet() -> Fernet:
+    """Return a ``Fernet`` instance using ``settings.OAUTH_ENCRYPTION_KEY``.
+
+    Raises ``RuntimeError`` if ``OAUTH_ENCRYPTION_KEY`` is not set — callers
+    must ensure this is configured before persisting OAuth tokens.
+    """
     key = settings.OAUTH_ENCRYPTION_KEY
     if not key:
         raise RuntimeError(
@@ -78,6 +103,15 @@ def _get_fernet() -> Fernet:
 
 
 def encrypt_token(token_info: dict) -> str:
+    """Fernet-encrypt an OAuth credential dict and return a base64 string.
+
+    Stores the full ``{access_token, refresh_token, token_uri, client_id,
+    client_secret, scopes, expiry}`` dict (or equivalent MSAL shape).
+
+    Raises:
+        RuntimeError: OAUTH_ENCRYPTION_KEY is not configured.
+        ValueError: ``token_info`` is not a non-empty dict.
+    """
     if not isinstance(token_info, dict) or not token_info:
         raise ValueError("token_info must be a non-empty dict")
     plaintext = json.dumps(token_info).encode("utf-8")
@@ -85,6 +119,13 @@ def encrypt_token(token_info: dict) -> str:
 
 
 def decrypt_token(encrypted: str) -> dict:
+    """Decrypt a Fernet-encrypted token string and return the credential dict.
+
+    Raises:
+        RuntimeError: OAUTH_ENCRYPTION_KEY is not configured.
+        ValueError: The encrypted string is invalid or was encrypted with a
+            different key.
+    """
     try:
         plaintext = _get_fernet().decrypt(encrypted.encode("utf-8"))
         return json.loads(plaintext)
@@ -92,10 +133,25 @@ def decrypt_token(encrypted: str) -> dict:
         raise ValueError(f"Failed to decrypt OAuth token: {exc}") from exc
 
 
+# ── Provider factory ──────────────────────────────────────────────────────
+
+
 def get_provider(
     provider: str,
     credentials_info: dict,
 ) -> "GmailClient | MSGraphClient":
+    """Return the correct provider client for *provider*.
+
+    Parameters
+    ----------
+    provider:
+        One of ``"gmail"``, ``"outlook"``, ``"teams"``.
+    credentials_info:
+        Decrypted OAuth credential dict (Google or Microsoft shape).
+
+    Raises:
+        ValueError: Unknown provider name.
+    """
     if provider == "gmail":
         from app.integrations.gmail import GmailClient
         return GmailClient(credentials_info)

app/integrations/gmail.py

@@ -1,7 +1,26 @@
 """Gmail API client for cloud agent integration.
 
-Adapted for Batch Agent Service: import from app.integrations instead of
+Wraps the Google Gmail REST API to fetch email messages matching a
-app.integrations (same relative path within the service).
+``filter_config`` dict.  Uses the official ``google-api-python-client``
+library (synchronous) wrapped in ``asyncio.to_thread()`` to avoid
+blocking the event loop.
+
+Token refresh is handled transparently: when the stored access token has
+expired, ``google.auth.transport.requests.Request`` will use the refresh
+token to obtain a fresh one.  The caller is responsible for persisting
+any refreshed credentials back to ``CloudScoutConfig.oauth_token_encrypted``
+(see ``agent_runner.run_cloud_agent``).
+
+Credential dict shape (Google OAuth2):
+    {
+        "token": "<access_token>",
+        "refresh_token": "<refresh_token>",
+        "token_uri": "https://oauth2.googleapis.com/token",
+        "client_id": "<client_id>",
+        "client_secret": "<client_secret>",
+        "scopes": ["https://www.googleapis.com/auth/gmail.readonly"],
+        "expiry": "2025-01-01T00:00:00Z"  # optional ISO-8601
+    }
 """
 
 from __future__ import annotations
@@ -19,8 +38,13 @@ from app.integrations import EmailMessage
 
 logger = logging.getLogger(__name__)
 
+# Gmail search date format — e.g. "after:2025/01/01"
 _GMAIL_DATE_FMT = "%Y/%m/%d"
+
+# Maximum characters of body text forwarded to the LLM.
 _BODY_TRUNCATE = 8_000
+
+# Maximum messages retrieved per run (prevents runaway quota usage).
 _MAX_MESSAGES = 200
 
 
@@ -28,9 +52,20 @@ def _build_gmail_query(
     filter_config: dict[str, Any] | None,
     since: datetime | None,
 ) -> str:
+    """Build a Gmail search query string from *filter_config* and *since*.
+
+    Supported ``filter_config`` keys:
+        labels (list[str]):  Gmail label names, e.g. ``["INBOX", "work"]``
+        senders (list[str]): Sender addresses or domains to include
+        date_range (dict):   ``{from: "<YYYY-MM-DD>", to: "<YYYY-MM-DD>"}``
+
+    A hard ``since`` date (from last run) always overrides ``date_range.from``
+    when it is earlier.
+    """
     parts: list[str] = []
     cfg = filter_config or {}
 
+    # Labels — joined with OR when multiple given.
     labels: list[str] = cfg.get("labels", [])
     if labels:
         if len(labels) == 1:
@@ -39,14 +74,17 @@ def _build_gmail_query(
             label_expr = " OR ".join(f"label:{lbl}" for lbl in labels)
             parts.append(f"({label_expr})")
 
+    # Senders — each prefixed with "from:".
     senders: list[str] = cfg.get("senders", [])
     for sender in senders:
         parts.append(f"from:{sender}")
 
+    # Date range.
     date_range: dict = cfg.get("date_range", {})
     from_str: str | None = date_range.get("from")
     to_str: str | None = date_range.get("to")
 
+    # Determine effective "from" date: most recent of filter_config.date_range.from and since.
     effective_since: datetime | None = since
     if from_str:
         try:
@@ -72,12 +110,18 @@ def _build_gmail_query(
 
 
 def _strip_html(raw_html: str) -> str:
+    """Remove HTML tags and decode entities to get plain text."""
     no_tags = re.sub(r"<[^>]+>", " ", raw_html)
     decoded = html.unescape(no_tags)
     return re.sub(r"\s+", " ", decoded).strip()
 
 
 def _parse_body(payload: dict[str, Any]) -> str:
+    """Recursively extract the plain-text body from a Gmail message payload.
+
+    Prefers ``text/plain``; falls back to ``text/html`` (stripped of tags).
+    Returns an empty string if no body can be extracted.
+    """
     mime_type: str = payload.get("mimeType", "")
     body: dict = payload.get("body", {})
     parts: list[dict] = payload.get("parts", [])
@@ -95,6 +139,7 @@ def _parse_body(payload: dict[str, Any]) -> str:
             return _strip_html(raw)
         return ""
 
+    # Multipart — prefer text/plain part, fall back to text/html.
     plain_fallback = ""
     for part in parts:
         part_mime = part.get("mimeType", "")
@@ -110,6 +155,7 @@ def _parse_body(payload: dict[str, Any]) -> str:
 
 
 def _parse_date(raw: str) -> datetime:
+    """Parse an RFC 2822 email date header into a UTC ``datetime``."""
     try:
         parsed = email.utils.parsedate_to_datetime(raw)
         if parsed.tzinfo is None:
@@ -120,6 +166,16 @@ def _parse_date(raw: str) -> datetime:
 
 
 class GmailClient:
+    """Fetch email messages from a Gmail account via the Gmail REST API.
+
+    Parameters
+    ----------
+    credentials_info:
+        Decrypted OAuth2 credential dict.  Must contain at minimum
+        ``token`` (access token) or ``refresh_token`` + ``token_uri`` +
+        ``client_id`` + ``client_secret``.
+    """
+
     def __init__(self, credentials_info: dict[str, Any]) -> None:
         from google.oauth2.credentials import Credentials
 
@@ -144,20 +200,38 @@ class GmailClient:
             expiry=expiry,
         )
 
+    # ── Public API ─────────────────────────────────────────────────────────
+
     async def fetch_messages(
         self,
         filter_config: dict[str, Any] | None = None,
         since: datetime | None = None,
     ) -> list[EmailMessage]:
+        """Return up to ``_MAX_MESSAGES`` emails matching *filter_config*.
+
+        Runs the synchronous Google API calls inside ``asyncio.to_thread()``
+        to avoid blocking the async event loop.
+
+        Token refresh is performed automatically when the access token has
+        expired.  After the call, ``self.refreshed_credentials`` may be
+        consulted to detect whether new credentials should be persisted.
+        """
         query = _build_gmail_query(filter_config, since)
         logger.debug("gmail: executing search query %r", query)
         return await asyncio.to_thread(self._fetch_sync, query)
 
     @property
     def refreshed_credentials(self) -> dict[str, Any] | None:
+        """Return updated credential dict if the access token was refreshed.
+
+        If the credentials were refreshed during ``fetch_messages()``, returns
+        a new dict that should be re-encrypted and written back to the DB.
+        Returns ``None`` if no refresh occurred.
+        """
         creds = self._credentials
         if not creds.valid and creds.expired:
             return None
+        # Check whether the token changed from what was stored.
         if creds.token != self._credentials_info.get("token"):
             result = {
                 "token": creds.token,
@@ -172,11 +246,15 @@ class GmailClient:
             return result
         return None
 
+    # ── Internal sync worker ───────────────────────────────────────────────
+
     def _fetch_sync(self, query: str) -> list[EmailMessage]:
+        """Synchronous worker — called inside ``asyncio.to_thread()``."""
         import googleapiclient.discovery
         import googleapiclient.errors
         from google.auth.transport.requests import Request
 
+        # Refresh token if needed before building the service.
         if self._credentials.expired and self._credentials.refresh_token:
             try:
                 self._credentials.refresh(Request())
@@ -186,8 +264,9 @@ class GmailClient:
         service = googleapiclient.discovery.build(
             "gmail", "v1", credentials=self._credentials, cache_discovery=False
         )
-        user_api = service.users()
+        user_api = service.users()  # type: ignore[attr-defined]
 
+        # ── List matching message IDs ──────────────────────────────────────
         ids: list[str] = []
         page_token: str | None = None
         while len(ids) < _MAX_MESSAGES:
@@ -214,10 +293,12 @@ class GmailClient:
                 break
 
         if not ids:
+            logger.debug("gmail: no messages matched query %r", query)
             return []
 
         logger.info("gmail: fetching %d message(s)", len(ids))
 
+        # ── Fetch individual message details ──────────────────────────────
         messages: list[EmailMessage] = []
         for msg_id in ids:
             try:
@@ -245,8 +326,10 @@ class GmailClient:
                     date=date,
                     labels=labels,
                 ))
+            except googleapiclient.errors.HttpError as exc:
+                logger.warning("gmail: skipping message %s — HTTP error: %s", msg_id, exc)
             except Exception as exc:
-                logger.warning("gmail: skipping message %s: %s", msg_id, exc)
+                logger.warning("gmail: skipping message %s — unexpected error: %s", msg_id, exc)
 
         logger.info("gmail: returned %d message(s)", len(messages))
         return messages

app/integrations/ms_graph.py

@@ -1,6 +1,24 @@
-"""Microsoft Graph API client for Outlook and Teams.
+"""Microsoft Graph API client for Outlook and Teams cloud agent integration.
 
-Adapted for Batch Agent Service: import settings from shared.config.
+Handles two data sources:
+
+* **Outlook email** (``provider="outlook"``) — ``fetch_emails()`` calls
+  ``/me/messages`` with an OData ``$filter`` built from ``filter_config``.
+* **Teams messages** (``provider="teams"``) — ``fetch_messages()`` calls
+  ``/me/chats/getAllMessages`` filtered by date.
+
+Authentication uses MSAL ``PublicClientApplication`` to acquire a token
+from a stored refresh token.  The ``httpx.AsyncClient`` (already a project
+dependency) is used for all API calls.
+
+Credential dict shape (Microsoft OAuth2 / MSAL):
+    {
+        "access_token":  "<access_token>",
+        "refresh_token": "<refresh_token>",
+        "token_type":    "Bearer",
+        "scope":         "Mail.Read ChannelMessage.Read.All offline_access",
+        "expires_in":    3600
+    }
 """
 
 from __future__ import annotations
@@ -12,19 +30,23 @@ from typing import Any
 
 import httpx
 
-from shared.config import settings
+from app.config.settings import settings
 from app.integrations import ChatMessage, EmailMessage
 
 logger = logging.getLogger(__name__)
 
 _GRAPH_BASE = "https://graph.microsoft.com/v1.0"
 
+# Max items fetched per run.
 _MAX_EMAILS = 200
 _MAX_MESSAGES = 200
+
+# Max characters of body forwarded to the LLM.
 _BODY_TRUNCATE = 8_000
 
 
 def _strip_html(raw: str) -> str:
+    """Strip HTML tags and collapse whitespace."""
     no_tags = re.sub(r"<[^>]+>", " ", raw)
     import html as _html
     decoded = _html.unescape(no_tags)
@@ -32,6 +54,7 @@ def _strip_html(raw: str) -> str:
 
 
 def _odata_datetime(dt: datetime) -> str:
+    """Format a datetime as an OData datetime literal (UTC, ISO 8601)."""
     utc = dt.astimezone(timezone.utc)
     return utc.strftime("%Y-%m-%dT%H:%M:%SZ")
 
@@ -40,14 +63,29 @@ def _build_email_filter(
     filter_config: dict[str, Any] | None,
     since: datetime | None,
 ) -> str:
+    """Build an OData ``$filter`` expression for the ``/me/messages`` endpoint.
+
+    Supported ``filter_config`` keys:
+        senders (list[str]):  Sender email addresses.
+        date_range (dict):    ``{from: "<ISO-8601>", to: "<ISO-8601>"}``
+        folders (list[str]):  Folder display names (not directly filterable
+                              via OData, so ignored here — callers iterate
+                              folder IDs separately if needed; listed for
+                              completeness).
+
+    A hard ``since`` date always overrides ``date_range.from`` when it is
+    earlier.
+    """
     clauses: list[str] = []
     cfg = filter_config or {}
 
+    # Senders.
     senders: list[str] = cfg.get("senders", [])
     if senders:
         sender_clauses = [f"from/emailAddress/address eq '{s}'" for s in senders]
         clauses.append("(" + " or ".join(sender_clauses) + ")")
 
+    # Date range.
     date_range: dict = cfg.get("date_range", {})
     from_str: str | None = date_range.get("from")
 
@@ -79,16 +117,33 @@ def _build_email_filter(
 
 
 class MSGraphClient:
+    """Fetch emails and Teams messages via the Microsoft Graph REST API.
+
+    Parameters
+    ----------
+    credentials_info:
+        Decrypted MSAL credential dict.
+    """
+
     def __init__(self, credentials_info: dict[str, Any]) -> None:
         self._credentials_info = credentials_info
         self._access_token: str = credentials_info.get("access_token", "")
         self._original_access_token: str = self._access_token
         self._refresh_token: str | None = credentials_info.get("refresh_token")
 
+    # ── Token management ───────────────────────────────────────────────────
+
     def _auth_headers(self) -> dict[str, str]:
         return {"Authorization": f"Bearer {self._access_token}"}
 
     async def _refresh_access_token(self) -> None:
+        """Use MSAL to exchange the refresh token for a fresh access token.
+
+        Updates ``self._access_token`` and ``self._credentials_info`` in-place.
+
+        Raises:
+            RuntimeError: MSAL reports an auth error.
+        """
         import msal
 
         app = msal.ConfidentialClientApplication(
@@ -109,6 +164,7 @@ class MSGraphClient:
             raise RuntimeError(f"MS Graph token refresh failed: {error}")
 
         self._access_token = result["access_token"]
+        # MSAL may issue a new refresh token.
         if "refresh_token" in result:
             self._refresh_token = result["refresh_token"]
             self._credentials_info["refresh_token"] = result["refresh_token"]
@@ -116,10 +172,16 @@ class MSGraphClient:
 
     @property
     def refreshed_credentials(self) -> dict[str, Any] | None:
+        """Return updated credential dict if the access token was refreshed.
+
+        Returns ``None`` if no change was made.
+        """
         if self._access_token != self._original_access_token:
             return {**self._credentials_info, "access_token": self._access_token}
         return None
 
+    # ── HTTP helpers ───────────────────────────────────────────────────────
+
     async def _get(
         self,
         client: httpx.AsyncClient,
@@ -128,8 +190,10 @@ class MSGraphClient:
         *,
         retry_on_401: bool = True,
     ) -> dict[str, Any]:
+        """GET *url* with auth; refresh token on 401 and retry once."""
         resp = await client.get(url, params=params, headers=self._auth_headers())
         if resp.status_code == 401 and retry_on_401 and self._refresh_token:
+            logger.debug("ms_graph: 401 on %s — refreshing token", url)
             await self._refresh_access_token()
             resp = await client.get(url, params=params, headers=self._auth_headers())
         if resp.status_code == 429:
@@ -137,11 +201,22 @@ class MSGraphClient:
         resp.raise_for_status()
         return resp.json()
 
+    # ── Public API ─────────────────────────────────────────────────────────
+
     async def fetch_emails(
         self,
         filter_config: dict[str, Any] | None = None,
         since: datetime | None = None,
     ) -> list[EmailMessage]:
+        """Return up to ``_MAX_EMAILS`` Outlook messages matching *filter_config*.
+
+        Parameters
+        ----------
+        filter_config:
+            Optional dict with ``senders``, ``date_range``, ``folders`` keys.
+        since:
+            Hard lower-bound on email date (from last agent run).
+        """
         odata_filter = _build_email_filter(filter_config, since)
         params: dict[str, Any] = {
             "$top": 50,
@@ -162,7 +237,7 @@ class MSGraphClient:
                     if len(emails) >= _MAX_EMAILS:
                         break
                 url = data.get("@odata.nextLink", "")
-                params = {}
+                params = {}  # nextLink already contains encoded params.
 
         logger.info("ms_graph: fetched %d Outlook email(s)", len(emails))
         return emails
@@ -172,6 +247,13 @@ class MSGraphClient:
         filter_config: dict[str, Any] | None = None,
         since: datetime | None = None,
     ) -> list[ChatMessage]:
+        """Return up to ``_MAX_MESSAGES`` Teams messages matching *filter_config*.
+
+        Fetches from ``/me/chats/getAllMessages`` (personal + group chats).
+        The ``filter_config.channels`` key is checked as a text-filter on
+        the channel name post-fetch (the API doesn't support channel OData
+        filter directly on ``getAllMessages``).
+        """
         cfg = filter_config or {}
         channel_filter: list[str] = [c.lower() for c in cfg.get("channels", [])]
         params: dict[str, Any] = {"$top": 50}
@@ -186,9 +268,11 @@ class MSGraphClient:
                 try:
                     data = await self._get(client, url, params if url.startswith(_GRAPH_BASE) else None)
                 except httpx.HTTPStatusError as exc:
+                    # getAllMessages requires specific licensing; degrade gracefully.
                     if exc.response.status_code in (403, 404):
                         logger.warning(
-                            "ms_graph: /me/chats/getAllMessages not available (%d)",
+                            "ms_graph: /me/chats/getAllMessages not available (%d) — "
+                            "check Teams license or permissions",
                             exc.response.status_code,
                         )
                         break
@@ -208,6 +292,8 @@ class MSGraphClient:
         logger.info("ms_graph: fetched %d Teams message(s)", len(messages))
         return messages
 
