fix(billing): auto-detect repo root for shared module import in local dev

- stripe_service: checkout sessions, webhook handling, subscription CRUD
- tier_manager: feature matrix (4 tiers), quota enforcement, rate limits
- routes: checkout, webhook (no auth), subscription, tier query, features
- Traefik header auth (X-User-Id) replaces get_current_user dependency
- /tier/{user_id} endpoint for internal service-to-service lookups
- /features and /features/{tier} for feature matrix queries
- Dockerfile: single worker, 30s timeout (lightweight service)

.env.example

@@ -2,69 +2,55 @@
 ENV=dev
 
 # ── Database ──────────────────────────────────────────────────────────────────
-DATABASE_URL=postgresql+asyncpg://postgres:postgres@localhost:5432/adiuvai
+DATABASE_URL=postgresql+asyncpg://postgres:postgres@localhost:5432/adiuva
 
-# ── Auth ──────────────────────────────────────────────────────────────────────
+# ── Redis ─────────────────────────────────────────────────────────────────────
-JWT_SECRET=replace-with-a-long-random-secret
+REDIS_URL=redis://localhost:6379/0
-JWT_ALGORITHM=HS256
+
+# ── Auth (JWT RS256) ──────────────────────────────────────────────────────────
+# Public key for optional local JWT verification (Traefik ForwardAuth handles
+# this in production — services trust X-User-* headers from Traefik).
+# 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_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=
-CEREBRAS_API_KEY=
+LLM_MODEL=gpt-4o
-
-# 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=
-
-# Setup-agent — guided journey to build an AgentConfig via WebSocket chat.
-LLM_MODEL_SETUP_AGENT=
 
 # ── 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
 
-# ── Langfuse (leave empty to disable observability) ───────────────────────────
+# ── Vector Store ──────────────────────────────────────────────────────────────
-LANGFUSE_SECRET_KEY=
+# Pinecone is used when PINECONE_API_KEY is set; otherwise falls back to Qdrant.
-LANGFUSE_PUBLIC_KEY=
+PINECONE_API_KEY=
-# LANGFUSE_BASE_URL=https://cloud.langfuse.com        # EU (default)
+PINECONE_INDEX=adiuva
-# LANGFUSE_BASE_URL=https://us.cloud.langfuse.com     # US
+QDRANT_URL=
-# LANGFUSE_BASE_URL=http://localhost:3000             # Self-hosted
+QDRANT_API_KEY=
+# 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

.gitea/workflows/deploy.yaml

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

.gitignore

@@ -13,6 +13,9 @@ env/
 # Environment variables
 .env
 
+# Cryptographic keys
+*.pem
+
 # IDE
 .vscode/
 .idea/
@@ -21,7 +24,6 @@ env/
 .pytest_cache/
 htmlcov/
 .coverage
-tests/fixtures/private*/
 
 # Docker
 *.log
@@ -33,3 +35,6 @@ Thumbs.db
 # Claude Code
 .claude/
 logs/
+
+# Eval private test data
+services/batch-agent/eval/fixtures/private_data/

README.md

@@ -0,0 +1,793 @@
+# Adiuva Cloud API
+
+**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
+
+```
+┌──────────────┐      ┌────────────────────────────────────────────────────────┐
+│  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 pool
+from sqlalchemy import engine_from_config, pool
 from sqlalchemy.ext.asyncio import create_async_engine
 
 # Alembic Config object (gives access to alembic.ini values).

alembic/versions/001_initial_schema.py

@@ -1,4 +1,5 @@
-"""Initial schema: users, refresh_tokens, subscriptions.
+"""Initial schema: users, refresh_tokens, subscriptions, storage_records,
+backup_metadata, plugins, plugin_installations, plugin_reviews, revenue_events.
 
 Revision ID: 001
 Revises:
@@ -27,6 +28,18 @@ 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(
@@ -75,10 +88,122 @@ 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

@@ -0,0 +1,92 @@
+"""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] = "001"
+down_revision: Union[str, None] = "002"
 branch_labels: Union[str, Sequence[str], None] = None
 depends_on: Union[str, Sequence[str], None] = None
 

alembic/versions/a3b9c0d1e2f3_add_agent_config_to_local_agents.py

@@ -1,107 +0,0 @@
-"""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

@@ -1,56 +0,0 @@
-"""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

@@ -1,31 +0,0 @@
-"""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/e04100e88ace_avatar_url_varchar_to_text.py

@@ -1,34 +0,0 @@
-"""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/filesystem_agent.py

@@ -7,31 +7,12 @@ 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:
@@ -102,93 +83,3 @@ FILESYSTEM_TOOLS: list[Any] = [
     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/note_agent.py

@@ -18,6 +18,21 @@ _UUID_RE = re.compile(
 def _is_uuid(value: str) -> bool:
     return bool(_UUID_RE.match(value))
 
+NOTE_SYSTEM_PROMPT = (
+    "You are a note-taking assistant. You help users create, retrieve, update,\n"
+    "and delete Markdown notes in their workspace.\n\n"
+    "Rules:\n"
+    "  - content is always Markdown; preserve formatting when updating\n"
+    "  - project_id is optional; link a note to a project when mentioned\n"
+    "  - When updating, call get_note first if you need to read existing content\n"
+    "    before appending or replacing sections\n"
+    "  - list_notes without project_id returns all notes; scope with project_id\n"
+    "    when the user is working within a specific project\n"
+    "  - project_id must be a UUID; if you only know a project name, do not pass it as project_id\n"
+    "  - Do not fabricate note content — reflect what the user provides or what\n"
+    "    is already in the note (retrieved via get_note)."
+)
+
 
 @tool
 async def list_notes(project_id: str = "") -> str:

app/agents/project_agent.py

@@ -8,6 +8,22 @@ from langchain_core.tools import tool
 
 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
 async def list_projects(

app/agents/task_agent.py

@@ -18,6 +18,23 @@ _UUID_RE = re.compile(
 def _is_uuid(value: str) -> bool:
     return bool(_UUID_RE.match(value))
 
+TASK_SYSTEM_PROMPT = (
+    "You are a task management assistant for a project workspace.\n"
+    "You create, update, list, and track tasks and their comments.\n\n"
+    "Rules:\n"
+    "  - status must be one of: todo, in_progress, done\n"
+    "  - priority must be one of: high, medium, low\n"
+    "  - due_date is a Unix timestamp in milliseconds; convert human dates\n"
+    "  - assignees is a JSON-encoded array of strings (e.g. '[\"Alice\",\"Bob\"]')\n"
+    "  - project_id is optional; link to a project when the user mentions one\n"
+    "  - is_ai_suggested: 1 only when proactively proposing a task the user\n"
+    "    did not explicitly request; 0 otherwise\n"
+    "  - is_ai_suggested: 1 only when proactively proposing a task the user did not explicitly request; 0 otherwise\n"
+    "  - Use list_tasks_due_today for 'what's due today' queries\n"
+    "  - For update_task, use -1 for integer fields you do not want to change\n"
+    "  - Always confirm the action in plain, user-friendly language."
+)
+
 
 # ── Task tools ────────────────────────────────────────────────────────
 

app/agents/timeline_agent.py

@@ -17,6 +17,20 @@ _UUID_RE = re.compile(
 def _is_uuid(value: str) -> bool:
     return bool(_UUID_RE.match(value))
 
+TIMELINE_SYSTEM_PROMPT = (
+    "You are a project timeline assistant. Timelines are milestone dates that\n"
+    "track progress on a project — they are not calendar events.\n\n"
+    "Rules:\n"
+    "  - project_id is REQUIRED for every create; confirm with the user if unknown\n"
+    "  - For listing, project_id must be a UUID; never pass plain names as project_id\n"
+    "  - date is a Unix timestamp in milliseconds; convert human-readable dates\n"
+    "  - is_ai_suggested: 1 when proactively proposing a timeline, 0 otherwise\n"
+    "  - is_ai_suggested: 1 when proactively proposing a timeline, 0 otherwise\n"
+    "  - For update_timeline, use -1 for integer fields you do not want to change\n"
+    "  - Listing without a project_id returns all timelines across projects\n"
+    "  - Always echo the title and formatted date in your confirmation."
+)
+
 
 @tool
 async def list_timelines(project_id: str = "") -> str:

app/api/middleware/auth.py

@@ -65,39 +65,16 @@ async def get_current_user(
     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.
+    # Fetch name/surname from user row.
     user_result = await db.execute(
-        select(
+        select(User.name, User.surname).where(User.id == user_id)
-            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/sanitizer.py

@@ -8,7 +8,8 @@ that could reveal server-side prompt IP:
   - Internal reasoning markers (<thinking>, <reasoning>, [INST], …)
   - Exact-match known prompt fingerprints
 
-The middleware only activates for paths under /api/v1/chat.
+Binary responses (storage blobs, backup data) are never touched — 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.

app/api/routes/agent_setup.py

@@ -1,11 +1,11 @@
-"""Chatbot Journey — WS-based guided conversation to build an AgentConfig.
+"""Chatbot Journey — WS-based guided conversation to build an agent prompt_template.
 