+    # ── Parsers ────────────────────────────────────────────────────────────
+
     @staticmethod
     def _parse_email(item: dict[str, Any]) -> EmailMessage:
         subject: str = item.get("subject", "(no subject)") or "(no subject)"

app/main.py

@@ -0,0 +1,242 @@
+from contextlib import asynccontextmanager
+import logging
+
+from fastapi import FastAPI
+from fastapi.middleware.cors import CORSMiddleware
+
+from app.api.middleware.rate_limit import TierRateLimitMiddleware
+from app.api.middleware.sanitizer import SanitizerMiddleware
+from app.config.settings import settings
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+)
+logging.getLogger("sqlalchemy.engine").setLevel(logging.WARNING)
+logging.getLogger("sqlalchemy.pool").setLevel(logging.WARNING)
+
+
+async def _memory_audit_cron_tick() -> None:
+    """Weekly cron: contradiction scan + label canonicalization for all users (Phase 7)."""
+    import logging  # noqa: PLC0415
+    _log = logging.getLogger(__name__)
+    _log.info("memory audit cron tick: starting")
+    try:
+        from app.db import async_session  # noqa: PLC0415
+        from app.core.memory_maintenance import audit_memory  # noqa: PLC0415
+        from app.models import User  # noqa: PLC0415
+        from sqlalchemy import select  # noqa: PLC0415
+
+        async with async_session() as db:
+            result = await db.execute(select(User.id))
+            user_ids: list[str] = list(result.scalars().all())
+
+        for uid in user_ids:
+            try:
+                async with async_session() as db:
+                    await audit_memory(db, uid)
+            except Exception as exc:
+                _log.warning("memory audit cron tick: audit_memory failed user=%s: %s", uid, exc)
+
+        _log.info("memory audit cron tick: done users=%d", len(user_ids))
+    except Exception as exc:
+        _log.warning("memory audit cron tick: failed: %s", exc)
+
+
+async def _memory_cron_tick() -> None:
+    """Hourly cron: drain Free-tier extraction queue + mine proactive patterns for Power+ users."""
+    import logging  # noqa: PLC0415
+    _log = logging.getLogger(__name__)
+    _log.info("memory cron tick: starting")
+    try:
+        from app.db import async_session  # noqa: PLC0415
+        from app.core.memory_maintenance import drain_extraction_queue, mine_proactive_patterns  # noqa: PLC0415
+        from app.billing.tier_manager import tier_manager  # noqa: PLC0415
+        from app.models import User  # noqa: PLC0415
+        from sqlalchemy import select  # noqa: PLC0415
+
+        async with async_session() as db:
+            await drain_extraction_queue(db)
+
+        # mine proactive patterns for every Power+ user
+        async with async_session() as db:
+            result = await db.execute(select(User.id))
+            user_ids: list[str] = list(result.scalars().all())
+
+        for uid in user_ids:
+            try:
+                async with async_session() as db:
+                    tier = await tier_manager.get_tier(uid, db)
+                    if tier_manager.check_feature(tier, "proactive_mining"):
+                        await mine_proactive_patterns(db, uid)
+            except Exception as exc:
+                _log.warning("memory cron tick: mine_proactive_patterns failed user=%s: %s", uid, exc)
+
+        _log.info("memory cron tick: done users=%d", len(user_ids))
+    except Exception as exc:
+        _log.warning("memory cron tick: failed: %s", exc)
+
+
+async def _scout_cron_tick() -> None:
+    """Every-15-min cron: poll enabled cloud scouts (cron-fallback; push is primary).
+
+    Skips any scout whose ``last_run_at`` is within the last 5 minutes so
+    a push notification and the fallback cron don't double-fire within the
+    same window.
+    """
+    import logging  # noqa: PLC0415
+    import uuid  # noqa: PLC0415
+    from datetime import datetime, timezone  # noqa: PLC0415
+
+    _log = logging.getLogger(__name__)
+    _log.info("scout cron tick: starting")
+    try:
+        from app.db import async_session  # noqa: PLC0415
+        from app.models import CloudScoutConfig  # noqa: PLC0415
+        from app.scouts.engine import ScoutEngine  # noqa: PLC0415
+        from sqlalchemy import select  # noqa: PLC0415
+
+        async with async_session() as session:
+            scouts = (await session.execute(
+                select(CloudScoutConfig).where(CloudScoutConfig.enabled == True)  # noqa: E712
+            )).scalars().all()
+
+        engine = ScoutEngine()
+        triggered = 0
+        for scout in scouts:
+            # Rate-limit guard: push is primary; skip if ran within 5 minutes.
+            if scout.last_run_at:
+                elapsed = (datetime.now(tz=timezone.utc) - scout.last_run_at).total_seconds()
+                if elapsed < 300:
+                    continue
+            try:
+                await engine.trigger_scout(uuid.UUID(str(scout.id)))
+                triggered += 1
+            except Exception as exc:
+                _log.warning("scout cron tick: trigger failed scout=%s: %s", scout.id, exc)
+
+        _log.info("scout cron tick: done triggered=%d total=%d", triggered, len(scouts))
+    except Exception as exc:
+        _log.warning("scout cron tick: failed: %s", exc)
+
+
+async def _scout_watch_renewal_tick() -> None:
+    """Every-24-hour cron: re-issue Gmail users.watch for scouts expiring within 24h.
+
+    Handles missing or misconfigured connectors gracefully — logs and continues.
+    """
+    import logging  # noqa: PLC0415
+    from datetime import datetime, timedelta, timezone  # noqa: PLC0415
+
+    _log = logging.getLogger(__name__)
+    _log.info("scout watch renewal tick: starting")
+    try:
+        from app.db import async_session  # noqa: PLC0415
+        from app.models import CloudScoutConfig  # noqa: PLC0415
+        from app.scouts.connectors.registry import get_connector  # noqa: PLC0415
+        from sqlalchemy import select  # noqa: PLC0415
+
+        threshold = datetime.now(tz=timezone.utc) + timedelta(hours=24)
+        renewed = 0
+        async with async_session() as session:
+            scouts = (await session.execute(
+                select(CloudScoutConfig).where(
+                    CloudScoutConfig.enabled == True,  # noqa: E712
+                    CloudScoutConfig.provider == "gmail",
+                    CloudScoutConfig.gmail_watch_expires_at <= threshold,
+                )
+            )).scalars().all()
+
+            for scout in scouts:
+                try:
+                    connector = get_connector("gmail")
+                    await connector.renew_watch(scout)
+                    renewed += 1
+                except Exception:
+                    _log.exception("scout watch renewal tick: renew failed scout=%s", scout.id)
+
+            await session.commit()
+
+        _log.info("scout watch renewal tick: done renewed=%d", renewed)
+    except Exception as exc:
+        _log.warning("scout watch renewal tick: failed: %s", exc)
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    # Startup: register source connectors.
+    from app.scouts.connectors.gmail import GmailConnector  # noqa: PLC0415
+    from app.scouts.connectors.registry import register_connector  # noqa: PLC0415
+    register_connector(GmailConnector())
+
+    # Startup: ensure agent tool modules are loaded.
+    import app.agents  # noqa: F401
+
+    scheduler = None
+    if settings.SCHEDULER_ENABLED:
+        from apscheduler.schedulers.asyncio import AsyncIOScheduler  # noqa: PLC0415
+
+        scheduler = AsyncIOScheduler()
+        scheduler.add_job(_memory_cron_tick, "interval", hours=1, id="memory_cron")
+        scheduler.add_job(_memory_audit_cron_tick, "interval", weeks=1, id="memory_audit_cron")
+        scheduler.add_job(
+            _scout_cron_tick, "interval", minutes=15,
+            id="scout_cron_tick", replace_existing=True,
+        )
+        scheduler.add_job(
+            _scout_watch_renewal_tick, "interval", hours=24,
+            id="scout_watch_renewal_tick", replace_existing=True,
+        )
+        scheduler.start()
+        logging.getLogger(__name__).info("memory cron scheduler started (interval=1h)")
+
+    yield
+
+    if scheduler is not None:
+        scheduler.shutdown(wait=False)
+
+    # Shutdown: dispose SQLAlchemy connection pool
+    from app.db import engine
+    await engine.dispose()
+
+
+def create_app() -> FastAPI:
+    app = FastAPI(
+        title="AdiuvAI Cloud API",
+        version="0.1.0",
+        docs_url="/docs" if settings.ENV == "dev" else None,
+        redoc_url=None,
+        lifespan=lifespan,
+    )
+
+    app.add_middleware(
+        CORSMiddleware,
+        allow_origins=settings.CORS_ORIGINS,
+        allow_credentials=True,
+        allow_methods=["*"],
+        allow_headers=["*"],
+    )
+    # Middleware stack (Starlette inserts at position 0, so last-added = outermost).
+    # Request flow:  TierRateLimit → Sanitizer → CORS → Router
+    # Response flow: Router → CORS → Sanitizer → TierRateLimit
+    app.add_middleware(SanitizerMiddleware)
+    app.add_middleware(TierRateLimitMiddleware)
+
+    from app.api.routes import scouts, auth, billing, chat, device_ws, memory, scout_webhooks
+
+    app.include_router(auth.router,           prefix="/api/v1")
+    app.include_router(chat.router,           prefix="/api/v1")
+    app.include_router(billing.router,        prefix="/api/v1")
+    app.include_router(scouts.router,         prefix="/api/v1")
+    app.include_router(scout_webhooks.router, prefix="/api/v1")
+    app.include_router(device_ws.router,      prefix="/api/v1")
+    app.include_router(memory.router,         prefix="/api/v1")
+
+    @app.get("/api/v1/health", tags=["health"])
+    async def health() -> dict:
+        return {"status": "ok", "version": app.version}
+
+    return app
+
+
+app = create_app()

app/models.py

@@ -1,14 +1,20 @@
 """SQLAlchemy ORM models for all persistent tables.
 
-Centralized here so that Alembic migrations and all services share
+Only auth, billing, scout config, and memory data live here.
-the same model definitions.  Each service only queries the tables it owns.
+User content (notes, tasks, etc.) lives exclusively on the client.
 
-Ownership:
+Table inventory:
-  Auth Service      → users, refresh_tokens, subscriptions
+  users               — account credentials + tier
-  Chat Service      → memory_core, memory_associative, memory_episodic, memory_proactive
+  refresh_tokens      — hashed refresh token store
-  Batch Agent       → local_agent_configs, cloud_agent_configs, agent_run_logs
+  subscriptions       — Stripe subscription records
-  Billing Service   → subscriptions (shared write with Auth)
+  local_scout_configs — per-device batch scout configs
-  (excluded MVP)    → storage_records, backup_metadata, plugins, plugin_*, revenue_events
+  cloud_scout_configs — OAuth-backed cloud scout configs
+  scout_run_logs      — execution history for all scouts
+  memory_core         — per-user persistent key/value preferences (encrypted)
+  memory_associative  — per-user semantic memory with embeddings (encrypted)
+  memory_episodic     — per-user session summaries (encrypted)
+  memory_proactive    — per-user behavioral patterns (encrypted)
+  memory_relations    — per-user entity/relation graph (Mem0g-light, Phase 3)
 """
 
 from __future__ import annotations
@@ -16,8 +22,8 @@ from __future__ import annotations
 import uuid
 from datetime import datetime, timezone
 
+from pgvector.sqlalchemy import Vector
 from sqlalchemy import (
-    BigInteger,
     Boolean,
     DateTime,
     Enum,
@@ -25,15 +31,17 @@ from sqlalchemy import (
     ForeignKey,
     Integer,
     JSON,
+    LargeBinary,
     String,
     Text,
     UniqueConstraint,
     Uuid,
     func,
+    text,
 )
 from sqlalchemy.orm import Mapped, mapped_column, relationship
 
-from shared.db import Base
+from app.db import Base
 
 # ── Helpers ──────────────────────────────────────────────────────────────
 
@@ -49,14 +57,12 @@ def _now() -> datetime:
 # ── Enum types ────────────────────────────────────────────────────────────
 
 TierEnum = Enum("free", "pro", "power", "team", name="billing_tier")
-PluginStatusEnum = Enum("pending_review", "approved", "rejected", name="plugin_status")
-ReviewDecisionEnum = Enum("approved", "rejected", name="review_decision")
 AgentTypeEnum = Enum("local", "cloud", name="agent_type")
 AgentStatusEnum = Enum("running", "success", "error", "partial", name="agent_run_status")
 CloudProviderEnum = Enum("gmail", "teams", "outlook", name="cloud_provider")
 
 
-# ── Auth models ───────────────────────────────────────────────────────────
+# ── Models ────────────────────────────────────────────────────────────────
 
 
 class User(Base):
@@ -68,13 +74,19 @@ class User(Base):
     email: Mapped[str] = mapped_column(String(255), unique=True, nullable=False, index=True)
     name: Mapped[str | None] = mapped_column(String(100), nullable=True)
     surname: Mapped[str | None] = mapped_column(String(100), nullable=True)
-    password_hash: Mapped[str] = mapped_column(String(255), nullable=False)
+    password_hash: Mapped[str | None] = mapped_column(String(255), nullable=True)
+    avatar_url: Mapped[str | None] = mapped_column(Text, nullable=True)
     tier: Mapped[str] = mapped_column(TierEnum, nullable=False, default="free")
     stripe_customer_id: Mapped[str | None] = mapped_column(String(255), nullable=True)
+    # Per-user Fernet key (base64-urlsafe, 44 chars). Generated on registration.
+    # Used to encrypt/decrypt all memory rows for this user.
     encryption_key: Mapped[str | None] = mapped_column(String(64), nullable=True)
     created_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()
     )
+    onboarding_completed_at: Mapped[datetime | None] = mapped_column(
+        DateTime(timezone=True), nullable=True, default=None
+    )
     updated_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now(), onupdate=func.now()
     )
@@ -85,6 +97,9 @@ class User(Base):
     subscription: Mapped[Subscription | None] = relationship(
         back_populates="user", uselist=False, cascade="all, delete-orphan"
     )
+    oauth_accounts: Mapped[list[OAuthAccount]] = relationship(
+        back_populates="user", cascade="all, delete-orphan"
+    )
 
 
 class RefreshToken(Base):
@@ -105,6 +120,25 @@ class RefreshToken(Base):
     user: Mapped[User] = relationship(back_populates="refresh_tokens")
 
 
+class OAuthAccount(Base):
+    __tablename__ = "oauth_accounts"
+
+    id: Mapped[str] = mapped_column(
+        Uuid(as_uuid=False), primary_key=True, default=_uuid
+    )
+    user_id: Mapped[str] = mapped_column(
+        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"), nullable=False, index=True
+    )
+    provider: Mapped[str] = mapped_column(String(50), nullable=False)
+    provider_user_id: Mapped[str] = mapped_column(String(255), nullable=False)
+    provider_email: Mapped[str | None] = mapped_column(String(255), nullable=True)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+
+    user: Mapped[User] = relationship(back_populates="oauth_accounts")
+
+
 class Subscription(Base):
     __tablename__ = "subscriptions"
 
@@ -126,161 +160,8 @@ class Subscription(Base):
     user: Mapped[User] = relationship(back_populates="subscription")
 
 
-# ── Storage models (excluded from MVP, kept for Alembic) ──────────────
+class LocalScoutConfig(Base):
-
+    __tablename__ = "local_scout_configs"
-
-class StorageRecord(Base):
-    __tablename__ = "storage_records"
-
-    id: Mapped[str] = mapped_column(
-        Uuid(as_uuid=False), primary_key=True, default=_uuid
-    )
-    user_id: Mapped[str] = mapped_column(
-        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"), nullable=False, index=True
-    )
-    table_name: Mapped[str] = mapped_column(String(100), nullable=False)
-    s3_key: Mapped[str] = mapped_column(String(500), nullable=False)
-    checksum: Mapped[str] = mapped_column(String(64), nullable=False)
-    size_bytes: Mapped[int] = mapped_column(Integer, nullable=False)
-    created_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )
-    updated_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now(), onupdate=func.now()
-    )
-
-
-class BackupMetadata(Base):
-    __tablename__ = "backup_metadata"
-
-    id: Mapped[str] = mapped_column(
-        Uuid(as_uuid=False), primary_key=True, default=_uuid
-    )
-    user_id: Mapped[str] = mapped_column(
-        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"), nullable=False, index=True
-    )
-    s3_key: Mapped[str] = mapped_column(String(500), nullable=False)
-    version: Mapped[int] = mapped_column(Integer, nullable=False)
-    timestamp: Mapped[int] = mapped_column(BigInteger, nullable=False)
-    checksum: Mapped[str] = mapped_column(String(64), nullable=False)
-    size_bytes: Mapped[int] = mapped_column(Integer, nullable=False)
-    created_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )
-
-
-# ── Plugin models (excluded from MVP, kept for Alembic) ───────────────
-
-
-class Plugin(Base):
-    __tablename__ = "plugins"
-
-    id: Mapped[str] = mapped_column(String(255), primary_key=True)
-    name: Mapped[str] = mapped_column(String(255), nullable=False)
-    description: Mapped[str] = mapped_column(Text, nullable=False, default="")
-    version: Mapped[str] = mapped_column(String(50), nullable=False, default="1.0.0")
-    author_id: Mapped[str | None] = mapped_column(
-        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="SET NULL"), nullable=True
-    )
-    author_name: Mapped[str] = mapped_column(String(255), nullable=False, default="")
-    category: Mapped[str] = mapped_column(String(100), nullable=False, default="")
-    price_cents: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
-    permissions: Mapped[str] = mapped_column(Text, nullable=False, default="[]")
-    status: Mapped[str] = mapped_column(PluginStatusEnum, nullable=False, default="pending_review")
-    s3_package_key: Mapped[str | None] = mapped_column(String(500), nullable=True)
-    install_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
-    avg_rating: Mapped[float] = mapped_column(Float, nullable=False, default=0.0)
-    rejection_reason: Mapped[str | None] = mapped_column(Text, nullable=True)
-    submitted_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )
-    created_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )
-
-    installations: Mapped[list[PluginInstallation]] = relationship(
-        back_populates="plugin", cascade="all, delete-orphan"
-    )
-    reviews: Mapped[list[PluginReview]] = relationship(
-        back_populates="plugin", cascade="all, delete-orphan"
-    )
-    revenue_events: Mapped[list[RevenueEvent]] = relationship(
-        back_populates="plugin", cascade="all, delete-orphan"
-    )
-
-
-class PluginInstallation(Base):
-    __tablename__ = "plugin_installations"
-    __table_args__ = (UniqueConstraint("plugin_id", "user_id", name="uq_plugin_user"),)
-
-    id: Mapped[str] = mapped_column(
-        Uuid(as_uuid=False), primary_key=True, default=_uuid
-    )
-    plugin_id: Mapped[str] = mapped_column(
-        String(255), ForeignKey("plugins.id", ondelete="CASCADE"), nullable=False, index=True
-    )
-    user_id: Mapped[str] = mapped_column(
-        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"), nullable=False, index=True
-    )
-    installed_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )
-
-    plugin: Mapped[Plugin] = relationship(back_populates="installations")
-
-
-class PluginReview(Base):
-    __tablename__ = "plugin_reviews"
-
-    id: Mapped[str] = mapped_column(
-        Uuid(as_uuid=False), primary_key=True, default=_uuid
-    )
-    plugin_id: Mapped[str] = mapped_column(
-        String(255), ForeignKey("plugins.id", ondelete="CASCADE"), nullable=False, index=True
-    )
-    reviewer_id: Mapped[str | None] = mapped_column(
-        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="SET NULL"), nullable=True
-    )
-    decision: Mapped[str] = mapped_column(ReviewDecisionEnum, nullable=False)
-    notes: Mapped[str | None] = mapped_column(Text, nullable=True)
-    reviewed_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )
-    created_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )
-
-    plugin: Mapped[Plugin] = relationship(back_populates="reviews")
-
-
-class RevenueEvent(Base):
-    __tablename__ = "revenue_events"
-
-    id: Mapped[str] = mapped_column(
-        Uuid(as_uuid=False), primary_key=True, default=_uuid
-    )
-    plugin_id: Mapped[str] = mapped_column(
-        String(255), ForeignKey("plugins.id", ondelete="CASCADE"), nullable=False, index=True
-    )
-    user_id: Mapped[str] = mapped_column(
-        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"), nullable=False, index=True
-    )
-    amount_cents: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
-    developer_share_cents: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
-    stripe_transfer_id: Mapped[str | None] = mapped_column(String(255), nullable=True)
-    paid_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
-    created_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )
-
-    plugin: Mapped[Plugin] = relationship(back_populates="revenue_events")
-
-
-# ── Agent models ──────────────────────────────────────────────────────────
-
-
-class LocalAgentConfig(Base):
-    __tablename__ = "local_agent_configs"
 