 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,
+  1. FE sends ``journey_start`` frame with basic agent config (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,
@@ -13,11 +13,10 @@ Journey flow:
   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 ``AgentConfig`` JSON
+  5. After 3-5 turns the LLM wraps up by emitting a ``prompt_template``
-     block delimited by ``AGENT_CONFIG_START`` / ``AGENT_CONFIG_END``.
+     block delimited by ``PROMPT_TEMPLATE_START`` / ``PROMPT_TEMPLATE_END``.
-  6. Server parses and validates the JSON with Pydantic, sends
+  6. Server parses the block, sends ``journey_reply`` with ``done=True``
-     ``journey_reply`` with ``done=True`` and the serialised config.
+     and the template.  FE stores it locally.
-     FE stores it locally.
 """
 
 from __future__ import annotations
@@ -31,10 +30,8 @@ from typing import Any
 
 from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
 
-from app.agents.filesystem_agent import make_directory_tools
+from app.agents.filesystem_agent import FILESYSTEM_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_llm
-from app.core.llm import get_agent_llm, model_for_agent
-from app.schemas import AgentConfig
 
 logger = logging.getLogger(__name__)
 
@@ -42,9 +39,9 @@ logger = logging.getLogger(__name__)
 
 _SESSION_TTL_SECONDS: int = 1800  # 30 minutes
 
-# Sentinel strings used to delimit the LLM-produced AgentConfig JSON.
+# Sentinel strings used to delimit the LLM-produced prompt_template.
-_CONFIG_START = "AGENT_CONFIG_START"
+_TEMPLATE_START = "PROMPT_TEMPLATE_START"
-_CONFIG_END = "AGENT_CONFIG_END"
+_TEMPLATE_END = "PROMPT_TEMPLATE_END"
 
 # Minimum turns before we consider nudging the LLM to wrap up.
 _MIN_TURNS_BEFORE_NUDGE: int = 3
@@ -65,7 +62,6 @@ class JourneySession:
     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:
@@ -87,76 +83,61 @@ def get_journey_session(session_id: str, user_id: str) -> JourneySession | None:
     return s
 
 
-# ── System prompt ─────────────────────────────────────────────────────────
+# ── System prompt builder ─────────────────────────────────────────────────
 
-_JOURNEY_SYSTEM_PROMPT = """\
+_SYSTEM_PROMPT_TEMPLATE = """\
 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
+Your job is to understand exactly what data the user wants to extract from their
-structured AgentConfig JSON that the extraction agent will use as its instruction set.
+local directory and produce a detailed prompt_template that a separate AI will use
+as its instruction set.
+
+The extraction agent already has this base behaviour built in:
+  - Reads each file using file-system tools.
+  - Creates records (tasks, notes, timelines, projects) via CRUD tools.
+  - Sets isAiSuggested=1 on every new record.
+  - Only extracts data explicitly present in the files — it never invents information.
+The user's custom prompt is appended AFTER this base behaviour, so focus on
+what to look for and how to map it — not on the general extraction mechanics.
 
 You have access to file-system tools to explore the user's directory:
-- list_directory: see folder structure and file names
+- list_directory: to see folder structure
-- read_file_content: peek at a file's content
+- read_file_content: to peek at file contents
-- get_file_metadata: check file size, extension, dates
+- get_file_metadata: to check file info
 
 The user's configured directory is: {directory}
 Target data types: {data_types}
 
-## Your process
+IMPORTANT — project assignment is handled automatically by the main agent runner
+before the custom prompt is ever used.  You MUST NOT ask the user about projects,
+projectId, or how to link records to projects.  Never include projectId logic or
+project creation instructions in the generated prompt_template.
 
-### Step 1 — Explore the directory
+Start by exploring the directory to understand its structure.  Then ask concise,
-Use list_directory and read_file_content to understand what types of files are present
+focused questions one at a time.  Cover these topics (not necessarily in this order):
-(HTML emails, plain-text documents, CSVs, etc.).
+  1. The type and format of the source content (confirmed by your exploration).
+  2. How fields should be mapped (e.g. filename → task title).
+  3. Priority or status rules (e.g. "urgent" keyword → high priority).
+  4. Any special handling, date extraction, or exclusions.
 
-### Step 2 — Identify content types
+Once you reach 90% confidence, output the final prompt_template between these exact
-For each distinct file type found, decide:
+markers on their own lines:
-- 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)
+{template_start}
-Cover these topics based on what you discovered:
+<the complete extraction prompt here>
-1. How to map content to entity types (task / note / timeline entry)
+{template_end}
-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 AgentConfig JSON
+The prompt_template must be a self-contained instruction for an AI that reads files
-Once you are ≥ 90% confident, output the final config between these exact markers
+and must perform CRUD operations using tools to create records.  It should specify:
-(each on its own line):
+  - What entity types to create (tasks, notes, timelines) — never projects.
-
+  - How to map file content to record fields (camelCase: title, status, priority,
-{config_start}
+    dueDate, content, etc.) — never include projectId.
-{{
+  - That isAiSuggested must be set to 1 on every new record.
-  "content_types": [
+  - Concrete examples of mappings based on what you discovered in the directory.
-    {{
-      "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}\
+Keep asking clarifying questions until you are at least 90% confident you have
+enough information to generate an accurate prompt_template.  Once you reach that
+confidence level, stop asking and produce the final template immediately.
 Begin by exploring the directory, then ask your first question.\
 """
 
@@ -164,53 +145,33 @@ 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,
+    existing_template: str | None = None,
-) -> tuple[str, Any]:
+) -> str:
-    """Return ``(compiled_system_prompt, langfuse_prompt_obj_or_None)``."""
     existing_section = (
-        "\nThe user already has the following AgentConfig — refine it based on their answers:\n"
+        f"\nThe user already has the following prompt_template — refine it based on their answers:\n"
-        f"```json\n{existing_config}\n```\n"
+        f"---\n{existing_template}\n---\n"
-        if existing_config
+        if existing_template
         else ""
     )
-    template, prompt_obj = get_prompt_or_fallback(
+    return _SYSTEM_PROMPT_TEMPLATE.format(
-        "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),
+        template_start=_TEMPLATE_START,
-        config_start=_CONFIG_START,
+        template_end=_TEMPLATE_END,
-        config_end=_CONFIG_END,
         existing_section=existing_section,
     )
-    return compiled, prompt_obj
 
 
-# ── AgentConfig extraction ────────────────────────────────────────────────
+# ── Template extraction ───────────────────────────────────────────────────
 
 
-def _extract_agent_config(text: str) -> str | None:
+def _extract_template(text: str) -> str | None:
-    """Return validated AgentConfig JSON string from between markers, or None.
+    """Return the text between PROMPT_TEMPLATE_START and PROMPT_TEMPLATE_END, or None."""
-
+    if _TEMPLATE_START not in text or _TEMPLATE_END not in text:
-    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 = AgentConfig.model_validate_json(raw)
-        return parsed.model_dump_json()
-    except Exception as exc:
-        logger.warning("agent_setup: failed to parse AgentConfig JSON: %s", exc)
         return None
+    start_idx = text.index(_TEMPLATE_START) + len(_TEMPLATE_START)
+    end_idx = text.index(_TEMPLATE_END)
+    return text[start_idx:end_idx].strip() or None
 
 
 # ── LLM call with tool support ───────────────────────────────────────────
@@ -238,17 +199,12 @@ 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":
@@ -256,59 +212,16 @@ async def _call_llm_with_tools(
         else:
             messages.append(AIMessage(content=turn["content"]))
 
-    llm = get_agent_llm("setup", temperature=0.4)
+    llm = get_llm(model=None, 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)
+    for _ in range(_MAX_TOOL_STEPS):
-    _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:
+            return _as_text(response.content)
-                    _span.update(output=resp_text)
-                return resp_text
 
         for call in response.tool_calls:
             call_name = str(call.get("name", ""))
@@ -334,19 +247,7 @@ async def _call_llm_with_tools(
 
     # Fallback: exceeded max steps.
     final = await llm.ainvoke(messages)
-        final_text = _as_text(final.content)
+    return _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) ──────────────────────────
@@ -364,12 +265,12 @@ async def handle_journey_start(
     agent_type = frame.get("agent_type", "local")
     directory = frame.get("directory", "")
     data_types = frame.get("data_types", [])
-    existing_config = frame.get("existing_config")
+    existing_template = frame.get("existing_template")
 
     # 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)
+    system_prompt = _build_system_prompt(directory, data_types, existing_template)
 
     session = JourneySession(
         session_id=session_id,
@@ -378,21 +279,19 @@ async def handle_journey_start(
         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
+    # The LLM will explore the directory using FILESYSTEM_TOOLS via the
-    # user/input message to be present.
+    # ws_context executor (already set by the WS handler before calling us).
+    # Seed with an initial user message — some providers (e.g. GitHub Copilot)
+    # 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),
+        tools=list(FILESYSTEM_TOOLS),
-        user_id=user_id,
-        session_id=session_id,
-        langfuse_prompt=langfuse_prompt,
     )
 
     session.history.extend(seed_history)
@@ -406,14 +305,14 @@ async def handle_journey_start(
         directory,
     )
 
-    # Check if the LLM produced the config on the first turn (unlikely but possible).
+    # Check if the LLM produced the template on the first turn (unlikely but possible).
-    agent_config = _extract_agent_config(ai_reply)
+    prompt_template = _extract_template(ai_reply)
-    done = agent_config is not None
+    done = prompt_template is not None
 
     display_message = ai_reply
     if done:
         display_message = (
-            ai_reply[: ai_reply.index(_CONFIG_START)].strip()
+            ai_reply[: ai_reply.index(_TEMPLATE_START)].strip()
             or "Here is your agent configuration. You can save it or continue refining."
         )
         _sessions.pop(session_id, None)
@@ -423,7 +322,7 @@ async def handle_journey_start(
         "session_id": session_id,
         "message": display_message,
         "done": done,
-        "agent_config": agent_config,
+        "prompt_template": prompt_template,
     }
 
 
@@ -446,59 +345,53 @@ async def handle_journey_message(
             "session_id": session_id,
             "message": "Journey session not found or expired. Please start a new setup.",
             "done": True,
-            "agent_config": None,
+            "prompt_template": 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,
+        tools=list(FILESYSTEM_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.
+    # Check if the LLM produced the final template.
-    agent_config = _extract_agent_config(ai_reply)
+    prompt_template = _extract_template(ai_reply)
-    done = agent_config is not None
+    done = prompt_template is not None
 
-    # If the LLM didn't produce a config, nudge it once it hits the hard safety cap.
+    # If the LLM didn't produce a template, nudge it once it has asked enough
+    # questions (>= _MIN_TURNS_BEFORE_NUDGE) or 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"AgentConfig JSON now, wrapped in {_CONFIG_START} / {_CONFIG_END} markers.]"
+                f"prompt_template now, wrapped in {_TEMPLATE_START} / {_TEMPLATE_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,
+                tools=list(FILESYSTEM_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)
+            prompt_template = _extract_template(nudge_reply)
-            if agent_config is not None:
+            if prompt_template 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()
+            ai_reply[: ai_reply.index(_TEMPLATE_START)].strip()
-            if _CONFIG_START in ai_reply
+            if _TEMPLATE_START in ai_reply
             else "Here is your agent configuration. You can save it or continue refining."
         )
         _sessions.pop(session_id, None)
@@ -509,5 +402,5 @@ async def handle_journey_message(
         "session_id": session_id,
         "message": display_message,
         "done": done,
-        "agent_config": agent_config,
+        "prompt_template": prompt_template,
     }

app/api/routes/agents.py

@@ -12,11 +12,8 @@ in backend agent-config tables.
 from __future__ import annotations
 
 import asyncio
-import logging
 import uuid
-from datetime import datetime, timezone
+from datetime import datetime, timedelta, timezone
-
-logger = logging.getLogger(__name__)
 
 from fastapi import APIRouter, Depends, HTTPException, status
 from sqlalchemy import func, select
@@ -180,11 +177,6 @@ async def trigger_agent_run(
     _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 = LocalAgentConfig(
         id=str(uuid.uuid4()),
         user_id=current_user.id,
@@ -192,12 +184,10 @@ async def trigger_agent_run(
         name="Local Directory Monitor",
         directory_paths=[body.directory],
         data_types=_to_data_types(body.what_to_extract),
-        prompt_template=body.custom_agent_prompt or "",
+        prompt_template=body.custom_agent_prompt,
-        agent_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.

app/api/routes/auth.py

@@ -1,68 +1,34 @@
-"""Auth routes: register, login, refresh, me, OAuth social login, onboarding.
+"""Auth routes: register, login, refresh, me.
 
 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 pydantic import BaseModel
 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.models import 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 ─────────────────────────────────────────────────
 
 
@@ -265,531 +231,5 @@ async def update_profile(
         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/backup.py

@@ -0,0 +1,171 @@
+"""Backup routes: upload, download, history, and delete E2E-encrypted backups.
+
+Blobs are stored in S3 via BlobStore. Backup metadata is persisted in the
+PostgreSQL ``backup_metadata`` table.
+
+IMPORTANT: GET /history must be declared BEFORE GET / to avoid FastAPI
+treating "history" as a ``{backup_id}`` path parameter.
+"""
+
+from __future__ import annotations
+
+import uuid
+from email.utils import parsedate_to_datetime
+
+from fastapi import APIRouter, Depends, Header, HTTPException, Request, Response, status
+from sqlalchemy import func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.api.deps import get_current_user
+from app.billing.tier_manager import tier_manager
+from app.db import get_session
+from app.models import BackupMetadata as BackupMetadataModel
+from app.schemas import BackupMetadata, UserProfile
+from app.storage.blob_store import BlobStore
+from app.storage.encryption import reject_if_tampered
+
+router = APIRouter(prefix="/backup", tags=["backup"])
+
+_blob_store = BlobStore()
+
+
+async def _current_backup_bytes(user_id: str, db: AsyncSession) -> int:
+    """Return total backup bytes stored by *user_id*."""
+    result = await db.execute(
+        select(func.coalesce(func.sum(BackupMetadataModel.size_bytes), 0)).where(
+            BackupMetadataModel.user_id == user_id
+        )
+    )
+    return int(result.scalar_one())
+
+
+async def _check_backup_quota(
+    user: UserProfile, size_bytes: int, db: AsyncSession
+) -> None:
+    """Raise HTTP 402 if the upload would exceed the tier's backup limit."""
+    current = await _current_backup_bytes(user.id, db)
+    tier_manager.enforce_backup_quota(
+        user.tier, current_bytes=current, additional_bytes=size_bytes
+    )
+
+
+@router.put("")
+async def upload_backup(
+    request: Request,
+    x_backup_version: int = Header(..., alias="X-Backup-Version"),
+    x_backup_timestamp: int = Header(..., alias="X-Backup-Timestamp"),
+    x_backup_checksum: str = Header(..., alias="X-Backup-Checksum"),
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Upload an E2E-encrypted backup blob.
+
+    Metadata is passed via custom headers; the raw body is the encrypted blob.
+    """
+    blob = await request.body()
+    reject_if_tampered(blob, x_backup_checksum)
+    await _check_backup_quota(current_user, len(blob), db)
+
+    s3_key = await _blob_store.upload(
+        current_user.id, "backup", str(x_backup_timestamp), blob, x_backup_checksum
+    )
+
+    row = BackupMetadataModel(
+        id=str(uuid.uuid4()),
+        user_id=current_user.id,
+        s3_key=s3_key,
+        version=x_backup_version,
+        timestamp=x_backup_timestamp,
+        checksum=x_backup_checksum,
+        size_bytes=len(blob),
+    )
+    db.add(row)
+    await db.commit()
+
+    return {"ok": True}
+
+
+@router.get("/history", response_model=list[BackupMetadata])
+async def backup_history(
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> list[BackupMetadata]:
+    """Return backup metadata records for the authenticated user (no blob bytes)."""
+    result = await db.execute(
+        select(BackupMetadataModel)
+        .where(BackupMetadataModel.user_id == current_user.id)
+        .order_by(BackupMetadataModel.timestamp.desc())
+    )
+    rows = result.scalars().all()
+    return [
+        BackupMetadata(
+            version=r.version,
+            timestamp=r.timestamp,
+            checksum=r.checksum,
+            chunk_count=1,
+        )
+        for r in rows
+    ]
+
+
+@router.get("")
+async def download_backup(
+    request: Request,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> Response:
+    """Download the latest backup blob. Supports ``If-Modified-Since``."""
+    result = await db.execute(
+        select(BackupMetadataModel)
+        .where(BackupMetadataModel.user_id == current_user.id)
+        .order_by(BackupMetadataModel.timestamp.desc())
+        .limit(1)
+    )
+    latest = result.scalar_one_or_none()
+    if latest is None:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="No backup found")
+
+    ims_header = request.headers.get("If-Modified-Since")
+    if ims_header:
+        try:
+            ims_dt = parsedate_to_datetime(ims_header)
+            ims_ms = int(ims_dt.timestamp() * 1000)
+            if latest.timestamp <= ims_ms:
+                return Response(status_code=status.HTTP_304_NOT_MODIFIED)
+        except Exception:
+            pass  # malformed header — ignore and serve the blob
+
+    blob = await _blob_store.download(current_user.id, latest.s3_key)
+    return Response(
+        content=blob,
+        media_type="application/octet-stream",
+        headers={
+            "X-Backup-Version": str(latest.version),
+            "X-Backup-Timestamp": str(latest.timestamp),
+            "X-Checksum": latest.checksum,
+        },
+    )
+
+
+@router.delete("/{backup_id}", response_model=dict)
+async def delete_backup(
+    backup_id: str,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Delete a specific backup by ID."""
+    result = await db.execute(
+        select(BackupMetadataModel).where(
+            BackupMetadataModel.id == backup_id,
+            BackupMetadataModel.user_id == current_user.id,
+        )
+    )
+    target = result.scalar_one_or_none()
+    if target is None:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Backup not found")
+
+    await _blob_store.delete(current_user.id, target.s3_key)
+    await db.delete(target)
+    await db.commit()
+
+    return {"ok": True}