     id: Mapped[str] = mapped_column(
         Uuid(as_uuid=False), primary_key=True, default=_uuid
@@ -293,6 +174,7 @@ class LocalAgentConfig(Base):
     directory_paths: Mapped[list] = mapped_column(JSON, nullable=False, default=list)
     data_types: Mapped[list] = mapped_column(JSON, nullable=False, default=list)
     prompt_template: Mapped[str] = mapped_column(Text, nullable=False, default="")
+    scout_config: Mapped[dict | None] = mapped_column(JSON, nullable=True)
     file_extensions: Mapped[list] = mapped_column(JSON, nullable=False, default=list)
     schedule_cron: Mapped[str] = mapped_column(String(100), nullable=False, default="0 */6 * * *")
     enabled: Mapped[bool] = mapped_column(Boolean, nullable=False, default=True)
@@ -304,17 +186,17 @@ class LocalAgentConfig(Base):
         DateTime(timezone=True), nullable=False, server_default=func.now(), onupdate=func.now()
     )
 
-    run_logs: Mapped[list[AgentRunLog]] = relationship(
+    run_logs: Mapped[list["ScoutRunLog"]] = relationship(
-        back_populates="local_agent",
+        back_populates="local_scout",
-        primaryjoin="and_(AgentRunLog.agent_id == LocalAgentConfig.id, AgentRunLog.agent_type == 'local')",
+        primaryjoin="and_(ScoutRunLog.scout_id == LocalScoutConfig.id, ScoutRunLog.scout_type == 'local')",
-        foreign_keys="AgentRunLog.agent_id",
+        foreign_keys="ScoutRunLog.scout_id",
         cascade="all, delete-orphan",
-        overlaps="run_logs,cloud_agent",
+        overlaps="run_logs,cloud_scout",
     )
 
 
-class CloudAgentConfig(Base):
+class CloudScoutConfig(Base):
-    __tablename__ = "cloud_agent_configs"
+    __tablename__ = "cloud_scout_configs"
 
     id: Mapped[str] = mapped_column(
         Uuid(as_uuid=False), primary_key=True, default=_uuid
@@ -337,55 +219,97 @@ class CloudAgentConfig(Base):
     updated_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now(), onupdate=func.now()
     )
+    auto_trash_spam: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False, server_default=text("false"))
+    gmail_history_id: Mapped[str | None] = mapped_column(String(64), nullable=True)
+    gmail_watch_expires_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
+    device_inactivity_pause_days: Mapped[int] = mapped_column(Integer, nullable=False, default=14, server_default="14")
 
-    run_logs: Mapped[list[AgentRunLog]] = relationship(
+    run_logs: Mapped[list["ScoutRunLog"]] = relationship(
-        back_populates="cloud_agent",
+        back_populates="cloud_scout",
-        primaryjoin="and_(AgentRunLog.agent_id == CloudAgentConfig.id, AgentRunLog.agent_type == 'cloud')",
+        primaryjoin="and_(ScoutRunLog.scout_id == CloudScoutConfig.id, ScoutRunLog.scout_type == 'cloud')",
-        foreign_keys="AgentRunLog.agent_id",
+        foreign_keys="ScoutRunLog.scout_id",
         cascade="all, delete-orphan",
-        overlaps="run_logs,local_agent",
+        overlaps="run_logs,local_scout",
     )
 
 
-class AgentRunLog(Base):
+class ScoutTriageQueue(Base):
-    __tablename__ = "agent_run_logs"
+    __tablename__ = "scout_triage_queue"
+    __table_args__ = (
+        UniqueConstraint("scout_id", "source_msg_ref", name="uq_scout_triage_queue_scout_msg"),
+    )
+
+    id: Mapped[str] = mapped_column(Uuid(as_uuid=False), primary_key=True, default=_uuid)
+    user_id: Mapped[str] = mapped_column(Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"), nullable=False, index=True)
+    scout_id: Mapped[str] = mapped_column(Uuid(as_uuid=False), ForeignKey("cloud_scout_configs.id", ondelete="CASCADE"), nullable=False)
+    source_type: Mapped[str] = mapped_column(String(50), nullable=False)
+    source_msg_ref: Mapped[str] = mapped_column(String(255), nullable=False)
+    triage_verdict: Mapped[str] = mapped_column(String(20), nullable=False)
+    triage_reason: Mapped[str | None] = mapped_column(Text, nullable=True)
+    status: Mapped[str] = mapped_column(String(20), nullable=False, default="queued", server_default="queued")
+    triaged_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), nullable=False, server_default=func.now())
+    delivered_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
+    acked_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
+    expires_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), nullable=False)
+
+
+class ScoutRunLog(Base):
+    __tablename__ = "scout_run_logs"
 
     id: Mapped[str] = mapped_column(
         Uuid(as_uuid=False), primary_key=True, default=_uuid
     )
-    agent_id: Mapped[str] = mapped_column(String(255), nullable=False, index=True)
+    # Plain string — not a FK because it references either local_scout_configs or cloud_scout_configs
-    agent_type: Mapped[str] = mapped_column(AgentTypeEnum, nullable=False)
+    # depending on scout_type. Query by (scout_id, scout_type) to locate the source config.
+    scout_id: Mapped[str] = mapped_column(String(255), nullable=False, index=True)
+    scout_type: Mapped[str] = mapped_column(AgentTypeEnum, nullable=False)
     user_id: Mapped[str] = mapped_column(
         Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"), nullable=False, index=True
     )
     status: Mapped[str] = mapped_column(AgentStatusEnum, nullable=False, default="running")
     items_processed: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
     items_created: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    tokens_used: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
     errors: Mapped[list | None] = mapped_column(JSON, nullable=True)
     started_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()
     )
     completed_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
 
-    local_agent: Mapped[LocalAgentConfig | None] = relationship(
+    local_scout: Mapped["LocalScoutConfig | None"] = relationship(
         back_populates="run_logs",
-        primaryjoin="and_(AgentRunLog.agent_id == LocalAgentConfig.id, AgentRunLog.agent_type == 'local')",
+        primaryjoin="and_(ScoutRunLog.scout_id == LocalScoutConfig.id, ScoutRunLog.scout_type == 'local')",
-        foreign_keys="AgentRunLog.agent_id",
+        foreign_keys="ScoutRunLog.scout_id",
-        overlaps="run_logs,cloud_agent",
+        overlaps="run_logs,cloud_scout",
     )
-    cloud_agent: Mapped[CloudAgentConfig | None] = relationship(
+    cloud_scout: Mapped["CloudScoutConfig | None"] = relationship(
         back_populates="run_logs",
-        primaryjoin="and_(AgentRunLog.agent_id == CloudAgentConfig.id, AgentRunLog.agent_type == 'cloud')",
+        primaryjoin="and_(ScoutRunLog.scout_id == CloudScoutConfig.id, ScoutRunLog.scout_type == 'cloud')",
-        foreign_keys="AgentRunLog.agent_id",
+        foreign_keys="ScoutRunLog.scout_id",
-        overlaps="run_logs,local_agent",
+        overlaps="run_logs,local_scout",
     )
 
 
-# ── Memory models ─────────────────────────────────────────────────────────
+class MonthlyTokenUsage(Base):
+    __tablename__ = "monthly_token_usage"
+
+    user_id: Mapped[str] = mapped_column(
+        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"), primary_key=True
+    )
+    year_month: Mapped[str] = mapped_column(String(7), primary_key=True)  # 'YYYY-MM'
+    feature: Mapped[str] = mapped_column(String(64), primary_key=True)
+    tokens_used: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
+
+
+# ── Memory models ─────────────────────────────────────────────────────────────
 
 
 class MemoryCore(Base):
-    """Per-user persistent key/value preferences, encrypted at rest."""
+    """Per-user persistent key/value preferences, encrypted at rest.
+
+    Examples: preferred_language, timezone, work_style.
+    Decrypted in-memory only using User.encryption_key.
+    """
 
     __tablename__ = "memory_core"
 
@@ -402,7 +326,11 @@ class MemoryCore(Base):
 
 
 class MemoryAssociative(Base):
-    """Per-user semantic memory: encrypted content + pgvector embedding."""
+    """Per-user semantic memory: encrypted content + pgvector embedding for similarity search.
+
+    Production: ``embedding`` column is ``vector(1536)`` via pgvector.
+    Tests (SQLite): stored as JSON list.
+    """
 
     __tablename__ = "memory_associative"
 
@@ -412,7 +340,8 @@ class MemoryAssociative(Base):
         nullable=False, index=True,
     )
     content_encrypted: Mapped[str] = mapped_column(Text, nullable=False)
-    embedding: Mapped[list | None] = mapped_column(JSON, nullable=True)
+    # vector(1536) via pgvector; SQLite tests use NULL embeddings so no dialect issue.
+    embedding: Mapped[list | None] = mapped_column(Vector(1536), nullable=True)
     entity_type: Mapped[str | None] = mapped_column(String(100), nullable=True)
     entity_id: Mapped[str | None] = mapped_column(String(255), nullable=True)
     updated_at: Mapped[datetime] = mapped_column(
@@ -421,7 +350,10 @@ class MemoryAssociative(Base):
 
 
 class MemoryEpisodic(Base):
-    """Per-user session summaries, encrypted at rest."""
+    """Per-user session summaries, encrypted at rest.
+
+    One row per session interaction; used to recall recent conversations.
+    """
 
     __tablename__ = "memory_episodic"
 
@@ -438,7 +370,11 @@ class MemoryEpisodic(Base):
 
 
 class MemoryProactive(Base):
-    """Per-user inferred behavioral patterns, encrypted at rest."""
+    """Per-user inferred behavioral patterns, encrypted at rest.
+
+    Confidence in [0.0, 1.0]; only patterns above threshold are injected.
+    Source: 'inferred' (from episodes) or 'explicit' (user-stated).
+    """
 
     __tablename__ = "memory_proactive"
 
@@ -453,3 +389,85 @@ class MemoryProactive(Base):
     created_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()
     )
+
+
+class ExtractionQueue(Base):
+    """Batch extraction queue for Free-tier users (Phase 2).
+
+    Pro/Power/Team users get realtime asyncio.create_task() extraction.
+    Free users get a queue row here; a daily cron (Phase 5) drains it.
+    """
+
+    __tablename__ = "extraction_queue"
+
+    id: Mapped[str] = mapped_column(Uuid(as_uuid=False), primary_key=True, default=_uuid)
+    user_id: Mapped[str] = mapped_column(
+        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"),
+        nullable=False, index=True,
+    )
+    episode_id: Mapped[str | None] = mapped_column(
+        Uuid(as_uuid=False), nullable=True,
+    )
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+
+
+class MemoryRelation(Base):
+    """Per-user entity/relation graph row (Mem0g-light, Phase 3).
+
+    subject_label/object_label are plaintext entity identifiers (not user content).
+    notes_encrypted is optional Fernet-encrypted per-user commentary.
+    confidence in [0.0, 1.0] — decays 5 % per 30 days since last_confirmed_at.
+    """
+
+    __tablename__ = "memory_relations"
+
+    id: Mapped[str] = mapped_column(Uuid(as_uuid=False), primary_key=True, default=_uuid)
+    user_id: Mapped[str] = mapped_column(
+        Uuid(as_uuid=False), ForeignKey("users.id", ondelete="CASCADE"),
+        nullable=False, index=True,
+    )
+    subject_label: Mapped[str] = mapped_column(String(128), nullable=False)
+    subject_type: Mapped[str] = mapped_column(String(32), nullable=False)
+    predicate: Mapped[str] = mapped_column(String(64), nullable=False)
+    object_label: Mapped[str] = mapped_column(String(128), nullable=False)
+    object_type: Mapped[str] = mapped_column(String(32), nullable=False)
+    confidence: Mapped[float] = mapped_column(Float, nullable=False, default=0.7)
+    source_episode_id: Mapped[str | None] = mapped_column(
+        Uuid(as_uuid=False),
+        ForeignKey("memory_episodic.id", ondelete="SET NULL"),
+        nullable=True,
+    )
+    notes_encrypted: Mapped[bytes | None] = mapped_column(LargeBinary, nullable=True)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )
+    updated_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now(), onupdate=func.now()
+    )
+    last_confirmed_at: Mapped[datetime | None] = mapped_column(
+        DateTime(timezone=True), nullable=True
+    )
+
+
+class Plugin(Base):
+    """Plugin marketplace catalog entry."""
+
+    __tablename__ = "plugins"
+
+    id: Mapped[str] = mapped_column(String(255), primary_key=True)
+    name: Mapped[str] = mapped_column(String(255), nullable=False)
+    description: Mapped[str] = mapped_column(Text, nullable=False)
+    version: Mapped[str] = mapped_column(String(50), nullable=False)
+    author_name: Mapped[str] = mapped_column(String(255), nullable=False)
+    category: Mapped[str] = mapped_column(String(100), nullable=False)
+    price_cents: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    permissions: Mapped[str] = mapped_column(Text, nullable=False, default="[]")
+    status: Mapped[str] = mapped_column(String(50), nullable=False, default="pending")
+    s3_package_key: Mapped[str | None] = mapped_column(String(500), nullable=True)
+    install_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    avg_rating: Mapped[float] = mapped_column(Float, nullable=False, default=0.0)
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )

app/schemas/__init__.py

@@ -1,7 +1,6 @@
 """Pydantic schemas — API request/response contracts.
 
-Shared across all services. Mirrors the TypeScript types from
+Mirrors the TypeScript types from the Electron app (src/shared/api-types.ts).
-the Electron app (src/shared/api-types.ts).
 """
 
 from __future__ import annotations
@@ -31,6 +30,16 @@ class UserProfile(BaseModel):
     name: str | None = None
     surname: str | None = None
     tier: BillingTier
+    avatar_url: str | None = None
+    has_password: bool = True
+    onboarding_completed_at: int | None = None  # epoch ms, null = not onboarded
+    memory: dict[str, str] = Field(default_factory=dict)  # decrypted core memory k/v
+
+
+class OAuthAccountInfo(BaseModel):
+    provider: str
+    provider_email: str | None = None
+    created_at: int  # epoch ms
 
 
 # ── Chat ─────────────────────────────────────────────────────────────
@@ -51,88 +60,6 @@ class ChatResponse(BaseModel):
     response: str
 
 
-# ── Backup ───────────────────────────────────────────────────────────
-
-class BackupMetadata(BaseModel):
-    version: int
-    timestamp: int
-    checksum: str
-    chunk_count: int
-
-
-# ── Cloud Storage (E2E encrypted blobs) ──────────────────────────────
-
-class StorageRecord(BaseModel):
-    id: str
-    user_id: str
-    table: str
-    blob: bytes
-    checksum: str
-    created_at: int
-    updated_at: int
-
-
-class StorageRecordCreate(BaseModel):
-    table: str
-    blob: bytes
-    checksum: str
-
-
-class StorageRecordUpdate(BaseModel):
-    blob: bytes
-    checksum: str
-
-
-# ── Cloud Vector Store (E2E encrypted vectors) ────────────────────────
-
-class VectorItem(BaseModel):
-    id: str
-    blob: bytes
-    checksum: str
-
-
-class VectorUpsertRequest(BaseModel):
-    vectors: list[VectorItem]
-
-
-class VectorSearchRequest(BaseModel):
-    query_blob: bytes
-    top_k: int = 10
-
-
-class VectorSearchResult(BaseModel):
-    id: str
-    score: float
-    blob: bytes
-
-
-class VectorSearchResponse(BaseModel):
-    results: list[VectorSearchResult]
-
-
-# ── Plugin Marketplace ────────────────────────────────────────────────
-
-class PluginManifest(BaseModel):
-    id: str
-    name: str
-    description: str
-    version: str
-    author: str
-    permissions: list[str]
-    category: str
-    price_cents: int = 0
-
-
-class PluginListResponse(BaseModel):
-    plugins: list[PluginManifest]
-    total: int
-    page: int
-
-
-class PluginInstallRequest(BaseModel):
-    plugin_id: str
-
-
 # ── WebSocket Frame Protocol ──────────────────────────────────────────
 
 class WsFrameType(str, Enum):
@@ -146,11 +73,9 @@ class WsFrameType(str, Enum):
     device_hello = "device_hello"
     # ── v3 frame types ─────────────────────────────────────────────────
     home_request = "home_request"
-    floating_request = "floating_request"
     stream_start = "stream_start"
     stream_text = "stream_text"
     stream_end = "stream_end"
-    floating_domain = "floating_domain"
     data_request = "data_request"
     data_response = "data_response"
     mutation = "mutation"
@@ -158,6 +83,24 @@ class WsFrameType(str, Enum):
     journey_start = "journey_start"
     journey_message = "journey_message"
     journey_reply = "journey_reply"
+    # ── v5 brief frame types ──────────────────────────────────────────
+    brief_request = "brief_request"
+    # ── v6 task brief frame types ─────────────────────────────────────
+    task_brief_request = "task_brief_request"
+    # ── v7 folder index frame types ───────────────────────────────────
+    index_session_start = "index_session_start"
+    index_file_batch = "index_file_batch"
+    index_session_cancel = "index_session_cancel"
+    index_file_result = "index_file_result"
+    index_session_progress = "index_session_progress"
+    index_session_done = "index_session_done"
+    # ── v8 contextual sidebar frame types ────────────────────────────
+    contextual_request = "contextual_request"
+    contextual_scope_update = "contextual_scope_update"
+    contextual_scope_ack = "contextual_scope_ack"
+    # ── v9 scout proposal frame types ────────────────────────────────
+    SCOUT_PROPOSAL = "scout_proposal"
+    SCOUT_PROPOSAL_ACK = "scout_proposal_ack"
 
 
 class WsToolCall(BaseModel):
@@ -207,16 +150,20 @@ class WsDeviceHello(BaseModel):
 
     type: Literal[WsFrameType.device_hello] = WsFrameType.device_hello
     device_id: str
-    agent_ids: list[str] = Field(default_factory=list)
+    scout_ids: list[str] = Field(default_factory=list)
+
 
 
 # ── WebSocket v3 Frame Models ─────────────────────────────────────────
 
-class WsFloatingScope(BaseModel):
+class FormatPrefsModel(BaseModel):
-    """Scope for a floating request."""
+    """User display preferences sent by Electron on each request."""
 