app/api/routes/billing.py

@@ -83,16 +83,3 @@ async def cancel_subscription(
     """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

app/api/routes/chat.py

@@ -1,4 +1,4 @@
-"""Chat routes: POST /chat (REST fallback) and POST /chat/embed (text → vector).
+"""Chat routes: POST /chat (REST fallback).
 
 WebSocket chat is handled by the unified device WS endpoint (/api/v1/ws/device).
 """
@@ -7,30 +7,14 @@ from __future__ import annotations
 
 from fastapi import APIRouter, Depends
 from fastapi.responses import JSONResponse
-from pydantic import BaseModel
 
 from app.api.deps import get_current_user
 from app.core.deep_agent import run_home
-from app.core.llm import embed
 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,
@@ -43,17 +27,3 @@ async def chat(
         context=body.context.model_dump(),
     )
     return JSONResponse(content={"response": response})
-
-
-@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/plugins.py

@@ -0,0 +1,148 @@
+"""Plugins routes: browse and install plugins from the marketplace.
+
+Backed by ``PluginRegistry`` and ``RevenueShare`` service classes that
+persist data in the PostgreSQL ``plugins`` and ``revenue_events`` tables.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Literal
+
+from fastapi import APIRouter, Depends, HTTPException, Query, status
+from pydantic import BaseModel
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.api.deps import get_current_user
+from app.db import get_session
+from app.marketplace.plugin_registry import registry
+from app.marketplace.revenue_share import revenue_share
+from app.models import PluginInstallation, PluginReview as PluginReviewModel
+from app.schemas import PluginInstallRequest, PluginListResponse, PluginManifest, UserProfile
+
+router = APIRouter(prefix="/plugins", tags=["plugins"])
+
+
+# ── Tier gate ─────────────────────────────────────────────────────────
+
+def _require_plugin_tier(user: UserProfile) -> None:
+    """Raise HTTP 403 for users below Power tier."""
+    if user.tier not in ("power", "team"):
+        raise HTTPException(
+            status_code=status.HTTP_403_FORBIDDEN,
+            detail="Plugin marketplace requires Power tier or above",
+        )
+
+
+# ── Local detail schema ────────────────────────────────────────────────
+
+class _PluginDetail(BaseModel):
+    plugin: PluginManifest
+    install_count: int
+    ratings: list[Any]
+
+
+# ── Routes ────────────────────────────────────────────────────────────
+
+@router.get("", response_model=PluginListResponse)
+async def list_plugins(
+    category: str | None = Query(default=None),
+    q: str | None = Query(default=None),
+    page: int = Query(default=1, ge=1),
+    sort: Literal["rating", "installs", "newest"] = Query(default="newest"),
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> PluginListResponse:
+    """Browse the plugin marketplace. Requires Power tier or above."""
+    _require_plugin_tier(current_user)
+    return await registry.list_plugins(db, category=category, query=q, page=page, sort=sort)
+
+
+@router.get("/{plugin_id}", response_model=_PluginDetail)
+async def get_plugin(
+    plugin_id: str,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> _PluginDetail:
+    """Get full plugin details including install count. Requires Power tier or above."""
+    _require_plugin_tier(current_user)
+    entry = await registry.get_plugin(db, plugin_id)
+    if entry is None:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Plugin not found")
+
+    # Fetch review ratings for this plugin
+    review_result = await db.execute(
+        select(PluginReviewModel).where(PluginReviewModel.plugin_id == plugin_id)
+    )
+    reviews = review_result.scalars().all()
+    ratings = [
+        {
+            "reviewer_id": r.reviewer_id,
+            "decision": r.decision,
+            "notes": r.notes,
+            "reviewed_at": int(r.reviewed_at.timestamp() * 1000) if r.reviewed_at else None,
+        }
+        for r in reviews
+    ]
+
+    return _PluginDetail(
+        plugin=entry["manifest"],
+        install_count=entry["install_count"],
+        ratings=ratings,
+    )
+
+
+@router.post("/{plugin_id}/install", response_model=dict)
+async def install_plugin(
+    plugin_id: str,
+    body: PluginInstallRequest,  # noqa: ARG001 — reserved for future fields
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, Any]:
+    """Install a plugin. Triggers Stripe Connect revenue split for paid plugins.
+
+    Requires Power tier or above.
+    """
+    _require_plugin_tier(current_user)
+    entry = await registry.get_plugin(db, plugin_id)
+    if entry is None:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Plugin not found")
+
+    # Record the installation in plugin_installations
+    installation = PluginInstallation(
+        plugin_id=plugin_id,
+        user_id=current_user.id,
+    )
+    db.add(installation)
+    await db.flush()
+
+    await revenue_share.record_install(
+        db,
+        plugin_id=plugin_id,
+        user_id=current_user.id,
+        amount_cents=entry["manifest"].price_cents,
+    )
+
+    download_url = f"https://cdn.adiuva.app/plugins/{plugin_id}/package.zip"
+    return {"ok": True, "download_url": download_url}
+
+
+@router.delete("/{plugin_id}/install", response_model=dict)
+async def uninstall_plugin(
+    plugin_id: str,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Unregister a plugin installation."""
+    result = await db.execute(
+        select(PluginInstallation).where(
+            PluginInstallation.plugin_id == plugin_id,
+            PluginInstallation.user_id == current_user.id,
+        )
+    )
+    installation = result.scalar_one_or_none()
+    if installation is not None:
+        await db.delete(installation)
+        await db.commit()
+    await registry.record_uninstall(db, plugin_id)
+    return {"ok": True}

app/api/routes/storage.py

@@ -0,0 +1,195 @@
+"""Storage routes: CRUD for E2E-encrypted cloud records.
+
+Blobs are stored in S3 via BlobStore. Record metadata is persisted in the
+PostgreSQL ``storage_records`` table.
+"""
+
+from __future__ import annotations
+
+import uuid
+
+from fastapi import APIRouter, Depends, HTTPException, Query, Response, status
+from pydantic import BaseModel
+from sqlalchemy import func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.api.deps import get_current_user
+from app.billing.tier_manager import tier_manager
+from app.db import get_session
+from app.models import StorageRecord
+from app.schemas import StorageRecordCreate, StorageRecordUpdate, UserProfile
+from app.storage.blob_store import BlobStore
+from app.storage.encryption import reject_if_tampered
+
+router = APIRouter(prefix="/storage", tags=["storage"])
+
+_blob_store = BlobStore()
+
+
+# ── Local response schemas ─────────────────────────────────────────────
+
+class _CreateResponse(BaseModel):
+    id: str
+    created_at: int
+
+
+class _RecordMeta(BaseModel):
+    id: str
+    table: str
+    checksum: str
+    created_at: int
+    updated_at: int
+
+
+# ── Helpers ────────────────────────────────────────────────────────────
+
+async def _current_usage_bytes(user_id: str, db: AsyncSession) -> int:
+    """Return total bytes stored by *user_id*."""
+    result = await db.execute(
+        select(func.coalesce(func.sum(StorageRecord.size_bytes), 0)).where(
+            StorageRecord.user_id == user_id
+        )
+    )
+    return int(result.scalar_one())
+
+
+async def _check_quota(user: UserProfile, additional_bytes: int, db: AsyncSession) -> None:
+    """Raise HTTP 402 if adding *additional_bytes* would exceed the tier limit."""
+    current = await _current_usage_bytes(user.id, db)
+    tier_manager.enforce_quota(user.tier, current_bytes=current, additional_bytes=additional_bytes)
+
+
+async def _get_record_for_user(
+    record_id: str, user_id: str, db: AsyncSession
+) -> StorageRecord:
+    """Look up a record and verify ownership. Returns 404 on mismatch
+    to prevent user enumeration attacks."""
+    result = await db.execute(
+        select(StorageRecord).where(
+            StorageRecord.id == record_id, StorageRecord.user_id == user_id
+        )
+    )
+    record = result.scalar_one_or_none()
+    if record is None:
+        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="Record not found")
+    return record
+
+
+# ── Routes ─────────────────────────────────────────────────────────────
+
+@router.post("/records", response_model=_CreateResponse, status_code=status.HTTP_201_CREATED)
+async def create_record(
+    body: StorageRecordCreate,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> _CreateResponse:
+    """Upload a new E2E-encrypted blob. Verifies checksum before storing."""
+    reject_if_tampered(body.blob, body.checksum)
+    await _check_quota(current_user, len(body.blob), db)
+
+    record_id = str(uuid.uuid4())
+
+    s3_key = await _blob_store.upload(
+        current_user.id, body.table, record_id, body.blob, body.checksum
+    )
+
+    record = StorageRecord(
+        id=record_id,
+        user_id=current_user.id,
+        table_name=body.table,
+        s3_key=s3_key,
+        checksum=body.checksum,
+        size_bytes=len(body.blob),
+    )
+    db.add(record)
+    await db.commit()
+    await db.refresh(record)
+
+    created_at_ms = int(record.created_at.timestamp() * 1000)
+    return _CreateResponse(id=record_id, created_at=created_at_ms)
+
+
+@router.get("/records", response_model=list[_RecordMeta])
+async def list_records(
+    table: str | None = Query(default=None),
+    page: int = Query(default=1, ge=1),
+    limit: int = Query(default=50, ge=1, le=200),
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> list[_RecordMeta]:
+    """List record metadata for the authenticated user. Blob bytes are never returned."""
+    query = select(StorageRecord).where(StorageRecord.user_id == current_user.id)
+    if table is not None:
+        query = query.where(StorageRecord.table_name == table)
+    query = query.offset((page - 1) * limit).limit(limit)
+
+    result = await db.execute(query)
+    rows = result.scalars().all()
+
+    return [
+        _RecordMeta(
+            id=r.id,
+            table=r.table_name,
+            checksum=r.checksum,
+            created_at=int(r.created_at.timestamp() * 1000),
+            updated_at=int(r.updated_at.timestamp() * 1000),
+        )
+        for r in rows
+    ]
+
+
+@router.get("/records/{record_id}")
+async def download_record(
+    record_id: str,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> Response:
+    """Download an E2E-encrypted blob. Returns raw bytes with ``X-Checksum`` header."""
+    record = await _get_record_for_user(record_id, current_user.id, db)
+    blob = await _blob_store.download(current_user.id, record.s3_key)
+    return Response(
+        content=blob,
+        media_type="application/octet-stream",
+        headers={"X-Checksum": record.checksum},
+    )
+
+
+@router.put("/records/{record_id}", response_model=dict)
+async def update_record(
+    record_id: str,
+    body: StorageRecordUpdate,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Replace the blob for an existing record. Verifies checksum before storing."""
+    record = await _get_record_for_user(record_id, current_user.id, db)
+    reject_if_tampered(body.blob, body.checksum)
+
+    delta = len(body.blob) - record.size_bytes
+    if delta > 0:
+        await _check_quota(current_user, delta, db)
+
+    s3_key = await _blob_store.upload(
+        current_user.id, record.table_name, record_id, body.blob, body.checksum
+    )
+
+    record.s3_key = s3_key
+    record.checksum = body.checksum
+    record.size_bytes = len(body.blob)
+    await db.commit()
+
+    return {"ok": True}
+
+
+@router.delete("/records/{record_id}", response_model=dict)
+async def delete_record(
+    record_id: str,
+    current_user: UserProfile = Depends(get_current_user),
+    db: AsyncSession = Depends(get_session),
+) -> dict[str, bool]:
+    """Delete a record and its S3 blob."""
+    record = await _get_record_for_user(record_id, current_user.id, db)
+    await _blob_store.delete(current_user.id, record.s3_key)
+    await db.delete(record)
+    await db.commit()
+    return {"ok": True}

app/api/routes/vectors.py

@@ -0,0 +1,79 @@
+"""Vectors routes: upsert, search, delete cloud vector store entries, and embed text."""
+
+from __future__ import annotations
+
+from fastapi import APIRouter, Depends
+from pydantic import BaseModel
+
+from app.api.deps import get_current_user
+from app.core.llm import embed
+from app.schemas import (
+    UserProfile,
+    VectorSearchRequest,
+    VectorSearchResponse,
+    VectorUpsertRequest,
+)
+from app.storage.encryption import reject_if_tampered
+from app.storage.vector_store import VectorStore
+
+router = APIRouter(prefix="/storage", tags=["vectors"])
+
+_vector_store = VectorStore()
+
+
+class _VectorDeleteRequest(BaseModel):
+    ids: list[str]
+
+
+class _EmbedRequest(BaseModel):
+    text: str
+
+
+class _EmbedResponse(BaseModel):
+    vector: list[float]
+
+
+@router.post("/vectors/upsert", response_model=dict)
+async def upsert_vectors(
+    body: VectorUpsertRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> dict[str, int]:
+    """Verify checksums and store encrypted vectors in the user-scoped namespace."""
+    for item in body.vectors:
+        reject_if_tampered(item.blob, item.checksum)
+    await _vector_store.upsert(current_user.id, body.vectors)
+    return {"upserted": len(body.vectors)}
+
+
+@router.post("/vectors/search", response_model=VectorSearchResponse)
+async def search_vectors(
+    body: VectorSearchRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> VectorSearchResponse:
+    """Search the user-scoped vector namespace with an encrypted query blob."""
+    results = await _vector_store.search(current_user.id, body.query_blob, body.top_k)
+    return VectorSearchResponse(results=results)
+
+
+@router.delete("/vectors", response_model=dict)
+async def delete_vectors(
+    body: _VectorDeleteRequest,
+    current_user: UserProfile = Depends(get_current_user),
+) -> dict[str, bool]:
+    """Delete vectors by ID, scoped to the authenticated user."""
+    await _vector_store.delete(current_user.id, body.ids)
+    return {"ok": True}
+
+
+@router.post("/vectors/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 backend tools (note_agent) and Electron (vectordb.ts) alike.
+    """
+    vector = await embed(body.text)
+    return _EmbedResponse(vector=vector)

app/auth/__init__.py

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

app/auth/oauth_providers.py

@@ -1,135 +0,0 @@
-"""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/stripe_service.py

@@ -43,8 +43,8 @@ class StripeService:
         self,
         user_id: str,
         tier: str,
-        success_url: str = "https://app.adiuvai.app/billing/success?session_id={CHECKOUT_SESSION_ID}",
+        success_url: str = "https://app.adiuva.app/billing/success?session_id={CHECKOUT_SESSION_ID}",
-        cancel_url: str = "https://app.adiuvai.app/billing/cancel",
+        cancel_url: str = "https://app.adiuva.app/billing/cancel",
     ) -> str:
         """Create a Stripe checkout session and return the URL.
 
@@ -200,45 +200,6 @@ 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(

app/billing/tier_manager.py

@@ -22,32 +22,44 @@ FEATURES: dict[str, dict[str, Any]] = {
         "agents": 3,
         "batch_active": 2,
         "batch_runs_per_day": 5,
+        "cloud_storage_gb": 0,
+        "backup_gb": 0,
         "providers": 1,
         "batch_builder": False,
+        "plugin_marketplace": False,
         "sso": False,
     },
     "pro": {
         "agents": -1,           # unlimited
         "batch_active": 10,
         "batch_runs_per_day": 50,
+        "cloud_storage_gb": 5,
+        "backup_gb": 5,
         "providers": -1,
         "batch_builder": False,
+        "plugin_marketplace": False,
         "sso": False,
     },
     "power": {
         "agents": -1,
         "batch_active": -1,     # unlimited
         "batch_runs_per_day": -1,  # unlimited
+        "cloud_storage_gb": 25,
+        "backup_gb": 25,
         "providers": -1,
         "batch_builder": True,
+        "plugin_marketplace": True,
         "sso": False,
     },
     "team": {
         "agents": -1,
         "batch_active": -1,
         "batch_runs_per_day": -1,  # unlimited
+        "cloud_storage_gb": -1,  # unlimited
+        "backup_gb": -1,         # unlimited
         "providers": -1,
         "batch_builder": True,
+        "plugin_marketplace": True,
         "sso": True,
     },
 }
@@ -113,6 +125,71 @@ class TierManager:
         """Return the requests-per-minute limit for ``tier``."""
         return RATE_LIMITS.get(tier, RATE_LIMITS["free"])
 