-    type: Literal["task", "project", "note", "timeline"]
+    timezone: str = "UTC"
-    id: str | None = None
+    date_format: str = "dd/MM/yyyy"
+    time_format: str = "24h"
+    locale: str = "en-US"
+    now_iso: str = ""
 
 
 class WsHomeRequest(BaseModel):
@@ -225,14 +172,18 @@ class WsHomeRequest(BaseModel):
     type: Literal[WsFrameType.home_request] = WsFrameType.home_request
     message: str
     conversation_history: list[dict[str, Any]] = Field(default_factory=list)
+    format_prefs: FormatPrefsModel | None = None
 
 
-class WsFloatingRequest(BaseModel):
+class WsBriefRequest(BaseModel):
-    """Client → Server: Floating chat message scoped to an entity."""
+    """Client → Server: Request a plain-text brief (home or project)."""
 
-    type: Literal[WsFrameType.floating_request] = WsFrameType.floating_request
+    type: Literal[WsFrameType.brief_request] = WsFrameType.brief_request
-    message: str
+    request_id: str | None = None
-    scope: WsFloatingScope
+    session_id: str | None = None
+    mode: Literal["home", "project"]
+    project_id: str | None = None
+    format_prefs: FormatPrefsModel | None = None
 
 
 class WsStreamStart(BaseModel):
@@ -255,57 +206,65 @@ class WsStreamEnd(BaseModel):
 
     type: Literal[WsFrameType.stream_end] = WsFrameType.stream_end
     request_id: str
+    error: str | None = None
+    mutations: list[dict[str, Any]] | None = None
 
 
-class WsDomain(BaseModel):
+# ── Scout Config V2 ───────────────────────────────────────────────────
-    """Structured floating domain payload for UI routing decisions."""
-
-    type: Literal["task", "timeline", "project", "node"]
-    id: str | None = None
-    section: Literal["task", "timeline", "note"] | None = None
 
 
-class WsFloatingDomain(BaseModel):
+class ScoutContentTypeConfig(BaseModel):
-    """Server → Client: domain determined for a floating request."""
+    """Per-type extraction config produced by the journey chatbot."""
 
-    type: Literal[WsFrameType.floating_domain] = WsFrameType.floating_domain
+    id: str
-    request_id: str
+    label: str = ""
-    domain: WsDomain
+    detection_hint: str = ""
+    preprocessing: str = "generic"  # handler name: "email_html", "plain_text", ...
+    extraction_prompt: str
 
 
-# ── Agent Catalog ─────────────────────────────────────────────────────
+class ScoutConfig(BaseModel):
+    """Structured scout configuration (replaces freeform prompt_template)."""
 
-class AgentCatalogItem(BaseModel):
+    content_types: list[ScoutContentTypeConfig] = []
+    global_rules: list[str] = []
+    data_types: list[str] = []
+
+
+# ── Scout Catalog ─────────────────────────────────────────────────────
+
+class ScoutCatalogItem(BaseModel):
     type: str
     name: str
     description: str
 
 
-class AgentCreationCheckRequest(BaseModel):
+class ScoutCreationCheckRequest(BaseModel):
     active_agents: int = Field(ge=0, default=0)
 
 
-class AgentCreationCheckResponse(BaseModel):
+class ScoutCreationCheckResponse(BaseModel):
     allowed: bool
     tier: BillingTier
     active_agents: int
     limit: int
 
 
-class AgentTriggerRequest(BaseModel):
+class ScoutTriggerRequest(BaseModel):
     directory: str = Field(min_length=1)
     device_id: str = Field(default="")
-    agent_id: str | None = None
+    agent_id: str | None = None  # FE stable agent ID (electron-store UUID)
     what_to_extract: list[str] = Field(min_length=1)
-    actions_by_type: dict[str, list[str]] | None = None
     batch_interval: str = Field(min_length=1)
-    custom_agent_prompt: str = Field(min_length=1)
+    custom_agent_prompt: str | None = None
+    agent_config: dict | None = None
     active_agents: int = Field(ge=0, default=0)
+    last_run_at: int | None = None  # epoch ms from FE — enables incremental scanning
 
 
-# ── Agent Run Log ─────────────────────────────────────────────────────
+# ── Scout Run Log ─────────────────────────────────────────────────────
 
-class AgentRunLogResponse(BaseModel):
+class ScoutRunLogResponse(BaseModel):
     id: str
     agent_id: str
     agent_type: Literal["local", "cloud"]
@@ -315,3 +274,29 @@ class AgentRunLogResponse(BaseModel):
     errors: list[str]
     started_at: int
     completed_at: int | None
+
+
+# ── Chatbot Journey ───────────────────────────────────────────────────
+
+
+# ── Scout Proposal Frame Models ───────────────────────────────────────
+
+class ScoutProposalPayload(BaseModel):
+    id: str
+    scout_id: str
+    source_type: str
+    source_msg_ref: str
+    raw_subject: str | None = None
+    raw_snippet: str | None = None
+    category: Literal["unprocessed"] = "unprocessed"
+    payload: dict | None = None
+
+
+class ScoutProposalFrame(BaseModel):
+    type: Literal[WsFrameType.SCOUT_PROPOSAL]
+    proposal: ScoutProposalPayload
+
+
+class ScoutProposalAckFrame(BaseModel):
+    type: Literal[WsFrameType.SCOUT_PROPOSAL_ACK]
+    proposal_id: str

app/schemas/contextual.py

@@ -0,0 +1,73 @@
+"""Contextual sidebar scope schema and prompt block renderer.
+
+ContextualScope mirrors the TypeScript ContextualScope type sent by the
+Electron renderer when the user opens the side chat anchored to a specific
+view.  The renderer ships camelCase keys; Pydantic's alias_generator maps
+them to snake_case Python attributes automatically.
+"""
+
+from __future__ import annotations
+
+from typing import Literal, Optional
+
+from pydantic import BaseModel, ConfigDict
+from pydantic.alias_generators import to_camel
+
+
+PageType = Literal[
+    "timeline",
+    "tasks",
+    "projects-list",
+    "project",
+    "note",
+]
+
+EntityType = Literal["project", "note", "task", "timeline_event"]
+
+
+class ContextualScope(BaseModel):
+    """Scope payload sent by the Electron renderer for contextual chat.
+
+    The renderer ships camelCase keys (entityType, entityId, ...).  Pydantic's
+    alias generator maps them to snake_case Python attrs.
+    """
+
+    model_config = ConfigDict(populate_by_name=True, alias_generator=to_camel)
+
+    page: PageType
+    entity_type: Optional[EntityType] = None
+    entity_id: Optional[str] = None
+    entity_name: Optional[str] = None
+    project_id: Optional[str] = None
+    char_count: Optional[int] = None
+    counts: Optional[dict[str, int]] = None
+    filters: Optional[dict] = None
+
+
+def render_scope_block(scope: ContextualScope) -> str:
+    """Produce a single-paragraph human-readable summary of the current view
+    for injection into the contextual agent system prompt.
+
+    Never emits internal ids — only names.  The LLM is told to use names in
+    prose; ids travel through tool calls.
+    """
+    if scope.entity_type == "project":
+        c = scope.counts or {}
+        return (
+            f"User is viewing the project {scope.entity_name!r}. "
+            f"{c.get('tasks', 0)} tasks, "
+            f"{c.get('notes', 0)} notes, "
+            f"{c.get('milestones', 0)} milestones."
+        )
+    if scope.entity_type == "note":
+        return (
+            f"User is viewing the note {scope.entity_name!r} "
+            f"({scope.char_count or 0} characters)."
+        )
+    if scope.page == "tasks":
+        return "User is viewing the global Tasks list (all projects)."
+    if scope.page == "timeline":
+        return "User is viewing the global Timeline view."
+    if scope.page == "projects-list":
+        return "User is viewing the Projects list."
+    return f"User is on page {scope.page}."

app/scouts/connectors/base.py

@@ -0,0 +1,56 @@
+"""Source connector Protocol and shared item types.
+
+A SourceConnector adapts a third-party data source (Gmail, Slack, ...) to the
+shared ScoutEngine interface. Each connector owns:
+
+  * how to enumerate new items since the last poll (``list_new``)
+  * how to fetch a single item's metadata cheaply (``fetch_metadata``)
+  * how to fetch a single item's full content for in-memory triage
+    (``fetch_content``) — this content MUST NOT be persisted by the engine
+  * how to archive/trash an item (``archive``) for spam handling
+  * optional push-notification setup (``setup_watch`` / ``renew_watch``)
+"""
+
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Literal, Protocol
+
+from pydantic import BaseModel, Field
+
+
+class ItemRef(BaseModel):
+    source_msg_ref: str
+    received_at: datetime | None = None
+
+
+class ItemMetadata(BaseModel):
+    subject: str | None = None
+    sender: str | None = None
+    snippet: str | None = None
+    received_at: datetime | None = None
+
+
+class ItemContent(BaseModel):
+    metadata: ItemMetadata
+    body_text: str
+    raw_headers: dict[str, str] = Field(default_factory=dict)
+
+
+class TriageVerdict(BaseModel):
+    verdict: Literal["relevant", "spam"]
+    reason: str
+    confidence: float = Field(ge=0.0, le=1.0)
+
+
+class SourceConnector(Protocol):
+    """Adapter for a third-party data source (Gmail, Slack, ...)."""
+
+    source_type: str  # e.g. "gmail"
+
+    async def list_new(self, scout) -> list[ItemRef]: ...
+    async def fetch_metadata(self, scout, ref: ItemRef) -> ItemMetadata: ...
+    async def fetch_content(self, scout, ref: ItemRef) -> ItemContent: ...
+    async def archive(self, scout, ref: ItemRef) -> None: ...
+    async def setup_watch(self, scout) -> None: ...
+    async def renew_watch(self, scout) -> None: ...

app/scouts/connectors/gmail.py

@@ -0,0 +1,213 @@
+"""Gmail SourceConnector — wraps the existing GmailClient.
+
+Responsibilities:
+  * list_new: incremental fetch since the scout's stored gmail_history_id
+  * fetch_metadata: subject + sender + snippet only (Gmail metadata format)
+  * fetch_content: full body text — transient, never persisted by engine
+  * archive: move a message to Gmail Trash (recoverable for 30 days)
+  * setup_watch / renew_watch: Gmail push notifications via Pub/Sub
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from datetime import datetime, timezone
+
+from app.config.settings import settings
+from app.integrations import decrypt_token
+from app.scouts.connectors.base import ItemContent, ItemMetadata, ItemRef
+
+logger = logging.getLogger(__name__)
+
+
+def _extract_plain_text_body(payload: dict) -> str:
+    """Recursively walk a Gmail message payload to find text/plain content."""
+    import base64
+    mime_type = payload.get("mimeType", "")
+    if mime_type == "text/plain":
+        data = payload.get("body", {}).get("data", "")
+        if data:
+            return base64.urlsafe_b64decode(data + "==").decode("utf-8", errors="replace")
+        return ""
+    if mime_type.startswith("multipart/"):
+        for part in payload.get("parts", []):
+            text = _extract_plain_text_body(part)
+            if text:
+                return text
+    # text/html fallback: strip tags rudimentarily if no text/plain part
+    if mime_type == "text/html":
+        data = payload.get("body", {}).get("data", "")
+        if data:
+            import re
+            html = base64.urlsafe_b64decode(data + "==").decode("utf-8", errors="replace")
+            return re.sub(r"<[^>]+>", " ", html)
+    return ""
+
+
+def _get_gmail_service(scout):
+    """Return a synchronous Google API client for low-level metadata/history calls."""
+    from googleapiclient.discovery import build
+    from google.oauth2.credentials import Credentials
+
+    creds_info = decrypt_token(scout.oauth_token_encrypted)
+    credentials = Credentials(
+        token=creds_info.get("token"),
+        refresh_token=creds_info.get("refresh_token"),
+        token_uri=creds_info.get("token_uri", "https://oauth2.googleapis.com/token"),
+        client_id=creds_info.get("client_id"),
+        client_secret=creds_info.get("client_secret"),
+        scopes=creds_info.get("scopes"),
+    )
+    return build("gmail", "v1", credentials=credentials, cache_discovery=False)
+
+
+class GmailConnector:
+    source_type = "gmail"
+
+    # ── list_new ──────────────────────────────────────────────────────────
+
+    async def list_new(self, scout) -> list[ItemRef]:
+        """Return new message refs since scout.gmail_history_id.
+
+        On first run (gmail_history_id is None/empty), records the current
+        historyId without backfilling — avoids flooding the user with old mail.
+        Updates scout.gmail_history_id in-place (caller must persist to DB).
+        """
+        def _sync() -> tuple[list[ItemRef], str | None]:
+            service = _get_gmail_service(scout)
+            history_id = scout.gmail_history_id
+            refs: list[ItemRef] = []
+            new_history_id = history_id
+
+            if history_id:
+                resp = (
+                    service.users()
+                    .history()
+                    .list(
+                        userId="me",
+                        startHistoryId=history_id,
+                        historyTypes=["messageAdded"],
+                    )
+                    .execute()
+                )
+                for entry in resp.get("history", []):
+                    for added in entry.get("messagesAdded", []):
+                        refs.append(ItemRef(source_msg_ref=added["message"]["id"]))
+                new_history_id = resp.get("historyId", history_id)
+            else:
+                # First run: capture baseline history id without backfilling.
+                profile = service.users().getProfile(userId="me").execute()
+                new_history_id = profile["historyId"]
+
+            return refs, new_history_id
+
+        refs, new_history_id = await asyncio.to_thread(_sync)
+        if new_history_id and new_history_id != scout.gmail_history_id:
+            scout.gmail_history_id = new_history_id
+        return refs
+
+    # ── fetch_metadata ────────────────────────────────────────────────────
+
+    async def fetch_metadata(self, scout, ref: ItemRef) -> ItemMetadata:
+        """Fetch subject, sender, snippet only — uses Gmail metadata format (no body)."""
+
+        def _sync() -> ItemMetadata:
+            service = _get_gmail_service(scout)
+            msg = (
+                service.users()
+                .messages()
+                .get(
+                    userId="me",
+                    id=ref.source_msg_ref,
+                    format="metadata",
+                    metadataHeaders=["Subject", "From", "Date"],
+                )
+                .execute()
+            )
+            headers = {
+                h["name"]: h["value"]
+                for h in msg.get("payload", {}).get("headers", [])
+            }
+            return ItemMetadata(
+                subject=headers.get("Subject"),
+                sender=headers.get("From"),
+                snippet=msg.get("snippet"),
+                received_at=None,
+            )
+
+        return await asyncio.to_thread(_sync)
+
+    # ── fetch_content ─────────────────────────────────────────────────────
+
+    async def fetch_content(self, scout, ref: ItemRef) -> ItemContent:
+        """Fetch full body text for a single message — transient, must not be persisted."""
+
+        def _sync() -> ItemContent:
+            service = _get_gmail_service(scout)
+            msg = service.users().messages().get(
+                userId="me", id=ref.source_msg_ref, format="full",
+            ).execute()
+            headers = {h["name"]: h["value"] for h in msg.get("payload", {}).get("headers", [])}
+            body_text = _extract_plain_text_body(msg.get("payload", {}))
+            return ItemContent(
+                metadata=ItemMetadata(
+                    subject=headers.get("Subject"),
+                    sender=headers.get("From"),
+                    snippet=msg.get("snippet"),
+                    received_at=None,
+                ),
+                body_text=body_text,
+                raw_headers=headers,
+            )
+
+        return await asyncio.to_thread(_sync)
+
+    # ── archive ───────────────────────────────────────────────────────────
+
+    async def archive(self, scout, ref: ItemRef) -> None:
+        """Move the message to Gmail Trash (recoverable for 30 days)."""
+
+        def _sync() -> None:
+            service = _get_gmail_service(scout)
+            service.users().messages().trash(
+                userId="me", id=ref.source_msg_ref
+            ).execute()
+
+        await asyncio.to_thread(_sync)
+
+    # ── watch management ──────────────────────────────────────────────────
+
+    async def setup_watch(self, scout) -> None:
+        """Register a Gmail Pub/Sub push watch for the INBOX label.
+
+        Requires ``settings.GMAIL_PUBSUB_TOPIC`` to be set to the full topic
+        resource name (e.g. ``projects/my-project/topics/gmail-push``).
+        Logs a warning and returns without error if the topic is not configured.
+        """
+        topic = settings.GMAIL_PUBSUB_TOPIC
+        if not topic:
+            logger.warning(
+                "setup_watch: GMAIL_PUBSUB_TOPIC is not configured — skipping watch setup"
+            )
+            return
+
+        def _sync() -> None:
+            service = _get_gmail_service(scout)
+            request_body = {
+                "labelIds": ["INBOX"],
+                "topicName": topic,
+            }
+            resp = service.users().watch(userId="me", body=request_body).execute()
+            scout.gmail_history_id = resp.get("historyId")
+            expiration_ms = resp.get("expiration")
+            if expiration_ms:
+                scout.gmail_watch_expires_at = datetime.fromtimestamp(
+                    int(expiration_ms) / 1000, tz=timezone.utc
+                )
+
+        await asyncio.to_thread(_sync)
+
+    async def renew_watch(self, scout) -> None:
+        """Renew an existing Gmail Pub/Sub watch (same as setup_watch)."""
+        await self.setup_watch(scout)