+    # ── Storage quota ────────────────────────────────────────────────────
+
+    def enforce_quota(
+        self,
+        tier: BillingTier,
+        current_bytes: int = 0,
+        additional_bytes: int = 0,
+    ) -> None:
+        """Raise ``HTTP 402`` if the user would exceed their cloud storage quota.
+
+        ``tier`` is the caller's current tier (from ``current_user.tier``).
+        ``current_bytes`` is the total bytes already stored (queried by caller).
+        """
+        limit_gb: int = FEATURES[tier]["cloud_storage_gb"]
+        if limit_gb == 0:
+            raise HTTPException(
+                status_code=status.HTTP_402_PAYMENT_REQUIRED,
+                detail=f"Cloud storage is not available on the '{tier}' tier",
+            )
+        if limit_gb == -1:
+            return  # unlimited
+        limit_bytes = limit_gb * 1024 ** 3
+        if current_bytes + additional_bytes > limit_bytes:
+            raise HTTPException(
+                status_code=status.HTTP_402_PAYMENT_REQUIRED,
+                detail=f"Storage quota exceeded for tier '{tier}'",
+            )
+
+    def enforce_backup_quota(
+        self,
+        tier: BillingTier,
+        current_bytes: int = 0,
+        additional_bytes: int = 0,
+    ) -> None:
+        """Raise ``HTTP 402`` if the user would exceed their backup quota."""
+        limit_gb: int = FEATURES[tier]["backup_gb"]
+        if limit_gb == 0:
+            raise HTTPException(
+                status_code=status.HTTP_402_PAYMENT_REQUIRED,
+                detail=f"Backup is not available on the '{tier}' tier",
+            )
+        if limit_gb == -1:
+            return  # unlimited
+        limit_bytes = limit_gb * 1024 ** 3
+        if current_bytes + additional_bytes > limit_bytes:
+            raise HTTPException(
+                status_code=status.HTTP_402_PAYMENT_REQUIRED,
+                detail=f"Backup quota exceeded for tier '{tier}'",
+            )
+
+    def check_quota(
+        self,
+        tier: BillingTier,
+        current_bytes: int = 0,
+        additional_bytes: int = 0,
+    ) -> bool:
+        """Return ``True`` if the user can store ``additional_bytes`` more data."""
+        limit_gb: int = FEATURES[tier]["cloud_storage_gb"]
+        if limit_gb == 0:
+            return False
+        if limit_gb == -1:
+            return True
+        limit_bytes = limit_gb * 1024 ** 3
+        return current_bytes + additional_bytes <= limit_bytes
+
 
 # Module-level singleton shared across the app.
 tier_manager = TierManager()

app/config/settings.py

@@ -3,7 +3,7 @@ from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
 class Settings(BaseSettings):
-    DATABASE_URL: str = "postgresql+asyncpg://postgres:postgres@localhost:5432/adiuvai"
+    DATABASE_URL: str = "postgresql+asyncpg://postgres:postgres@localhost:5432/adiuva"
     JWT_SECRET: str = "change-me-in-production"
     JWT_ALGORITHM: str = "HS256"
     JWT_ACCESS_TOKEN_EXPIRE_MINUTES: int = 30
@@ -12,22 +12,26 @@ class Settings(BaseSettings):
     STRIPE_SECRET_KEY: str = ""
     STRIPE_WEBHOOK_SECRET: str = ""
 
+    S3_BUCKET: str = ""
+    S3_REGION: str = "us-east-1"
+    S3_ENDPOINT_URL: str = ""
+    AWS_ACCESS_KEY_ID: str = ""
+    AWS_SECRET_ACCESS_KEY: str = ""
+
+    PINECONE_API_KEY: str = ""
+    PINECONE_INDEX: str = "adiuva"
+    QDRANT_URL: str = ""
+    QDRANT_API_KEY: str = ""
+
     OPENAI_API_KEY: str = ""
     ANTHROPIC_API_KEY: str = ""
     GOOGLE_API_KEY: str = ""
     CEREBRAS_API_KEY: str = ""
+    GITHUB_TOKEN: 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 = ""        # _infer_floating_domain (intent routing)
-    LLM_MODEL_HOME_AGENT: str = ""        # home-agent (run_single_agent / stream)
-    LLM_MODEL_FLOATING_AGENT: str = ""    # floating-agent (contextual chat)
-    LLM_MODEL_UNIFIED_PROCESSOR: str = "" # unified-processor (agent_runner)
-    LLM_MODEL_CLOUD_PROCESSOR: str = ""   # cloud-processor (agent_runner)
-    LLM_MODEL_SETUP_AGENT: str = ""       # agent-setup journey
-
     # 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.
@@ -41,37 +45,18 @@ class Settings(BaseSettings):
     # 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"
-
     # 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] = [
+    CORS_ORIGINS: list[str] = ["app://.", "http://localhost:3000", "http://localhost:5173"]
-        "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"
 
     ENV: Literal["dev", "prod"] = "dev"
 
-    model_config = SettingsConfigDict(env_file=".env", env_file_encoding="utf-8")
+    model_config = SettingsConfigDict(
+        env_file=".env", env_file_encoding="utf-8", extra="ignore"
+    )
 
 
 settings = Settings()

app/core/agent_runner.py

@@ -2,12 +2,12 @@
 
 Drives two agent types:
 
-* **Local directory agent** — V2 unified flow per file:
+* **Local directory agent** — two-step execution per file:
-  Phase A (Detect + Preprocess, zero LLM): Python detects the content type
+  Step 1 (Classification) uses code to fetch all projects and asks the LLM
-  and strips markup/noise, producing clean text + metadata.
+  to identify which project the file belongs to and which domains are relevant.
-  Phase B (Single LLM call with tools): the LLM identifies the project,
+  Step 2 (Processing) fetches existing entities for that project/domains via
-  checks for duplicates via list_* tools, and creates/updates records.
+  code and runs an LLM with tools — existing data in context enforces
-  ``items_created`` is counted from ``create_*`` tool calls.
+  update-first naturally.
 
 * **Cloud connector agent** — fetches data from third-party APIs (Gmail,
   Teams, Outlook) and pushes extracted items to Electron.
@@ -29,7 +29,7 @@ from __future__ import annotations
 import asyncio
 import json
 import logging
-import os
+import uuid
 from datetime import datetime, timedelta, timezone
 from typing import Any
 
@@ -43,9 +43,7 @@ from app.agents.project_agent import PROJECT_TOOLS
 from app.agents.task_agent import TASK_TOOLS
 from app.agents.timeline_agent import TIMELINE_TOOLS
 from app.core.device_manager import DeviceConnectionManager
-from app.core.langfuse_client import compile_prompt, extract_usage, get_langfuse, get_prompt_or_fallback, langfuse_context
+from app.core.llm import get_llm
-from app.core.llm import get_agent_llm, model_for_agent
-from app.core.preprocessors import detect_content_type, preprocess
 from app.core.ws_context import clear_client_executor, execute_on_client, set_client_executor
 from app.db import async_session
 from app.models import AgentRunLog, CloudAgentConfig, LocalAgentConfig
@@ -72,52 +70,97 @@ _MAX_PROCESSING_STEPS: int = 12
 _MAX_SCAN_DEPTH: int = 5
 
 # ── Data-type to tool mapping ─────────────────────────────────────────────
+# NOTE: "projects" is intentionally excluded — project creation/assignment is
+# handled in code by the runner, never delegated to the Step 2 LLM.
 
 _DATA_TYPE_TOOLS: dict[str, list[Any]] = {
     "tasks": TASK_TOOLS,
     "notes": NOTE_TOOLS,
     "timelines": TIMELINE_TOOLS,
-    "timelineEvents": TIMELINE_TOOLS,
-    "projects": PROJECT_TOOLS,
 }
 
-# ── V2: Unified processing prompt (hot-swappable via Langfuse "unified_processing") ──
+# ── Step 1: Classification prompt ─────────────────────────────────────────
 
-_UNIFIED_PROCESSING_PROMPT = """\
+_DOMAIN_DESCRIPTIONS: dict[str, str] = {
+    "tasks": (
+        "Action items, to-dos, deliverables — anything that describes work to be done, "
+        "assigned to someone, or tracked with a due date or status."
+    ),
+    "notes": (
+        "Documentation, meeting notes, summaries, reference material — "
+        "written content meant to be read and referenced rather than acted on."
+    ),
+    "timelines": (
+        "Project milestones, deadlines, scheduled events — "
+        "specific dates that mark a point in the progress of a project."
+    ),
+    "projects": (
+        "High-level project entities — only relevant if the file clearly introduces "
+        "a new project or updates the scope of an existing one."
+    ),
+}
+
+_STEP1_SYSTEM_PROMPT = """\
+You are a file classifier for a freelance project management tool.
+
+Your job is to match a file to an existing project and identify which data domains to extract.
+
+## Project matching rules (STRICT — follow in order)
+
+1. Search the file content for any mention of a project name, client name, acronym, or topic
+   that overlaps with the existing projects listed below.
+2. The match does NOT need to be exact — partial name, abbreviation, or topic similarity is enough.
+3. STRONGLY PREFER matching an existing project. Only return "new" as an absolute last resort
+   when the file has zero meaningful connection to any listed project.
+4. When in doubt, pick the closest match from the list.
+
+## Response format
+
+Respond ONLY with a JSON object — no markdown, no explanation:
+
+{{"project_id": "<exact id from the list below, or new>", "new_project_name": "<concise 2-5 word name, only when project_id is new>", "domains": ["tasks", "notes"]}}
+
+## Domain definitions (only consider domains in the allowed list)
+
+{domain_definitions}
+
+## Existing projects
+
+{projects_list}
+"""
+
+# ── Step 2: Processing prompt ─────────────────────────────────────────────
+
+_PROCESSING_SYSTEM_PROMPT = """\
 You are a data extraction assistant for a freelance project management tool.
 