app/scouts/connectors/registry.py

@@ -0,0 +1,32 @@
+"""Connector registry — single source of truth for source_type -> connector."""
+
+from __future__ import annotations
+
+from typing import Any
+
+_CONNECTORS: dict[str, Any] = {}
+
+
+def register_connector(connector: Any) -> None:
+    """Register a SourceConnector instance under its ``source_type``.
+
+    Calling twice with the same ``source_type`` replaces the prior entry —
+    useful for tests and hot-reload, but in production each connector
+    should be registered exactly once at startup.
+    """
+    if not getattr(connector, "source_type", None):
+        raise ValueError("Connector must declare a non-empty source_type")
+    _CONNECTORS[connector.source_type] = connector
+
+
+def get_connector(source_type: str) -> Any:
+    """Return the registered connector for ``source_type`` or raise KeyError."""
+    try:
+        return _CONNECTORS[source_type]
+    except KeyError as exc:
+        raise KeyError(f"No connector registered for source_type {source_type!r}") from exc
+
+
+def _reset_for_tests() -> None:
+    """Clear the registry — for use in pytest fixtures only."""
+    _CONNECTORS.clear()

app/scouts/engine.py

@@ -0,0 +1,270 @@
+"""ScoutEngine — orchestrates triage, queueing, and delivery for cloud scouts.
+
+Triage flow per scout:
+  1. Resolve scout config from the DB.
+  2. Skip if device hasn't connected within ``device_inactivity_pause_days``.
+  3. Ask the connector to ``list_new`` — fresh items since last poll.
+  4. For each item:
+     - skip if already in the queue (idempotent on (scout_id, source_msg_ref))
+     - fetch the full content via the connector (transient, never persisted)
+     - run the triage LLM call → relevant | spam
+     - spam + auto_trash_spam → connector.archive
+     - relevant → INSERT scout_triage_queue row
+  5. Update scout.last_run_at.
+
+Delivery flow on Electron WS reconnect:
+  - drain ``status='queued'`` rows for the user
+  - fetch metadata-only for each (subject + snippet)
+  - send a ``scout_proposal`` frame
+  - flip status to ``delivered`` on ack
+"""
+
+from __future__ import annotations
+
+import logging
+import uuid
+from datetime import datetime, timedelta, timezone
+
+from sqlalchemy import select
+from sqlalchemy.exc import IntegrityError
+
+from app.core.langfuse_client import extract_usage, get_langfuse, get_prompt_or_fallback
+from app.core.llm import get_llm
+from app.db import async_session
+from app.models import CloudScoutConfig, ScoutTriageQueue
+from app.scouts.connectors.base import ItemContent, ItemRef, TriageVerdict
+from app.scouts.connectors.registry import get_connector
+
+logger = logging.getLogger(__name__)
+
+QUEUE_TTL_DAYS = 30
+
+
+class ScoutEngine:
+    def __init__(self, session_factory=None) -> None:
+        self._session_factory = session_factory or async_session
+
+    async def trigger_scout(self, scout_id: uuid.UUID) -> None:
+        async with self._session_factory() as session:
+            scout = await session.get(CloudScoutConfig, str(scout_id))
+            if scout is None:
+                logger.warning("trigger_scout: no such scout id=%s", scout_id)
+                return
+            if not scout.enabled:
+                return
+            # Device-inactivity pause check is a simple heuristic on last_run_at —
+            # the device-online signal lives in the DeviceConnectionManager and is
+            # consulted at delivery time. For triage, we only check that the
+            # configured pause threshold isn't suppressing the run.
+            connector = get_connector(scout.provider)
+            try:
+                refs = await connector.list_new(scout)
+            except Exception:
+                logger.exception("scout %s: list_new failed", scout.id)
+                return
+
+            for ref in refs:
+                await self._process_item(session, scout, connector, ref)
+
+            scout.last_run_at = datetime.now(tz=timezone.utc)
+            await session.commit()
+
+    async def _process_item(
+        self,
+        session,
+        scout: CloudScoutConfig,
+        connector,
+        ref: ItemRef,
+    ) -> None:
+        # Idempotency check
+        existing = await session.execute(
+            select(ScoutTriageQueue.id).where(
+                ScoutTriageQueue.scout_id == scout.id,
+                ScoutTriageQueue.source_msg_ref == ref.source_msg_ref,
+            )
+        )
+        if existing.first() is not None:
+            return
+
+        try:
+            content = await connector.fetch_content(scout, ref)
+        except Exception:
+            logger.exception("scout %s: fetch_content failed for %s", scout.id, ref.source_msg_ref)
+            return
+
+        try:
+            verdict = await self._triage_llm(scout, content)
+        except Exception:
+            logger.exception("scout %s: triage_llm failed for %s", scout.id, ref.source_msg_ref)
+            return
+
+        if verdict.verdict == "spam":
+            if scout.auto_trash_spam:
+                try:
+                    await connector.archive(scout, ref)
+                except Exception:
+                    logger.exception("scout %s: archive failed for %s", scout.id, ref.source_msg_ref)
+            return
+
+        now = datetime.now(tz=timezone.utc)
+        row = ScoutTriageQueue(
+            id=str(uuid.uuid4()),
+            user_id=scout.user_id,
+            scout_id=scout.id,
+            source_type=connector.source_type,
+            source_msg_ref=ref.source_msg_ref,
+            triage_verdict=verdict.verdict,
+            triage_reason=verdict.reason,
+            status="queued",
+            triaged_at=now,
+            expires_at=now + timedelta(days=QUEUE_TTL_DAYS),
+        )
+        session.add(row)
+        try:
+            # Use a savepoint so an IntegrityError on race doesn't poison the
+            # outer session — works on both PostgreSQL (SAVEPOINT) and SQLite.
+            async with session.begin_nested():
+                await session.flush()
+        except IntegrityError:
+            # Race: another worker inserted between our SELECT and INSERT.
+            # The unique constraint did its job; safe to ignore.
+            logger.debug(
+                "scout %s: idempotent skip for %s (race on unique constraint)",
+                scout.id,
+                ref.source_msg_ref,
+            )
+
+    async def deliver_pending(self, user_id: uuid.UUID, ws) -> None:
+        """Drain status='queued' rows for user, send scout_proposal WS frames, flip to 'delivered'."""
+        from app.scouts.connectors.base import ItemRef  # noqa: PLC0415
+        async with self._session_factory() as session:
+            rows = (await session.execute(
+                select(ScoutTriageQueue).where(
+                    ScoutTriageQueue.user_id == str(user_id),
+                    ScoutTriageQueue.status == "queued",
+                )
+            )).scalars().all()
+
+            for row in rows:
+                try:
+                    connector = get_connector(row.source_type)
+                except KeyError:
+                    logger.warning("deliver_pending: no connector for %s", row.source_type)
+                    continue
+                scout = await session.get(CloudScoutConfig, row.scout_id)
+                if scout is None:
+                    continue
+                try:
+                    meta = await connector.fetch_metadata(scout, ItemRef(source_msg_ref=row.source_msg_ref))
+                except Exception:
+                    logger.exception("deliver_pending: fetch_metadata failed")
+                    continue
+
+                payload = {
+                    "type": "scout_proposal",
+                    "proposal": {
+                        "id": row.id,
+                        "scout_id": row.scout_id,
+                        "source_type": row.source_type,
+                        "source_msg_ref": row.source_msg_ref,
+                        "raw_subject": meta.subject,
+                        "raw_snippet": meta.snippet,
+                        "category": "unprocessed",
+                        "payload": None,
+                    },
+                }
+                await ws.send_json(payload)
+                row.status = "delivered"
+                row.delivered_at = datetime.now(tz=timezone.utc)
+
+            await session.commit()
+
+    async def ack_proposal(self, proposal_id: str) -> None:
+        """Flip a delivered proposal to acked. Idempotent — no-op if already acked."""
+        async with self._session_factory() as session:
+            row = await session.get(ScoutTriageQueue, proposal_id)
+            if row is None:
+                return
+            row.status = "acked"
+            row.acked_at = datetime.now(tz=timezone.utc)
+            await session.commit()
+
+    async def _triage_llm(self, scout: CloudScoutConfig, content: ItemContent) -> TriageVerdict:
+        """Call the scout-triage-system Langfuse prompt to classify an item as relevant or spam.
+
+        Uses gpt-4o-mini with JSON mode. Wraps the LLM call in a Langfuse generation
+        observation when Langfuse is configured.
+        """
+        import json  # noqa: PLC0415
+
+        from langchain_core.messages import HumanMessage, SystemMessage  # noqa: PLC0415
+
+        _TRIAGE_FALLBACK = (
+            "You are a triage classifier for an executive-assistant scout that watches a "
+            "{source_type} feed.\n"
+            'The scout\'s purpose is: "{scout_purpose}".\n\n'
+            "Given one item, decide whether it is RELEVANT (worth surfacing to the user as a "
+            "potential task / event / note / project) or SPAM (advertising, mass marketing, "
+            "phishing, bulk notifications with no actionable content).\n\n"
+            "Item:\n"
+            "  - Subject: {item_subject}\n"
+            "  - From:    {item_sender}\n"
+            "  - Body (truncated): {item_body_truncated_2k}\n\n"
+            'Return JSON only, matching this schema:\n'
+            '  {{"verdict": "relevant" | "spam", "reason": <short string>, "confidence": <0..1>}}\n\n'
+            "Be conservative on \"spam\" — if a message could plausibly be a personal/work "
+            "email, mark it relevant."
+        )
+
+        template, prompt_obj = get_prompt_or_fallback("scout-triage-system", _TRIAGE_FALLBACK)
+
+        body_trunc = (content.body_text or "")[:2000]
+        variables = dict(
+            source_type=scout.provider,
+            scout_purpose=scout.prompt_template or "",
+            item_subject=content.metadata.subject or "",
+            item_sender=content.metadata.sender or "",
+            item_body_truncated_2k=body_trunc,
+        )
+
+        if prompt_obj is not None:
+            try:
+                system_text = prompt_obj.compile(**variables)
+                if isinstance(system_text, list):
+                    system_text = "\n".join(
+                        m.get("content", "") for m in system_text if isinstance(m, dict)
+                    )
+            except Exception as exc:
+                logger.warning("scout triage: compile failed: %s", exc)
+                system_text = template.replace("{{source_type}}", variables["source_type"]) \
+                    .replace("{{scout_purpose}}", variables["scout_purpose"]) \
+                    .replace("{{item_subject}}", variables["item_subject"]) \
+                    .replace("{{item_sender}}", variables["item_sender"]) \
+                    .replace("{{item_body_truncated_2k}}", variables["item_body_truncated_2k"])
+        else:
+            system_text = template.format(**variables)
+
+        llm = get_llm(model="gpt-4o-mini", temperature=0)
+        llm_json = llm.bind(response_format={"type": "json_object"})  # type: ignore[attr-defined]
+
+        messages = [
+            SystemMessage(content=system_text),
+            HumanMessage(content="Classify this item."),
+        ]
+
+        lf = get_langfuse()
+        if lf:
+            with lf.start_as_current_observation(
+                as_type="generation",
+                name="scout-triage",
+                model="gpt-4o-mini",
+                prompt=prompt_obj,
+                input=messages,
+            ) as gen:
+                response = await llm_json.ainvoke(messages)
+                gen.update(output=response.content, usage=extract_usage(response))
+        else:
+            response = await llm_json.ainvoke(messages)
+
+        data = json.loads(response.content)
+        return TriageVerdict(**data)

docker-compose.yml

@@ -1,34 +1,27 @@
-# ── Adiuva Microservices ─────────────────────────────────────────────
-# docker compose up --build
-# docker compose up --build auth ws-gateway chat   # subset
-
 services:
-
+  app:
-  # ═══════════════════════════════════════════════════════════════════
+    build: .
-  #  Infrastructure
-  # ═══════════════════════════════════════════════════════════════════
-
-  traefik:
-    image: traefik:v3.1
     ports:
-      - "80:80"
+      - "8080:8000"
-      - "443:443"
+    env_file:
-      - "8080:8080"   # dashboard (dev only)
+      - path: .env
+        required: false
     environment:
-      CF_DNS_API_TOKEN: ${CF_DNS_API_TOKEN:-}
+      DATABASE_URL: postgresql+asyncpg://postgres:postgres@db:5432/adiuvai
+      GITHUB_COPILOT_TOKEN_DIR: /root/.config/litellm/github_copilot
     volumes:
-      - /var/run/docker.sock:/var/run/docker.sock:ro
+      - copilot_tokens:/root/.config/litellm/github_copilot
-      - ./traefik/traefik.yml:/etc/traefik/traefik.yml:ro
+    depends_on:
-      - ./traefik/dynamic:/etc/traefik/dynamic:ro
+      db:
-      - traefik_acme:/etc/traefik/acme
+        condition: service_healthy
     restart: unless-stopped
 
   db:
     image: pgvector/pgvector:pg16
     environment:
-      POSTGRES_USER: ${POSTGRES_USER:-postgres}
+      POSTGRES_USER: postgres
-      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:-postgres}
+      POSTGRES_PASSWORD: postgres
-      POSTGRES_DB: ${POSTGRES_DB:-adiuva}
+      POSTGRES_DB: adiuvai
     volumes:
       - postgres_data:/var/lib/postgresql/data
     healthcheck:
@@ -38,161 +31,11 @@ services:
       retries: 5
     restart: unless-stopped
 
-  redis:
+  # Optional Redis for future rate-limit or caching needs
-    image: redis:7-alpine
+  # redis:
-    command: redis-server --maxmemory 256mb --maxmemory-policy allkeys-lru
+  #   image: redis:7-alpine
-    volumes:
-      - redis_data:/data
-    healthcheck:
-      test: ["CMD", "redis-cli", "ping"]
-      interval: 5s
-      timeout: 3s
-      retries: 5
-    restart: unless-stopped
-
-  # ── Optional infrastructure (uncomment as needed) ────────────────
-
-  # minio:
-  #   image: minio/minio:latest
-  #   command: server /data --console-address ":9001"
-  #   ports:
-  #     - "9000:9000"
-  #     - "9001:9001"
-  #   environment:
-  #     MINIO_ROOT_USER: minioadmin
-  #     MINIO_ROOT_PASSWORD: minioadmin
-  #   volumes:
-  #     - minio_data:/data
-  #   healthcheck:
-  #     test: ["CMD", "mc", "ready", "local"]
-  #     interval: 5s
-  #     timeout: 5s
-  #     retries: 5
   #   restart: unless-stopped
 
-  # qdrant:
-  #   image: qdrant/qdrant:latest
-  #   ports:
-  #     - "6333:6333"
-  #     - "6334:6334"
-  #   volumes:
-  #     - qdrant_data:/qdrant/storage
-  #   restart: unless-stopped
-
-  # ═══════════════════════════════════════════════════════════════════
-  #  Migrations (run once, then exit)
-  # ═══════════════════════════════════════════════════════════════════
-
-  migrate:
-    build:
-      context: .
-      dockerfile: Dockerfile
-    command: ["python", "-m", "alembic", "upgrade", "head"]
-    env_file:
-      - path: .env
-        required: false
-    environment:
-      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:-postgres}@db:5432/${POSTGRES_DB:-adiuva}
-    depends_on:
-      db:
-        condition: service_healthy
-    restart: "no"
-
-  # ═══════════════════════════════════════════════════════════════════
-  #  Application Services
-  # ═══════════════════════════════════════════════════════════════════
-
-  auth:
-    build:
-      context: .
-      dockerfile: services/auth/Dockerfile
-    env_file:
-      - path: .env
-        required: false
-    environment:
-      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:-postgres}@db:5432/${POSTGRES_DB:-adiuva}
-      REDIS_URL: redis://redis:6379/0
-    depends_on:
-      db:
-        condition: service_healthy
-      migrate:
-        condition: service_completed_successfully
-    restart: unless-stopped
-
-  ws-gateway:
-    build:
-      context: .
-      dockerfile: services/ws-gateway/Dockerfile
-    env_file:
-      - path: .env
-        required: false
-    environment:
-      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:-postgres}@db:5432/${POSTGRES_DB:-adiuva}
-      REDIS_URL: redis://redis:6379/0
-    depends_on:
-      redis:
-        condition: service_healthy
-      auth:
-        condition: service_started
-    restart: unless-stopped
-
-  chat:
-    build:
-      context: .
-      dockerfile: services/chat/Dockerfile
-    env_file:
-      - path: .env
-        required: false
-    environment:
-      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:-postgres}@db:5432/${POSTGRES_DB:-adiuva}
-      REDIS_URL: redis://redis:6379/0
-    depends_on:
-      db:
-        condition: service_healthy
-      redis:
-        condition: service_healthy
-      migrate:
-        condition: service_completed_successfully
-    restart: unless-stopped
-
-  batch-agent:
-    build:
-      context: .
-      dockerfile: services/batch-agent/Dockerfile
-    env_file:
-      - path: .env
-        required: false
-    environment:
-      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:-postgres}@db:5432/${POSTGRES_DB:-adiuva}
-      REDIS_URL: redis://redis:6379/0
-    depends_on:
-      db:
-        condition: service_healthy
-      redis:
-        condition: service_healthy
-      migrate:
-        condition: service_completed_successfully
-    restart: unless-stopped
-
-  billing:
-    build:
-      context: .
-      dockerfile: services/billing/Dockerfile
-    env_file:
-      - path: .env
-        required: false
-    environment:
-      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:-postgres}@db:5432/${POSTGRES_DB:-adiuva}
-    depends_on:
-      db:
-        condition: service_healthy
-      migrate:
-        condition: service_completed_successfully
-    restart: unless-stopped
-
 volumes:
   postgres_data:
-  redis_data:
+  copilot_tokens:
-  traefik_acme:
-  # minio_data:
-  # qdrant_data:

docs/MICROSERVICES_ARCHITECTURE.md

@@ -1,941 +0,0 @@
-# Adiuva — Architettura Microservizi (MVP)
-
-## Panoramica
-
-Il monolite viene suddiviso in **4 servizi MVP** + un **API Gateway (Traefik)**, orchestrati con Docker Compose su un singolo VPS raggiungibile via Cloudflare.
-
-> **Fuori dall'MVP**: Storage Service (S3/backup CRUD) e Plugin Service (marketplace). Verranno aggiunti come servizi indipendenti in una fase successiva.
-
-```
-                          ┌──────────────┐
-                          │  Cloudflare  │
-                          │  (DNS + CDN) │
-                          └──────┬───────┘
-                                 │ HTTPS / WSS
-                          ┌──────▼───────┐
-                          │   Traefik    │
-                          │ API Gateway  │
-                          │  (routing,   │
-                          │   TLS, rate  │
-                          │   limiting)  │
-                          └──────┬───────┘
-                                 │
-          ┌──────────┬───────────┼───────────┐
-          │          │           │           │
-    ┌─────▼────┐ ┌───▼───┐ ┌────▼────┐ ┌────▼───┐
-    │  Auth    │ │  Chat │ │  Agent  │ │Billing │
-    │ Service  │ │Service│ │ Service │ │Service │
-    └─────┬────┘ └───┬───┘ └────┬────┘ └────┬───┘
-          │          │          │           │
-    ┌─────▼──────────▼──────────▼───────────▼────┐
-    │              Infrastruttura                 │
-    │  PostgreSQL  │  Redis  │  Qdrant            │
-    └─────────────────────────────────────────────┘
-```
-
----
-
-## 1. Suddivisione dei Servizi
-
-### 1.1 Auth Service (`auth-service`)
-
-**Responsabilità**: Registrazione, login, refresh token, profilo utente, encryption key.
-
-| Endpoint originale | Metodo |
-|---|---|
-| `/api/v1/auth/register` | POST |
-| `/api/v1/auth/login` | POST |
-| `/api/v1/auth/refresh` | POST |
-| `/api/v1/auth/me` | GET / PUT |
-
-**Database**: Tabelle `users`, `refresh_tokens` (PostgreSQL condiviso, schema `auth`).
-
-**Modifica chiave — JWT con RS256**:
-Il monolite usa un `SECRET_KEY` simmetrico (HS256). Con i microservizi, passare a **RS256** (asimmetrico):
-- L'Auth Service firma i JWT con la **chiave privata**.
-- Tutti gli altri servizi verificano i JWT con la **chiave pubblica** senza mai contattare l'Auth Service.
-- La chiave pubblica viene esposta via `GET /api/v1/auth/.well-known/jwks.json` oppure montata come volume condiviso.
-
-```python
-# auth-service/app/auth/jwt.py
-from cryptography.hazmat.primitives.asymmetric import rsa
-from jose import jwt
-
-PRIVATE_KEY = ...  # Da env/secret
-PUBLIC_KEY = ...   # Derivata o da env
-
-def create_access_token(user_id: str, tier: str) -> str:
-    return jwt.encode(
-        {"sub": user_id, "tier": tier, "exp": ...},
-        PRIVATE_KEY,
-        algorithm="RS256",
-    )
-```
-
-```python
-# shared/auth.py  (usato da tutti gli altri servizi)
-from jose import jwt
-
-PUBLIC_KEY = ...  # Volume montato o fetched da JWKS endpoint
-
-def verify_token(token: str) -> dict:
-    return jwt.decode(token, PUBLIC_KEY, algorithms=["RS256"])
-```
-
-**Scaling**: 2 repliche sufficienti, stateless. Rate-limit dedicato su `/login` e `/register`.
-
----
-
-### 1.2 Chat Service (`chat-service`) ⭐ Real-time
-
-**Responsabilità**: WebSocket device connection, home chat, floating chat, memory middleware, streaming LLM responses verso il client.
-
-Questo servizio gestisce la **connessione persistente** con l'app Electron e le interazioni **real-time** dell'utente (chat home, floating chat). È il proprietario della WebSocket.
-
-| Endpoint | Tipo |
-|---|---|
-| `/api/v1/ws/device` | WebSocket (connessione persistente) |
-| `/api/v1/chat` | POST (REST fallback) |
-
-**Moduli inclusi**: `deep_agent`, `memory_middleware`, `ws_context`, `device_manager` (Redis-backed), `output_formatter`, `llm`, tutti gli agent tools (`task_agent`, `project_agent`, `note_agent`, `timeline_agent`).
-
-**Perché separato dall'Agent Service**: Il Chat Service tiene la WebSocket aperta e risponde in tempo reale (streaming). Scalare aggiungendo repliche è semplice con sticky sessions + Redis pub/sub per il cross-instance routing dei tool_call.
-
-**Scaling**: 2–N repliche. Sticky cookies per le WS + Redis per cross-instance.
-
----
-
-### 1.3 Agent Service (`agent-service`) ⭐ Batch
-
-**Responsabilità**: Batch agent processing (directory scanning, file classification, entity extraction), agent setup journeys, agent configuration CRUD.
-
-Questo servizio gestisce i processi **long-running** e **CPU-intensive**: scansione filesystem, classificazione file con LLM, estrazione entità in batch. Non possiede la WebSocket — comunica con il device dell'utente tramite **Redis pub/sub** passando per il Chat Service.
-
-| Endpoint | Tipo |
-|---|---|
-| `/api/v1/agents/catalog` | GET |
-| `/api/v1/agents/can-create` | POST |
-| `/api/v1/agents/trigger` | POST |
-| `/api/v1/agents/journey/start` | POST (o WS relay) |
-| `/api/v1/agents/journey/message` | POST (o WS relay) |
-
-**Moduli inclusi**: `agent_runner`, `agent_registry`, `filesystem_agent`, `llm`.
-
-**Flusso tool-call cross-service** (l'Agent Service non ha la WS):
-
-```
-┌──────────────┐            ┌──────────────┐            ┌──────────┐
-│ Agent Service│            │    Redis     │            │  Chat    │
-│ (batch run)  │            │              │            │ Service  │
-│              │            │              │            │ (ha WS)  │
-│ 1. Needs to  │  PUBLISH   │              │ SUBSCRIBE  │          │
-│    read file ├───────────►│tool_call:u123├───────────►│ 2. Invia │
-│    from      │            │              │            │    al    │
-│    device    │            │              │            │    device│
-│              │            │              │            │    via WS│
-│              │  SUBSCRIBE │              │  PUBLISH   │          │
-│ 4. Riceve   ◄────────────┤tool_result:id│◄───────────┤ 3. Device│
-│    risultato │            │              │            │    reply │
-└──────────────┘            └──────────────┘            └──────────┘
-```
-
-**Scaling**: 1–N repliche. Completamente stateless, scala indipendentemente dalla chat. Ogni replica processa batch job diversi. Può essere scalato a 0 se non ci sono agent attivi (risparmio risorse).
-
-**Vantaggio dello split**: Se 50 utenti triggerano agenti batch contemporaneamente, il Chat Service non ne risente — le risposte real-time rimangono veloci.
-
----
-
-### 1.4 Billing Service (`billing-service`)
-
-**Responsabilità**: Stripe checkout, webhook, subscription management.
-
-| Endpoint originale | Metodo |
-|---|---|
-| `/api/v1/billing/checkout` | POST |
-| `/api/v1/billing/webhook` | POST |
-| `/api/v1/billing/subscription` | GET / DELETE |
-
-**Database**: Tabelle `subscriptions` (schema `billing`).
-
-**Comunicazione inter-servizio**: Quando Stripe invia un webhook e il tier cambia, il Billing Service pubblica un evento su **Redis pub/sub** channel `tier_changed:{user_id}`. L'Auth Service aggiorna il campo `tier` nella tabella users. Al prossimo token refresh il JWT conterrà il tier aggiornato.
-
-**Scaling**: 1 replica sufficiente. Basso traffico.
-
----
-
-### 1.5 Servizi esclusi dall'MVP
-
-I seguenti servizi verranno aggiunti post-MVP come servizi indipendenti:
-
-| Servizio | Responsabilità | Note |
-|---|---|---|
-| **Storage Service** | S3 blobs CRUD, vector ops, backup | Le funzionalità vector/embed possono restare nel Chat Service per il MVP |
-| **Plugin Service** | Marketplace, install, revenue split | Feature non critica per il lancio |
-
----
-
-## 2. Tier Check — Dove e Come
-
-Il tier dell'utente (free/pro/power/team) determina rate-limiting, quote e accesso a funzionalità. Con i microservizi, **ogni servizio controlla il tier autonomamente** senza chiamare l'Auth Service.
-
-### Strategia: Tier nel JWT
-
-L'Auth Service include il `tier` come claim nel JWT al momento del login/refresh:
-
-```json
-{
-  "sub": "user_123",
-  "tier": "pro",
-  "exp": 1742515200,
-  "iat": 1742511600
-}
-```
-
-Ogni servizio:
-1. Decodifica il JWT con la chiave pubblica (già lo fa per l'auth)
-2. Legge `payload["tier"]` — **zero chiamate extra**
-3. Applica le sue regole di enforcement localmente
-
-```python
-# shared/auth.py — dependency FastAPI condivisa
-from fastapi import Depends, HTTPException, Request
-from jose import jwt
-
-PUBLIC_KEY = ...
-
-class CurrentUser:
-    def __init__(self, user_id: str, tier: str):
-        self.user_id = user_id
-        self.tier = tier
-
-async def get_current_user(request: Request) -> CurrentUser:
-    token = request.headers.get("Authorization", "").removeprefix("Bearer ")
-    payload = jwt.decode(token, PUBLIC_KEY, algorithms=["RS256"])
-    return CurrentUser(user_id=payload["sub"], tier=payload["tier"])
-
-def require_tier(*allowed_tiers: str):
-    """Dependency che blocca se il tier non è tra quelli ammessi."""
-    async def check(user: CurrentUser = Depends(get_current_user)):
-        if user.tier not in allowed_tiers:
-            raise HTTPException(403, "Tier insufficient")
-        return user
-    return check
-```
-
-### Cosa succede quando il tier cambia (upgrade/downgrade)?
-
-```
-┌──────────┐  Stripe webhook   ┌──────────┐  tier_changed   ┌──────────┐
-│  Stripe  │ ─────────────────►│ Billing  │ ───────────────►│   Auth   │
-│          │                    │ Service  │  (Redis pub/sub) │ Service  │
-└──────────┘                    └──────────┘                  └────┬─────┘
-                                                                   │
-                                                          UPDATE users
-                                                          SET tier = 'power'
-                                                                   │
-                                                    Al prossimo /refresh
-                                                    il JWT conterrà tier='power'
-```
-
-**Latenza del cambio**: Il tier si propaga al prossimo token refresh (tipicamente 15–30 min, o il client può forzare un refresh immediato dopo il checkout). Per il billing webhook, il downgrade può essere forzato invalidando il refresh token su Redis → il client è obbligato a ri-autenticarsi.
-
-### Dove si applica in ciascun servizio
-
-| Servizio | Enforcement |
-|---|---|
-| **Auth Service** | Nessuno (è lui che scrive il tier) |
-| **Chat Service** | Rate-limit per tier (req/min), quota messaggi |
-| **Agent Service** | Max agent configs, max runs/day, max concurrent batches |
-| **Billing Service** | Nessuno (gestisce i tier, non li consuma) |
-
-### Rate-limit distribuito via Redis
-
-Poiché ogni servizio ha le sue repliche, il rate-limiting deve essere **condiviso** via Redis:
-
-```python
-# shared/middleware/rate_limit.py
-import redis.asyncio as aioredis
-
-class DistributedRateLimiter:
-    def __init__(self, redis: aioredis.Redis):
-        self._redis = redis
-
-    async def check(self, user_id: str, tier: str, service: str) -> bool:
-        limits = {"free": 20, "pro": 60, "power": 120, "team": 200}
-        max_req = limits.get(tier, 20)
-        key = f"rate:{service}:{user_id}"
-
-        pipe = self._redis.pipeline()
-        pipe.incr(key)
-        pipe.expire(key, 60)
-        count, _ = await pipe.execute()
-
-        return count <= max_req
-```
-
----
-
-## 3. WebSocket con Scaling Orizzontale — Il Problema Chiave
-
-`DeviceConnectionManager` è un **singleton in-memory**:
-
-```python
-class DeviceConnectionManager:
-    def __init__(self):
-        self._connections: dict[str, DeviceConnection] = {}  # ← In-memory!
-```
-
-Con N istanze del Chat Service, il device si connette a **una sola** istanza. Quando un'altra istanza deve inviare un `tool_call` a quel device (es. un agent trigger da un'API call), non trova la connessione.
-
-### La soluzione: Redis Pub/Sub + Registry
-
-```
-┌──────────────────────────────────────────────────────────────┐
-│                     Redis                                    │
-│                                                              │
-│  Hash: ws:connections                                        │
-│    user_123 → instance_A                                     │
-│    user_456 → instance_B                                     │
-│                                                              │
-│  Pub/Sub channels:                                           │
-│    tool_call:{user_id}  → tool call payloads                 │
-│    tool_result:{call_id} → tool result payloads              │
-│    stream:{user_id}     → text_chunk streaming               │
-└──────────────────────────────────────────────────────────────┘
-
- Instance A (ha WS di user_123)     Instance B (deve chiamare tool su user_123)
- ┌───────────────────────┐          ┌───────────────────────┐
- │  1. Sottoscrive a     │          │  1. Lookup Redis Hash │
- │     tool_call:user_123│          │     → user_123 è su A │
- │                       │          │                       │
- │  2. Riceve tool_call  │◄─────────│  2. PUBLISH           │
- │     da Redis channel  │          │    tool_call:user_123 │
- │                       │          │    {id, action, ...}  │
- │  3. Invia al device   │          │                       │
- │     via WS            │          │  4. SUBSCRIBE         │
- │                       │          │    tool_result:{id}   │
- │  4. Device risponde   │          │                       │
- │     tool_result       │──────────│► 5. Riceve risultato  │
- │                       │          │                       │
- │  5. PUBLISH           │          │                       │
- │    tool_result:{id}   │          │                       │
- └───────────────────────┘          └───────────────────────┘
-```
-
-### Implementazione: `RedisDeviceManager`
-
-```python
-# chat-service/app/core/device_manager.py
-
-import asyncio
-import json
-import os
-import redis.asyncio as aioredis
-from dataclasses import dataclass, field
-from fastapi import WebSocket
-
-INSTANCE_ID = os.environ.get("INSTANCE_ID", os.urandom(8).hex())
-
-@dataclass
-class LocalConnection:
-    ws: WebSocket
-    device_id: str
-    pending_calls: dict[str, asyncio.Future[dict]] = field(default_factory=dict)
-
-
-class RedisDeviceManager:
-    """Device manager backed by Redis for cross-instance communication."""
-
-    def __init__(self, redis_url: str = "redis://redis:6379"):
-        self._redis = aioredis.from_url(redis_url)
-        self._pubsub = self._redis.pubsub()
-        self._local: dict[str, LocalConnection] = {}  # Solo connessioni locali
-        self._remote_futures: dict[str, asyncio.Future[dict]] = {}
-
-    async def start(self):
-        """Avvia il listener Redis per tool_call in arrivo."""
-        asyncio.create_task(self._listen_tool_calls())
-
-    # ── Registrazione ──
-
-    async def register(self, user_id: str, device_id: str, ws: WebSocket):
-        # Registra localmente
-        self._local[user_id] = LocalConnection(ws=ws, device_id=device_id)
-        # Registra in Redis quale istanza ha la connessione
-        await self._redis.hset("ws:connections", user_id, INSTANCE_ID)
-        # Sottoscrivi ai tool_call per questo utente
-        await self._pubsub.subscribe(f"tool_call:{user_id}")
-
-    async def unregister(self, user_id: str):
-        conn = self._local.pop(user_id, None)
-        if conn:
-            for fut in conn.pending_calls.values():
-                if not fut.done():
-                    fut.cancel()
-        await self._redis.hdel("ws:connections", user_id)
-        await self._pubsub.unsubscribe(f"tool_call:{user_id}")
-
-    # ── Presenza ──
-
-    async def is_online(self, user_id: str) -> bool:
-        return await self._redis.hexists("ws:connections", user_id)
-
-    # ── Tool-call round-trip (cross-instance) ──
-
-    async def execute_tool_call(self, user_id: str, payload: dict) -> dict:
-        """
-        Invia un tool_call al device dell'utente.
-        Funziona sia che la WS sia locale che su un'altra istanza.
-        """
-        call_id = payload["id"]
-
-        # Caso 1: connessione locale → invio diretto
-        if user_id in self._local:
-            conn = self._local[user_id]
-            loop = asyncio.get_event_loop()
-            fut: asyncio.Future[dict] = loop.create_future()
-            conn.pending_calls[call_id] = fut
-            await conn.ws.send_text(json.dumps({"type": "tool_call", **payload}))
-            return await asyncio.wait_for(fut, timeout=30.0)
-
-        # Caso 2: connessione remota → Redis pub/sub
-        loop = asyncio.get_event_loop()
-        fut = loop.create_future()
-        self._remote_futures[call_id] = fut
-
-        # Sottoscrivi al canale di risposta
-        result_channel = f"tool_result:{call_id}"
-        await self._pubsub.subscribe(result_channel)
-
-        # Pubblica il tool_call
-        await self._redis.publish(
-            f"tool_call:{user_id}",
-            json.dumps(payload),
-        )
-
-        try:
-            return await asyncio.wait_for(fut, timeout=30.0)
-        finally:
-            self._remote_futures.pop(call_id, None)
-            await self._pubsub.unsubscribe(result_channel)
-
-    # ── Risoluzione tool_result (da WS locale) ──
-
-    def resolve_local(self, user_id: str, call_id: str, result: dict):
-        conn = self._local.get(user_id)
-        if conn:
-            fut = conn.pending_calls.pop(call_id, None)
-            if fut and not fut.done():
-                fut.set_result(result)
-
-    async def resolve_and_publish(self, user_id: str, call_id: str, result: dict):
-        """Chiamato quando il device locale invia un tool_result."""
-        self.resolve_local(user_id, call_id, result)
-        # Pubblica anche su Redis per l'istanza remota che aspetta
-        await self._redis.publish(
-            f"tool_result:{call_id}",
-            json.dumps(result),
-        )
-
-    # ── Listener Redis ──
-
-    async def _listen_tool_calls(self):
-        """Loop che ascolta i tool_call in arrivo da altre istanze."""
-        async for message in self._pubsub.listen():
-            if message["type"] != "message":
-                continue
-            channel = message["channel"]
-            if isinstance(channel, bytes):
-                channel = channel.decode()
-
-            data = json.loads(message["data"])
-
-            if channel.startswith("tool_call:"):
-                # Un'altra istanza vuole che inviamo un tool_call al nostro device
-                user_id = channel.split(":", 1)[1]
-                conn = self._local.get(user_id)
-                if conn:
-                    await conn.ws.send_text(json.dumps({"type": "tool_call", **data}))
-
-            elif channel.startswith("tool_result:"):
-                # Risposta a un tool_call che abbiamo inviato tramite Redis
-                call_id = channel.split(":", 1)[1]
-                fut = self._remote_futures.pop(call_id, None)
-                if fut and not fut.done():
-                    fut.set_result(data)
-
-    # ── Stream cross-instance ──
-
-    async def publish_stream_chunk(self, user_id: str, chunk: dict):
-        """Pubblica un chunk di streaming su Redis (per REST→WS relay)."""
-        await self._redis.publish(f"stream:{user_id}", json.dumps(chunk))
-```
-
----
-
-## 4. Struttura Directory Proposta (MVP)
-
-```
-adiuva-api/
-├── docker-compose.yml          # Orchestrazione completa
-├── docker-compose.dev.yml      # Override per sviluppo locale
-├── shared/                     # Codice condiviso (montato come volume)
-│   ├── auth.py                 # JWT verification (chiave pubblica)
-│   ├── schemas.py              # Pydantic schemas condivisi
-│   ├── middleware/
-│   │   ├── rate_limit.py       # DistributedRateLimiter (Redis)
-│   │   └── sanitizer.py
-│   └── models/
-│       └── base.py             # SQLAlchemy base condivisa
-│
-├── auth-service/
-│   ├── Dockerfile
-│   ├── requirements.txt
-│   └── app/
-│       ├── main.py
-│       ├── config.py
-│       ├── db.py
-│       ├── models.py           # users, refresh_tokens
-│       ├── routes/
-│       │   └── auth.py
-│       └── services/
-│           ├── jwt_service.py  # RS256 signing
-│           └── user_service.py
-│
-├── chat-service/
-│   ├── Dockerfile
-│   ├── requirements.txt
-│   └── app/
-│       ├── main.py
-│       ├── config.py
-│       ├── db.py
-│       ├── models.py           # memory_*
-│       ├── routes/
-│       │   ├── device_ws.py    # WS connection owner
-│       │   └── chat.py         # REST fallback
-│       ├── core/
-│       │   ├── device_manager.py   # RedisDeviceManager
-│       │   ├── deep_agent.py       # Home + floating chat
-│       │   ├── memory_middleware.py
-│       │   ├── ws_context.py
-│       │   ├── output_formatter.py
-│       │   └── llm.py
-│       └── agents/                 # Tool definitions (used by deep_agent)
-│           ├── task_agent.py
-│           ├── project_agent.py
-│           ├── note_agent.py
-│           └── timeline_agent.py
-│
-├── agent-service/
-│   ├── Dockerfile
-│   ├── requirements.txt
-│   └── app/
-│       ├── main.py
-│       ├── config.py
-│       ├── db.py
-│       ├── models.py           # agent_run_logs, local/cloud_agent_configs
-│       ├── routes/
-│       │   ├── agents.py       # catalog, can-create, trigger
-│       │   └── agent_setup.py  # journey start/message
-│       ├── core/
-│       │   ├── agent_runner.py     # Batch classify → process
-│       │   ├── agent_registry.py
-│       │   ├── redis_executor.py   # execute_on_client via Redis pub/sub
-│       │   └── llm.py
-│       └── agents/
-│           ├── task_agent.py       # Tool definitions (batch context)
-│           ├── project_agent.py
-│           ├── note_agent.py
-│           ├── timeline_agent.py
-│           └── filesystem_agent.py
-│
-├── billing-service/
-│   ├── Dockerfile
-│   ├── requirements.txt
-│   └── app/
-│       ├── main.py
-│       ├── config.py
-│       ├── db.py
-│       ├── models.py           # subscriptions
-│       ├── routes/
-│       │   └── billing.py
-│       └── services/
-│           ├── stripe_service.py
-│           └── tier_manager.py
-│
-└── infra/
-    ├── traefik/
-    │   └── traefik.yml
-    ├── keys/
-    │   ├── jwt_private.pem     # Solo auth-service
-    │   └── jwt_public.pem      # Tutti i servizi
-    └── alembic/                # Migrazioni condivise o per-servizio
-```
-
----
-
-## 5. Docker Compose — Configurazione MVP
-
-```yaml
-# docker-compose.yml
-
-services:
-
-  # ══════════════════════════════════════════════════════════
-  # API Gateway
-  # ══════════════════════════════════════════════════════════
-  traefik:
-    image: traefik:v3.2
-    command:
-      - "--api.insecure=true"
-      - "--providers.docker=true"
-      - "--providers.docker.exposedbydefault=false"
-      - "--entrypoints.web.address=:80"
-      - "--entrypoints.websecure.address=:443"
-      - "--entrypoints.web.http.redirections.entrypoint.to=websecure"
-    ports:
-      - "80:80"
-      - "443:443"
-      - "8080:8080"   # Dashboard Traefik (disabilitare in prod)
-    volumes:
-      - /var/run/docker.sock:/var/run/docker.sock:ro
-      - ./infra/certs:/certs:ro
-    restart: unless-stopped
-
-  # ══════════════════════════════════════════════════════════
-  # Auth Service (2 repliche)
-  # ══════════════════════════════════════════════════════════
-  auth-service:
-    build: ./auth-service
-    deploy:
-      replicas: 2
-    env_file: .env
-    environment:
-      DATABASE_URL: postgresql+asyncpg://postgres:postgres@db:5432/adiuva
-      REDIS_URL: redis://redis:6379
-      JWT_PRIVATE_KEY_FILE: /run/secrets/jwt_private_key
-      SERVICE_NAME: auth
-    secrets:
-      - jwt_private_key
-      - jwt_public_key
-    labels:
-      - "traefik.enable=true"
-      - "traefik.http.routers.auth.rule=PathPrefix(`/api/v1/auth`)"
-      - "traefik.http.services.auth.loadbalancer.server.port=8000"
-    depends_on:
-      db:
-        condition: service_healthy
-      redis:
-        condition: service_healthy
-
-  # ══════════════════════════════════════════════════════════
-  # Chat Service — Real-time WS + Chat (scalabile)
-  # ══════════════════════════════════════════════════════════
-  chat-service:
-    build: ./chat-service
-    deploy:
-      replicas: 2
-    env_file: .env
-    environment:
-      DATABASE_URL: postgresql+asyncpg://postgres:postgres@db:5432/adiuva
-      REDIS_URL: redis://redis:6379
-      JWT_PUBLIC_KEY_FILE: /run/secrets/jwt_public_key
-      SERVICE_NAME: chat
-    secrets:
-      - jwt_public_key
-    labels:
-      - "traefik.enable=true"
-      # REST chat endpoint
-      - "traefik.http.routers.chat.rule=PathPrefix(`/api/v1/chat`)"
-      - "traefik.http.services.chat.loadbalancer.server.port=8000"
-      # WebSocket route con sticky session
-      - "traefik.http.routers.ws.rule=PathPrefix(`/api/v1/ws`)"
-      - "traefik.http.routers.ws.service=chat-ws"
-      - "traefik.http.services.chat-ws.loadbalancer.server.port=8000"
-      - "traefik.http.services.chat-ws.loadbalancer.sticky.cookie.name=ws_affinity"
-      - "traefik.http.services.chat-ws.loadbalancer.sticky.cookie.httpOnly=true"
-    depends_on:
-      db:
-        condition: service_healthy
-      redis:
-        condition: service_healthy
-
-  # ══════════════════════════════════════════════════════════
-  # Agent Service — Batch processing (scalabile indipendentemente)
-  # ══════════════════════════════════════════════════════════
-  agent-service:
-    build: ./agent-service
-    deploy:
-      replicas: 2
-    env_file: .env
-    environment:
-      DATABASE_URL: postgresql+asyncpg://postgres:postgres@db:5432/adiuva
-      REDIS_URL: redis://redis:6379
-      JWT_PUBLIC_KEY_FILE: /run/secrets/jwt_public_key
-      SERVICE_NAME: agent
-    secrets:
-      - jwt_public_key
-    labels:
-      - "traefik.enable=true"
-      - "traefik.http.routers.agents.rule=PathPrefix(`/api/v1/agents`)"
-      - "traefik.http.services.agents.loadbalancer.server.port=8000"
-    depends_on:
-      db:
-        condition: service_healthy
-      redis:
-        condition: service_healthy
-
-  # ══════════════════════════════════════════════════════════
-  # Billing Service (1 replica)
-  # ══════════════════════════════════════════════════════════
-  billing-service:
-    build: ./billing-service
-    deploy:
-      replicas: 1
-    env_file: .env
-    environment:
-      DATABASE_URL: postgresql+asyncpg://postgres:postgres@db:5432/adiuva
-      REDIS_URL: redis://redis:6379
-      JWT_PUBLIC_KEY_FILE: /run/secrets/jwt_public_key
-      SERVICE_NAME: billing
-    secrets:
-      - jwt_public_key
-    labels:
-      - "traefik.enable=true"
-      - "traefik.http.routers.billing.rule=PathPrefix(`/api/v1/billing`)"
-      - "traefik.http.services.billing.loadbalancer.server.port=8000"
-    depends_on:
-      db:
-        condition: service_healthy
-      redis:
-        condition: service_healthy
-
-  # ══════════════════════════════════════════════════════════
-  # Infrastruttura
-  # ══════════════════════════════════════════════════════════
-  db:
-    image: pgvector/pgvector:pg16
-    environment:
-      POSTGRES_USER: postgres
-      POSTGRES_PASSWORD: postgres
-      POSTGRES_DB: adiuva
-    volumes:
-      - postgres_data:/var/lib/postgresql/data
-    healthcheck:
-      test: ["CMD-SHELL", "pg_isready -U postgres"]
-      interval: 5s
-      timeout: 5s
-      retries: 5
-    restart: unless-stopped
-
-  redis:
-    image: redis:7-alpine
-    command: redis-server --maxmemory 256mb --maxmemory-policy allkeys-lru
-    volumes:
-      - redis_data:/data
-    healthcheck:
-      test: ["CMD", "redis-cli", "ping"]
-      interval: 5s
-      timeout: 3s
-      retries: 5
-    restart: unless-stopped
-
-  qdrant:
-    image: qdrant/qdrant:latest
-    volumes:
-      - qdrant_data:/qdrant/storage
-    restart: unless-stopped
-
-secrets:
-  jwt_private_key:
-    file: ./infra/keys/jwt_private.pem
-  jwt_public_key:
-    file: ./infra/keys/jwt_public.pem
-
-volumes:
-  postgres_data:
-  redis_data:
-  qdrant_data:
-```
-
----
-
-## 6. Configurazione Cloudflare + VPS
-
-### 6.1 DNS
-
-```
-api.tuodominio.com  →  A record  →  IP del VPS
-                    →  Proxy: ON (orange cloud)
-```
-
-### 6.2 Cloudflare Settings
-
-| Setting | Valore | Motivo |
-|---------|--------|--------|
-| SSL/TLS mode | **Full (Strict)** | Cloudflare ↔ VPS con certificato valido |
-| WebSocket | **ON** | Necessario per `/api/v1/ws/device` |
-| Proxy timeout | **100s** (Enterprise) o default | Le LLM calls possono durare 30s+ |
-| Under Attack Mode | Off (attivare se necessario) | |
-
-### 6.3 TLS sul VPS
-
-Due opzioni:
-- **Opzione A (consigliata)**: Cloudflare Origin Certificate → montato in Traefik
-- **Opzione B**: Let's Encrypt via Traefik (con DNS challenge Cloudflare)
-
-```yaml
-# traefik.yml — con Cloudflare Origin Certificate
-entryPoints:
-  websecure:
-    address: ":443"
-
-tls:
-  certificates:
-    - certFile: /certs/origin.pem
-      keyFile: /certs/origin-key.pem
-```
-
-### 6.4 Rete VPS
-
-```bash
-# UFW firewall — solo Cloudflare può raggiungere le porte 80/443
-# https://www.cloudflare.com/ips/
-ufw default deny incoming
-ufw allow from 173.245.48.0/20 to any port 443
-ufw allow from 103.21.244.0/22 to any port 443
-# ... (tutti gli IP range di Cloudflare)
-ufw allow ssh
-ufw enable
-```
-
----
-
-## 7. Comunicazione Inter-Servizio
-
-### 7.1 Redis Pub/Sub — Event Bus
-
-```
-┌──────────┐  tier_changed:user_123   ┌──────────┐
-│ Billing  │ ────────────────────────► │   Auth   │
-│ Service  │                           │ Service  │
-└──────────┘                           └──────────┘
-
-┌──────────┐  tool_call:user_123      ┌──────────┐
-│  Agent   │ ────────────────────────► │   Chat   │
-│ Service  │                           │ Service  │
-│ (batch)  │ ◄────────────────────────│ (ha WS)  │
-└──────────┘  tool_result:{call_id}    └──────────┘
-```
-
-### 7.2 Health Checks e Service Discovery
-
-Traefik gestisce automaticamente il service discovery via Docker labels. I servizi non devono conoscersi tra loro — comunicano solo via:
-- **Redis pub/sub** (tool-call cross-instance, tier events)
-- **Redis hash** (stato condiviso: `ws:connections`, rate-limit counters)
-- **PostgreSQL** (dati persistenti condivisi)
-
----
-
-## 8. Piano di Migrazione Incrementale (MVP)
-
-### Fase 1 — Preparazione (nel monolite attuale)
-1. Aggiungere Redis al `docker-compose.yml` attuale
-2. Migrare JWT da HS256 → RS256 (backward-compatible: accetta entrambi per un periodo)
-3. Implementare `RedisDeviceManager` come drop-in replacement del singleton in-memory
-4. Estrarre `shared/` con auth verification, schemas, middleware
-
-### Fase 2 — Auth Service (primo split)
-1. Estrarre `auth.py` routes + models in `auth-service/`
-2. Verificare che i JWT firmati da `auth-service` vengano validati dal monolite
-3. Aggiungere Traefik e routare `/api/v1/auth/*` al nuovo servizio
-4. Il monolite continua a servire tutto il resto
-
-### Fase 3 — Billing Service
-1. Estrarre billing routes, Stripe service, tier manager
-2. Configurare Redis pub/sub per `tier_changed` events
-3. Routare via Traefik
-
-### Fase 4 — Split Chat + Agent (il più delicato)
-1. Il monolite residuo contiene WS + chat + agents
-2. Separare Agent Service: estrarre `agent_runner`, `agent_registry`, `agent_setup`, route `/agents/*`
-3. Implementare `redis_executor.py` nell'Agent Service per tool-call via Redis
-4. Il Chat Service resta proprietario della WS e sottoscrive i canali `tool_call:{user_id}`
-5. Testare: trigger agent dall'Agent Service → tool_call via Redis → Chat Service → WS → device → risposta
-
-### Fase 5 — Scaling test
-1. Scalare Chat Service a 2 repliche, verificare sticky sessions
-2. Scalare Agent Service a 2 repliche, verificare batch processing distribuito
-3. Monitoring (Prometheus + Grafana) per ogni servizio
-
----
-
-## 9. Monitoraggio e Logging
-
-```yaml
-# Aggiungere al docker-compose.yml
-
-  prometheus:
-    image: prom/prometheus:latest
-    volumes:
-      - ./infra/prometheus/prometheus.yml:/etc/prometheus/prometheus.yml
-    restart: unless-stopped
-
-  grafana:
-    image: grafana/grafana:latest
-    ports:
-      - "3000:3000"
-    volumes:
-      - grafana_data:/var/lib/grafana
-    restart: unless-stopped
-
-  loki:
-    image: grafana/loki:latest
-    restart: unless-stopped
-```
-
-Ogni servizio espone `/metrics` (Prometheus) e scrive log strutturati (JSON) raccolti da Loki.
-
----
-
-## 10. Sizing VPS Minimo Consigliato (MVP)
-
-| Componente | CPU | RAM | Note |
-|---|---|---|---|
-| Traefik | 0.25 | 128MB | |
-| Auth Service ×2 | 0.25 ×2 | 128MB ×2 | Stateless, leggero |
-| Chat Service ×2 | 1.0 ×2 | 1GB ×2 | WS + streaming LLM |
-| Agent Service ×2 | 0.75 ×2 | 512MB ×2 | Batch LLM, CPU-bound |
-| Billing Service | 0.25 | 128MB | |
-| PostgreSQL | 1.0 | 1GB | |
-| Redis | 0.25 | 256MB | |
-| Qdrant | 0.5 | 512MB | |
-| **Totale MVP** | **~5.5 vCPU** | **~5 GB** | |
-
-**Raccomandazione**: VPS con **8 vCPU / 16 GB RAM** per avere margine. Hetzner CPX41 (~€30/mese) o equivalente. Senza Storage/Plugin si risparmia ~1 vCPU e 512MB rispetto alla versione completa.
-
----
-
-## Riepilogo Architettura MVP
-
-| Servizio | Repliche | Proprietario di |
-|---|---|---|
-| **Traefik** | 1 | Routing, TLS, sticky sessions |
-| **Auth Service** | 2 | JWT RS256, registrazione, login, profilo |
-| **Chat Service** | 2–N | WebSocket, home/floating chat, streaming |
-| **Agent Service** | 2–N | Batch processing, directory scan, agent setup |
-| **Billing Service** | 1 | Stripe, subscriptions, tier management |
-
-| Decisione | Scelta | Motivazione |
-|---|---|---|
-| API Gateway | Traefik | Nativo Docker, WebSocket support, service discovery automatico |
-| JWT | RS256 (asimmetrico) | Verifica distribuita senza contattare Auth Service |
-| Tier check | Claim nel JWT | Ogni servizio verifica localmente, zero roundtrip |
-| WebSocket scaling | Redis pub/sub + sticky cookies | Cross-instance tool-call routing |
-| Chat ↔ Agent split | Servizi separati | Batch CPU-bound non impatta real-time chat |
-| Agent → Device comms | Redis pub/sub via Chat Service | Agent non possiede la WS, usa un relay |
-| Rate limiting | Redis contatori distribuiti | Sliding window condivisa tra repliche |
-| Database | PostgreSQL condiviso | Semplicità MVP; split DB futuro facile |
-| TLS | Cloudflare Origin Certificate | Zero maintenance |
-| Orchestrazione | Docker Compose | Sufficiente per un singolo VPS |
-| Storage / Plugin | Post-MVP | Non critici per il lancio |