-## Your process (follow this exact order)
+Your task: extract structured data from the file content and persist it using the available tools.
 
-### 1. Identify the project
+## Mandatory process — follow this order for EVERY item you extract
-File: {filename}
-{metadata_section}
 
-Existing projects:
+1. READ the existing records listed below for the relevant domain.
-{projects_list}
+2. SEARCH for a match by title, topic, or semantic similarity.
+3. If a match exists → call the update_* tool with the existing record's id.
+4. If no match exists → call the create_* tool and set isAiSuggested=1.
 
-Match this file to an existing project using the filename and content clues.
+NEVER call create_* without first checking the existing records.
-If no project matches, {no_match_behavior}.
+NEVER duplicate a record that already exists under a different wording.
 
-### 2. Check existing records
+## Existing records (source of truth)
-Once you identify the project, use list_tasks / list_notes / list_timelines
-(filtered by projectId) to see what already exists.
-NEVER create a record that already exists under the same or similar title.
 
-### 3. Extract and create / update
+{existing_context}
-{extraction_rules}
 
-### Rules
+## Context
-- Set isAiSuggested=1 on every new record.
+
-- Set projectId on every record (use the id from the project list above).
+Project: {project_context}
-- Update existing records when a match is found by title or topic.
+Domains to extract: {data_types}
-- Do NOT invent data — only extract what is clearly stated in the content.
+
-- Target entity types: {data_types}.
+{custom_prompt_section}
-{global_rules}
 """
 
 # ── Cloud processing prompt (kept separate for cloud agent) ───────────────
 
-_BATCH_CLOUD_PROCESSING_PROMPT = """\
+_CLOUD_PROCESSING_PROMPT = """\
 You are a data extraction and management assistant for a freelance project
 management tool.
 
@@ -225,19 +268,9 @@ async def _run_agent_with_tools(
     user_message: str,
     tools: list[Any],
     max_steps: int,
-    user_id: str = "",
-    session_id: str = "",
-    langfuse_prompt: Any = None,
-    agent_name: str = "batch-agent",
-    _tool_calls_out: list[str] | None = None,
 ) -> str:
-    """Run an LLM agent with tool-calling, returning the final text response.
+    """Run an LLM agent with tool-calling, returning the final text response."""
-
+    llm = get_llm()
-    If *_tool_calls_out* is provided, the name of every tool called during the
-    run is appended to it (used by the caller to count ``create_*`` calls).
-    """
-    lf = get_langfuse()
-    llm = get_agent_llm(agent_name)
     llm_with_tools = llm.bind_tools(tools)
     messages: list[Any] = [
         SystemMessage(content=system_prompt),
@@ -246,45 +279,12 @@ async def _run_agent_with_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=agent_name,
-            metadata={"user_id": user_id} if user_id else None,
-            input=user_message,
-        )
-        if lf else None
-    )
-    _span = _span_ctx.__enter__() if _span_ctx else None
-
-    try:
     for _ in range(max_steps):
-            _gen_ctx = (
-                lf.start_as_current_observation(
-                    as_type="generation",
-                    name=f"{agent_name}-llm",
-                    model=model_for_agent(agent_name),
-                    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)
-
         messages.append(response)
 
         if not response.tool_calls:
-                final_text = _as_text(response.content)
+            return _as_text(response.content)
-                if _span:
-                    _span.update(output=final_text)
-                return final_text
 
         for call in response.tool_calls:
             call_id = str(call.get("id", ""))
@@ -296,9 +296,6 @@ async def _run_agent_with_tools(
                 json.dumps(call_args, ensure_ascii=True)[:800],
             )
 
-                if _tool_calls_out is not None:
-                    _tool_calls_out.append(call_name)
-
             tool_fn = tool_map.get(call_name)
             if tool_fn is None:
                 tool_output = f"Unknown tool: {call_name}"
@@ -313,16 +310,7 @@ async def _run_agent_with_tools(
             messages.append(ToolMessage(content=str(tool_output), tool_call_id=call["id"]))
 
     final = await llm.ainvoke(messages)
-        final_text = _as_text(final.content)
+    return _as_text(final.content)
-        if _span:
-            _span.update(output=final_text)
-        return final_text
-    finally:
-        if _span_ctx:
-            _span_ctx.__exit__(None, None, None)
-        _lf_ctx.__exit__(None, None, None)
-        if lf:
-            lf.flush()
 
 
 # ── Tool list builder ─────────────────────────────────────────────────────
@@ -389,8 +377,7 @@ async def _scan_directories(
     for file_path in all_files:
         try:
             meta = await execute_on_client(action="get_file_metadata", data={"path": file_path})
-            # FE sends snake_case keys on the wire (toSnakeCase transform)
+            modified_at = meta.get("modifiedAt")
-            modified_at = meta.get("modified_at") or meta.get("modifiedAt")
             if modified_at is None:
                 filtered.append(file_path)
                 continue
@@ -492,66 +479,83 @@ def _format_entities_for_context(domain: str, rows: list[dict]) -> str:
     return f"Existing {domain}:\n" + "\n".join(lines)
 
 
-# ── V2 helper functions ───────────────────────────────────────────────────
+# ── Step 1: LLM file classifier ───────────────────────────────────────────
 
 
-def _format_projects(projects: list[dict]) -> str:
+async def _classify_file(
-    """Format the project list for the unified system prompt."""
+    file_path: str,
-    if not projects:
+    file_content: str,
-        return "  (no projects yet)"
+    projects: list[dict],
-    lines: list[str] = []
+    config_data_types: list[str],
-    for p in projects:
+) -> tuple[str, list[str], str | None]:
+    """Call the LLM to classify a file by project and relevant domains.
+
+    Returns ``(project_id_or_"new", domains, new_project_name_or_None)``.
+    - ``project_id`` is an existing project UUID, or ``"new"`` when no match found.
+    - ``new_project_name`` is only set when ``project_id == "new"``.
+    Falls back to ``("new", config_data_types, None)`` on any error.
+    """
+    fallback: tuple[str, list[str], str | None] = ("new", list(config_data_types), None)
+
+    if not file_content.strip():
+        return fallback
+
+    valid_project_ids = {p["id"] for p in projects}
+
+    def _fmt_project(p: dict) -> str:
         summary = (p.get("aiSummary") or p.get("ai_summary") or "").strip()
         summary_part = f" — {summary[:100]}" if summary else ""
-        lines.append(
+        return f"  - id={p['id']} | name={p.get('name', '')} | status={p.get('status', '')}{summary_part}"
-            f"  - id={p['id']} | name={p.get('name', '')} | "
-            f"status={p.get('status', '')}{summary_part}"
-        )
-    return "\n".join(lines)
 
+    projects_list = "\n".join(_fmt_project(p) for p in projects) or "  (none yet)"
 
-def _format_metadata(metadata: dict) -> str:
+    domain_definitions = "\n".join(
-    """Format preprocessor metadata as a compact context block."""
+        f"  - {d}: {_DOMAIN_DESCRIPTIONS[d]}"
-    if not metadata:
+        for d in config_data_types
-        return ""
+        if d in _DOMAIN_DESCRIPTIONS
-    parts: list[str] = []
-    for key in ("subject", "from", "to", "date"):
-        if metadata.get(key):
-            parts.append(f"{key.capitalize()}: {metadata[key]}")
-    # any remaining keys
-    for key, val in metadata.items():
-        if key not in ("subject", "from", "to", "date") and val:
-            parts.append(f"{key}: {val}")
-    return "\n".join(parts)
-
-
-def _get_extraction_rules(agent_config: dict, content_type: str) -> str:
-    """Return the extraction_prompt for *content_type* from *agent_config*.
-
-    Falls back to a generic instruction when the type is not configured.
-    """
-    for ct in agent_config.get("content_types", []):
-        if ct.get("id") == content_type:
-            prompt = ct.get("extraction_prompt", "").strip()
-            if prompt:
-                return prompt
-    return (
-        "Extract relevant information as tasks (action items), notes "
-        "(informational content), or timelines (dated events)."
     )
 