requirements.txt

@@ -0,0 +1,43 @@
+fastapi>=0.115.0
+uvicorn[standard]>=0.34.0
+gunicorn>=22.0.0
+langchain>=0.3.0
+langchain-openai>=0.3.0
+langchain-litellm>=0.1.0
+litellm>=1.50.0
+pydantic>=2.10.0
+pydantic-settings>=2.7.0
+python-jose[cryptography]>=3.3.0
+stripe>=11.0.0
+boto3>=1.35.0
+slowapi>=0.1.9
+sqlalchemy>=2.0.0
+asyncpg>=0.30.0
+alembic>=1.14.0
+bcrypt>=4.2.0
+python-dotenv>=1.0.0
+httpx>=0.28.0
+websockets>=14.0
+psycopg2-binary>=2.9.0
+pytest>=8.0.0
+pytest-asyncio>=0.24.0
+aiosqlite>=0.20.0
+moto[s3]>=5.0.0
+pinecone>=5.0.0
+qdrant-client>=1.7.0
+croniter>=3.0.0
+google-api-python-client>=2.130.0
+google-auth>=2.29.0
+google-auth-oauthlib>=1.2.0
+google-auth-httplib2>=0.2.0
+msal>=1.28.0
+cryptography>=42.0.0
+pgvector>=0.2.5
+langfuse>=3.3.1
+beautifulsoup4>=4.12.0
+lxml>=5.0.0
+PyYAML>=6.0.0
+apscheduler>=3.10.0
+ruff>=0.8.0
+pypdf>=4.0
+python-docx>=1.1

services/auth/.env.example

@@ -1,19 +0,0 @@
-# ── Auth Service ──────────────────────────────────────────────────────────────
-# This file contains env vars specific to the Auth Service.
-# Shared vars (DATABASE_URL, REDIS_URL, etc.) come from the root .env
-# or from docker-compose environment.
-
-# ── JWT RS256 Keys ────────────────────────────────────────────────────────────
-# Generate keypair:
-#   openssl genpkey -algorithm RSA -out private.pem -pkeyopt rsa_keygen_bits:2048
-#   openssl rsa -in private.pem -pubout -out public.pem
-#
-# Paste PEM content with literal \n for newlines:
-#   JWT_PRIVATE_KEY=-----BEGIN PRIVATE KEY-----\nMIIEvQ...
-#   JWT_PUBLIC_KEY=-----BEGIN PUBLIC KEY-----\nMIIBIj...
-
-# PRIVATE KEY — used to SIGN JWTs. NEVER share outside this service.
-JWT_PRIVATE_KEY=
-
-# PUBLIC KEY — used to VERIFY JWTs.
-JWT_PUBLIC_KEY=

services/auth/Dockerfile

@@ -1,36 +0,0 @@
-# ── builder ──────────────────────────────────────────────────────────────────
-FROM python:3.12-slim AS builder
-
-WORKDIR /build
-
-# Install shared + service deps in one layer
-COPY services/auth/requirements.txt ./requirements.txt
-RUN pip install --upgrade pip && \
-    pip install --no-cache-dir --prefix=/install -r requirements.txt
-
-# ── runtime ──────────────────────────────────────────────────────────────────
-FROM python:3.12-slim AS runtime
-
-RUN addgroup --system appgroup && adduser --system --ingroup appgroup appuser
-
-WORKDIR /app
-
-COPY --from=builder /install /usr/local
-
-# Copy shared module (available to all services)
-COPY shared/ shared/
-
-# Copy service source
-COPY services/auth/app/ app/
-
-RUN chown -R appuser:appgroup /app
-
-USER appuser
-
-EXPOSE 8000
-
-CMD ["gunicorn", "app.main:app", \
-     "-k", "uvicorn.workers.UvicornWorker", \
-     "--bind", "0.0.0.0:8000", \
-     "--workers", "2", \
-     "--timeout", "30"]

services/auth/README.md

@@ -1,16 +0,0 @@
-# Auth Service
-
-Owns: user registration, login, JWT RS256 issuance, token refresh, `/me` endpoint.
-
-## Tables owned
-- `users`
-- `refresh_tokens`
-- `subscriptions` (read; Billing Service writes)
-
-## Endpoints
-- `POST /auth/register`
-- `POST /auth/login`
-- `POST /auth/refresh`
-- `GET /auth/me`
-- `PUT /auth/me`
-- `GET /auth/verify` (ForwardAuth for Traefik)

services/auth/app/config.py

@@ -1,34 +0,0 @@
-"""Auth Service — local configuration.
-
-Contains secrets that ONLY the Auth Service needs (e.g., JWT private key).
-These are NOT in shared/config.py to prevent other services from accessing them.
-"""
-
-from pydantic import field_validator
-from pydantic_settings import BaseSettings, SettingsConfigDict
-
-
-class AuthSettings(BaseSettings):
-    # RS256 private key (PEM format). Used to SIGN JWTs.
-    # Only the Auth Service has this. Generate with:
-    #   openssl genpkey -algorithm RSA -out private.pem -pkeyopt rsa_keygen_bits:2048
-    # Then set the env var (newlines as \n):
-    #   JWT_PRIVATE_KEY="-----BEGIN PRIVATE KEY-----\nMIIEv..."
-    JWT_PRIVATE_KEY: str = ""
-
-    # RS256 public key (PEM format). Used to VERIFY JWTs.
-    # Derived from the private key:
-    #   openssl rsa -in private.pem -pubout -out public.pem
-    JWT_PUBLIC_KEY: str = ""
-
-    @field_validator("JWT_PRIVATE_KEY", "JWT_PUBLIC_KEY", mode="before")
-    @classmethod
-    def _expand_pem_newlines(cls, v: str) -> str:
-        if isinstance(v, str) and r"\n" in v:
-            return v.replace(r"\n", "\n")
-        return v
-
-    model_config = SettingsConfigDict(env_file=".env", env_file_encoding="utf-8")
-
-
-auth_settings = AuthSettings()

services/auth/app/deps.py

@@ -1,69 +0,0 @@
-"""Auth dependencies — JWT validation for the Auth Service.
-
-This is the canonical get_current_user used by protected endpoints
-within the Auth Service itself (/me, /me PUT).
-"""
-
-from __future__ import annotations
-
-from fastapi import Depends, HTTPException, status
-from fastapi.security import OAuth2PasswordBearer
-from jose import JWTError, jwt
-from sqlalchemy import select
-from sqlalchemy.ext.asyncio import AsyncSession
-
-from shared.config import settings
-from shared.db import get_session
-from shared.models import Subscription, User
-from shared.schemas import UserProfile
-
-from app.config import auth_settings
-
-oauth2_scheme = OAuth2PasswordBearer(tokenUrl="/api/v1/auth/login")
-
-
-async def get_current_user(
-    token: str = Depends(oauth2_scheme),
-    db: AsyncSession = Depends(get_session),
-) -> UserProfile:
-    """Validate a Bearer JWT and return the authenticated user.
-
-    The JWT is used for identity and expiry.  Tier is fetched live from the
-    subscriptions table so upgrades/downgrades take effect immediately.
-    """
-    credentials_exc = HTTPException(
-        status_code=status.HTTP_401_UNAUTHORIZED,
-        detail="Could not validate credentials",
-        headers={"WWW-Authenticate": "Bearer"},
-    )
-    try:
-        payload = jwt.decode(
-            token, auth_settings.JWT_PUBLIC_KEY, algorithms=["RS256"]
-        )
-        user_id: str | None = payload.get("sub")
-        email: str | None = payload.get("email")
-        if not user_id or not email:
-            raise credentials_exc
-    except JWTError:
-        raise credentials_exc
-
-    # Live tier lookup
-    result = await db.execute(
-        select(Subscription.tier).where(Subscription.user_id == user_id)
-    )
-    default_tier = "power" if settings.ENV == "dev" else "free"
-    tier: str = result.scalar_one_or_none() or default_tier
-
-    # Fetch name/surname
-    user_result = await db.execute(
-        select(User.name, User.surname).where(User.id == user_id)
-    )
-    user_row = user_result.one_or_none()
-
-    return UserProfile(
-        id=user_id,
-        email=email,
-        name=user_row.name if user_row else None,
-        surname=user_row.surname if user_row else None,
-        tier=tier,
-    )  # type: ignore[arg-type]

services/auth/app/main.py

@@ -1,62 +0,0 @@
-"""Auth Service — JWT issuance, user management, ForwardAuth verification.
-
-Standalone FastAPI service extracted from the adiuva-api monolith.
-Owns: users, refresh_tokens, subscriptions (read).
-"""
-
-import sys
-from contextlib import asynccontextmanager
-from pathlib import Path
-
-# Ensure the repo root is on sys.path so "shared" is importable.
-# In Docker, COPY shared/ puts it at /app/shared/ (already importable).
-# In local dev, we need to add the repo root (two levels up from this file).
-_repo_root = str(Path(__file__).resolve().parents[3])
-if _repo_root not in sys.path:
-    sys.path.insert(0, _repo_root)
-
-from fastapi import FastAPI
-from fastapi.middleware.cors import CORSMiddleware
-
-from shared.config import settings
-
-
-@asynccontextmanager
-async def lifespan(app: FastAPI):
-    yield
-    from shared.db import engine
-
-    await engine.dispose()
-
-
-def create_app() -> FastAPI:
-    app = FastAPI(
-        title="Adiuva Auth Service",
-        version="0.1.0",
-        docs_url="/docs" if settings.ENV == "dev" else None,
-        redoc_url=None,
-        lifespan=lifespan,
-    )
-
-    app.add_middleware(
-        CORSMiddleware,
-        allow_origins=settings.CORS_ORIGINS,
-        allow_credentials=True,
-        allow_methods=["*"],
-        allow_headers=["*"],
-    )
-
-    from app.routes import router
-    from app.verify import router as verify_router
-
-    app.include_router(router, prefix="/api/v1")
-    app.include_router(verify_router, prefix="/api/v1")
-
-    @app.get("/api/v1/health", tags=["health"])
-    async def health() -> dict:
-        return {"status": "ok", "service": "auth", "version": app.version}
-
-    return app
-
-
-app = create_app()

services/auth/app/routes.py

@@ -1,249 +0,0 @@
-"""Auth routes: register, login, refresh, me.
-
-Extracted from app/api/routes/auth.py — uses shared.* imports instead of app.*.
-"""
-
-from __future__ import annotations
-
-import hashlib
-import time
-import uuid
-from datetime import datetime, timedelta, timezone
-
-import bcrypt
-from cryptography.fernet import Fernet
-from fastapi import APIRouter, Depends, HTTPException, status
-from jose import jwt
-from pydantic import BaseModel
-from sqlalchemy import select
-from sqlalchemy.ext.asyncio import AsyncSession
-
-from shared.config import settings
-from shared.db import get_session
-from shared.models import RefreshToken, Subscription, User
-from shared.schemas import AuthTokens, UserProfile
-
-from app.config import auth_settings
-from app.deps import get_current_user
-
-router = APIRouter(prefix="/auth", tags=["auth"])
-
-
-# ── Internal helpers ─────────────────────────────────────────────────
-
-
-def _hash_password(password: str) -> str:
-    return bcrypt.hashpw(password.encode(), bcrypt.gensalt()).decode()
-
-
-def _verify_password(password: str, hashed: str) -> bool:
-    return bcrypt.checkpw(password.encode(), hashed.encode())
-
-
-def _hash_token(plain_token: str) -> str:
-    """SHA-256 of the plain refresh token string."""
-    return hashlib.sha256(plain_token.encode()).hexdigest()
-
-
-def _make_access_token(user_id: str, email: str, tier: str) -> tuple[str, int]:
-    """Return (RS256-signed JWT, expires_at_ms)."""
-    now = int(time.time())
-    exp = now + settings.JWT_ACCESS_TOKEN_EXPIRE_MINUTES * 60
-    payload = {
-        "sub": user_id,
-        "email": email,
-        "tier": tier,
-        "exp": exp,
-        "iat": now,
-    }
-    token = jwt.encode(payload, auth_settings.JWT_PRIVATE_KEY, algorithm="RS256")
-    return token, exp * 1000  # ms for client
-
-
-async def _get_live_tier(db: AsyncSession, user_id: str) -> str:
-    """Fetch authoritative tier from subscriptions table."""
-    result = await db.execute(
-        select(Subscription.tier).where(Subscription.user_id == user_id)
-    )
-    default_tier = "power" if settings.ENV == "dev" else "free"
-    return result.scalar_one_or_none() or default_tier
-
-
-# ── Request bodies ────────────────────────────────────────────────────
-
-
-class _RegisterRequest(BaseModel):
-    email: str
-    password: str
-    name: str | None = None
-    surname: str | None = None
-
-
-class _LoginRequest(BaseModel):
-    email: str
-    password: str
-
-
-class _RefreshRequest(BaseModel):
-    refresh_token: str
-
-
-class _UpdateProfileRequest(BaseModel):
-    name: str | None = None
-    surname: str | None = None
-
-
-# ── Routes ────────────────────────────────────────────────────────────
-
-
-@router.post("/register", response_model=AuthTokens, status_code=status.HTTP_201_CREATED)
-async def register(
-    body: _RegisterRequest,
-    db: AsyncSession = Depends(get_session),
-) -> AuthTokens:
-    """Create a new account and return JWT tokens."""
-    existing = await db.execute(select(User).where(User.email == body.email))
-    if existing.scalar_one_or_none() is not None:
-        raise HTTPException(status.HTTP_409_CONFLICT, "Email already registered")
-
-    user = User(
-        id=str(uuid.uuid4()),
-        email=body.email,
-        name=body.name,
-        surname=body.surname,
-        password_hash=_hash_password(body.password),
-        tier="free",
-        encryption_key=Fernet.generate_key().decode(),
-    )
-    db.add(user)
-    await db.flush()
-
-    plain_token = str(uuid.uuid4())
-    expires_at = datetime.now(timezone.utc) + timedelta(
-        days=settings.JWT_REFRESH_TOKEN_EXPIRE_DAYS
-    )
-    rt = RefreshToken(
-        user_id=user.id,
-        token_hash=_hash_token(plain_token),
-        expires_at=expires_at,
-    )
-    db.add(rt)
-    await db.commit()
-
-    access_token, expires_at_ms = _make_access_token(user.id, user.email, user.tier)
-    return AuthTokens(
-        access_token=access_token,
-        refresh_token=plain_token,
-        expires_at=expires_at_ms,
-    )
-
-
-@router.post("/login", response_model=AuthTokens)
-async def login(
-    body: _LoginRequest,
-    db: AsyncSession = Depends(get_session),
-) -> AuthTokens:
-    """Validate credentials and return JWT tokens."""
-    result = await db.execute(select(User).where(User.email == body.email))
-    user = result.scalar_one_or_none()
-    if user is None or not _verify_password(body.password, user.password_hash):
-        raise HTTPException(status.HTTP_401_UNAUTHORIZED, "Invalid credentials")
-
-    # Fetch live tier for the JWT claim
-    tier = await _get_live_tier(db, user.id)
-
-    plain_token = str(uuid.uuid4())
-    expires_at = datetime.now(timezone.utc) + timedelta(
-        days=settings.JWT_REFRESH_TOKEN_EXPIRE_DAYS
-    )
-    rt = RefreshToken(
-        user_id=user.id,
-        token_hash=_hash_token(plain_token),
-        expires_at=expires_at,
-    )
-    db.add(rt)
-    await db.commit()
-
-    access_token, expires_at_ms = _make_access_token(user.id, user.email, tier)
-    return AuthTokens(
-        access_token=access_token,
-        refresh_token=plain_token,
-        expires_at=expires_at_ms,
-    )
-
-
-@router.post("/refresh", response_model=AuthTokens)
-async def refresh(
-    body: _RefreshRequest,
-    db: AsyncSession = Depends(get_session),
-) -> AuthTokens:
-    """Rotate a refresh token and return a new token pair."""
-    token_hash = _hash_token(body.refresh_token)
-    result = await db.execute(
-        select(RefreshToken).where(RefreshToken.token_hash == token_hash)
-    )
-    rt = result.scalar_one_or_none()
-
-    now = datetime.now(timezone.utc)
-    if rt is None or rt.expires_at.replace(tzinfo=timezone.utc) < now:
-        raise HTTPException(status.HTTP_401_UNAUTHORIZED, "Invalid or expired refresh token")
-
-    await db.delete(rt)
-
-    user_result = await db.execute(select(User).where(User.id == rt.user_id))
-    user = user_result.scalar_one_or_none()
-    if user is None:
-        raise HTTPException(status.HTTP_401_UNAUTHORIZED, "User not found")
-
-    # Fetch live tier for the new JWT
-    tier = await _get_live_tier(db, user.id)
-
-    plain_token = str(uuid.uuid4())
-    new_expires = now + timedelta(days=settings.JWT_REFRESH_TOKEN_EXPIRE_DAYS)
-    new_rt = RefreshToken(
-        user_id=user.id,
-        token_hash=_hash_token(plain_token),
-        expires_at=new_expires,
-    )
-    db.add(new_rt)
-    await db.commit()
-
-    access_token, expires_at_ms = _make_access_token(user.id, user.email, tier)
-    return AuthTokens(
-        access_token=access_token,
-        refresh_token=plain_token,
-        expires_at=expires_at_ms,
-    )
-
-
-@router.get("/me", response_model=UserProfile)
-async def me(current_user: UserProfile = Depends(get_current_user)) -> UserProfile:
-    """Return the profile for the authenticated user."""
-    return current_user
-
-
-@router.put("/me", response_model=UserProfile)
-async def update_profile(
-    body: _UpdateProfileRequest,
-    current_user: UserProfile = Depends(get_current_user),
-    db: AsyncSession = Depends(get_session),
-) -> UserProfile:
-    """Update the authenticated user's name and surname."""
-    result = await db.execute(select(User).where(User.id == current_user.id))
-    user = result.scalar_one()
-
-    if body.name is not None:
-        user.name = body.name
-    if body.surname is not None:
-        user.surname = body.surname
-
-    await db.commit()
-    await db.refresh(user)
-
-    return UserProfile(
-        id=user.id,
-        email=user.email,
-        name=user.name,
-        surname=user.surname,
-        tier=current_user.tier,
-    )

services/auth/app/verify.py

@@ -1,66 +0,0 @@
-"""ForwardAuth verification endpoint for Traefik.
-
-Traefik calls GET /api/v1/auth/verify on every request to a protected
-service.  This endpoint validates the JWT from the Authorization header
-and returns identity headers that Traefik injects into downstream requests.
-
-Downstream services NEVER validate JWTs themselves — they trust the
-X-User-Id, X-User-Email, X-User-Tier headers injected by Traefik.
-"""
-
-from __future__ import annotations
-
-from fastapi import APIRouter, Request, Response
-from fastapi import status as http_status
-from jose import JWTError, jwt
-from sqlalchemy import select
-
-from shared.config import settings
-from shared.db import async_session
-from shared.models import Subscription
-
-from app.config import auth_settings
-
-router = APIRouter(tags=["auth"])
-
-
-@router.get("/auth/verify")
-async def verify(request: Request) -> Response:
-    """Validate JWT and return identity headers for Traefik ForwardAuth.
-
-    Returns 200 with X-User-* headers on success, 401 on failure.
-    Traefik copies response headers to the downstream request.
-    """
-    auth_header = request.headers.get("Authorization", "")
-    if not auth_header.startswith("Bearer "):
-        return Response(status_code=http_status.HTTP_401_UNAUTHORIZED)
-
-    token = auth_header[7:]  # strip "Bearer "
-
-    try:
-        payload = jwt.decode(
-            token, auth_settings.JWT_PUBLIC_KEY, algorithms=["RS256"]
-        )
-        user_id: str | None = payload.get("sub")
-        email: str | None = payload.get("email")
-        if not user_id or not email:
-            return Response(status_code=http_status.HTTP_401_UNAUTHORIZED)
-    except JWTError:
-        return Response(status_code=http_status.HTTP_401_UNAUTHORIZED)
-
-    # Live tier lookup from subscriptions table
-    async with async_session() as db:
-        result = await db.execute(
-            select(Subscription.tier).where(Subscription.user_id == user_id)
-        )
-        default_tier = "power" if settings.ENV == "dev" else "free"
-        tier: str = result.scalar_one_or_none() or default_tier
-
-    return Response(
-        status_code=http_status.HTTP_200_OK,
-        headers={
-            "X-User-Id": user_id,
-            "X-User-Email": email,
-            "X-User-Tier": tier,
-        },
-    )

services/auth/requirements.txt

@@ -1,11 +0,0 @@
-fastapi>=0.115.0
-uvicorn[standard]>=0.34.0
-gunicorn>=22.0.0
-pydantic>=2.10.0
-pydantic-settings>=2.7.0
-python-jose[cryptography]>=3.3.0
-sqlalchemy>=2.0.0
-asyncpg>=0.30.0
-bcrypt>=4.2.0
-cryptography>=42.0.0
-python-dotenv>=1.0.0