+    system = _STEP1_SYSTEM_PROMPT.format(
+        domain_definitions=domain_definitions,
+        projects_list=projects_list,
+    )
 
-def _get_no_match_behavior(agent_config: dict) -> str:
+    llm = get_llm()
-    """Derive the 'no project match' instruction from global_rules."""
+    try:
-    rules = agent_config.get("global_rules", [])
+        response = await llm.ainvoke([
-    for rule in rules:
+            SystemMessage(content=system),
-        lower = rule.lower()
+            HumanMessage(content=f"File: {file_path}\n\nContent:\n{file_content[:4000]}"),
-        if "no project" in lower or "no match" in lower or "skip" in lower:
+        ])
-            return rule
+        raw = _as_text(response.content).strip()
-    return "create a new project with a concise name derived from the file content"
+        # Strip markdown fences if the model wraps the JSON.
+        if raw.startswith("```"):
+            raw = raw.split("```")[1]
+            if raw.startswith("json"):
+                raw = raw[4:]
+        parsed = json.loads(raw.strip())
+        raw_project_id: str = str(parsed.get("project_id") or "new")
+        # Reject hallucinated UUIDs — only accept ids that exist in the fetched list.
+        project_id = raw_project_id if raw_project_id in valid_project_ids else "new"
+        new_project_name: str | None = (
+            str(parsed["new_project_name"]).strip() or None
+            if project_id == "new" and parsed.get("new_project_name")
+            else None
+        )
+        domains: list[str] = [
+            d for d in parsed.get("domains", [])
+            if d in config_data_types
+        ]
+        if not domains:
+            domains = list(config_data_types)
+        return project_id, domains, new_project_name
+    except Exception as exc:
+        logger.warning(
+            "agent_runner: step1 classification failed for %r: %s", file_path, exc
+        )
+        return fallback
 
 
-# ── Local agent runner (V2 — unified per-file flow) ───────────────────────
+# ── Local agent runner (two-step per file) ────────────────────────────────
 
 
 async def run_local_agent(
@@ -561,17 +565,16 @@ async def run_local_agent(
     device_mgr: DeviceConnectionManager,
     run_context: dict | None = None,
 ) -> None:
-    """Execute a local directory agent run — V2 unified flow.
+    """Execute a local directory agent run using a two-step approach per file.
 
-    Phase A — Detect + Preprocess (zero LLM, per file):
+    Step 1 — Classification (code + 1 LLM call per file, no tools):
-        Python detects the content type from filename + content patterns and
+        Code scans directories and fetches all projects via WS.
-        runs the appropriate handler (e.g. email_html) to produce clean text
+        For each file, LLM identifies the project and relevant domains.
-        and structured metadata.
 
-    Phase B — Single LLM call with tools (per file):
+    Step 2 — Processing (code + 1 LLM call per file, with tools):
-        One LLM call handles project identification, duplicate checking, and
+        Code fetches existing entities for the identified project/domains.
-        record creation/update.  ``create_*`` tool calls are counted to
+        LLM receives file content + existing entities in context and uses
-        produce the accurate ``items_created`` metric.
+        tools to update existing records or create new ones.
     """
     run_id = run_log.id
     agent_id = (run_context or {}).get("agent_id") or config.id
@@ -606,11 +609,16 @@ async def run_local_agent(
     errors: list[str] = []
     items_processed = 0
     items_created = 0
-    agent_config: dict = config.agent_config or {}
+
-    processing_tools = _build_processing_tools(config.data_types)
+    custom_section = (
+        f"User instructions:\n{config.prompt_template}"
+        if config.prompt_template
+        else ""
+    )
 
     try:
         # ── Code: scan directories ───────────────────────────────────
+        logger.info("agent_runner: run=%s scanning directories user=%s", run_id, user_id)
         file_paths = await _scan_directories(
             paths=config.directory_paths,
             extensions=config.file_extensions or [],
@@ -626,89 +634,108 @@ async def run_local_agent(
 
         # ── Code: fetch all projects once ────────────────────────────
         projects = await _fetch_projects()
-        projects_block = _format_projects(projects)
-
-        # Prompt template + Langfuse version linking (hot-swappable from UI).
-        unified_template, prompt_obj = get_prompt_or_fallback(
-            "unified_processing", _UNIFIED_PROCESSING_PROMPT
-        )
 
         for file_path in file_paths:
             try:
-                # ── Phase A: read + detect + preprocess ─────────────
+                # Read file content via code.
                 file_result = await execute_on_client(
                     action="read_file_content", data={"path": file_path}
                 )
-                raw_content: str = file_result.get("content", "")
+                file_content: str = file_result.get("content", "")
-                if not raw_content.strip():
+                if not file_content:
-                    logger.debug(
+                    logger.debug("agent_runner: run=%s skipping empty file %r", run_id, file_path)
-                        "agent_runner: run=%s skipping empty file %r", run_id, file_path
-                    )
                     continue
 
                 items_processed += 1
-                filename = os.path.basename(file_path)
-                content_type = detect_content_type(filename, raw_content)
-                preprocessed = preprocess(content_type, raw_content)
 
+                # Step 1 — classify file.
+                project_id, domains, new_project_name = await _classify_file(
+                    file_path=file_path,
+                    file_content=file_content,
+                    projects=projects,
+                    config_data_types=config.data_types,
+                )
                 logger.info(
-                    "agent_runner: run=%s file=%r content_type=%s clean_len=%d",
+                    "agent_runner: run=%s file=%r → project=%s new_name=%r domains=%s",
-                    run_id, file_path, content_type, len(preprocessed.clean_text),
+                    run_id,
+                    file_path,
+                    project_id,
+                    new_project_name,
+                    domains,
                 )
 
-                # ── Phase B: single LLM call ─────────────────────────
+                # Step 2 — resolve project_id via CODE, then fetch entities.
-                extraction_rules = _get_extraction_rules(agent_config, content_type)
+                # Project creation is NEVER delegated to the Step 2 LLM.
-                no_match_behavior = _get_no_match_behavior(agent_config)
+                if project_id == "new":
-                global_rules_lines = "\n".join(
+                    proj_name = new_project_name or "Untitled Project"
-                    f"- {r}" for r in agent_config.get("global_rules", [])
+                    try:
+                        proj_result = await execute_on_client(
+                            action="insert",
+                            table="projects",
+                            data={"name": proj_name, "clientId": None},
                         )
-                metadata_section = _format_metadata(preprocessed.metadata)
+                        created = proj_result.get("row", {})
-
+                        effective_project_id = created.get("id", "standalone")
-                system_prompt = compile_prompt(
+                        # Add to local list so subsequent files can match it.
-                    unified_template,
+                        if "id" in created:
-                    prompt_obj,
+                            projects.append(created)
-                    filename=filename,
+                        logger.info(
-                    metadata_section=metadata_section,
+                            "agent_runner: run=%s created project %r id=%s",
-                    projects_list=projects_block,
+                            run_id, proj_name, effective_project_id,
-                    no_match_behavior=no_match_behavior,
+                        )
-                    extraction_rules=extraction_rules,
+                    except Exception as exc:
-                    global_rules=global_rules_lines,
+                        logger.warning(
-                    data_types=", ".join(config.data_types),
+                            "agent_runner: run=%s failed to create project %r: %s",
+                            run_id, proj_name, exc,
+                        )
+                        effective_project_id = "standalone"
+                        proj_name = "unknown"
+                    project_context = (
+                        f"Project: {proj_name} (id: {effective_project_id}). "
+                        "Always set projectId to this id on every record you create."
+                    )
+                else:
+                    effective_project_id = project_id
+                    proj = next((p for p in projects if p["id"] == project_id), None)
+                    proj_name = proj.get("name", project_id) if proj else project_id
+                    project_context = (
+                        f"Project: {proj_name} (id: {project_id}). "
+                        "Always set projectId to this id on every record you create."
                     )
 
-                user_message = (
+                # "projects" domain is never passed to Step 2 — handled above in code.
-                    f"Process this file and extract relevant information.\n\n"
+                domains = [d for d in domains if d != "projects"]
-                    f"File: {file_path}\n\n"
+
-                    f"Content:\n{preprocessed.clean_text}"
+                existing_blocks: list[str] = []
+                for domain in domains:
+                    rows = await _fetch_domain_entities(domain, effective_project_id)
+                    existing_blocks.append(_format_entities_for_context(domain, rows))
+
+                existing_context = "\n\n".join(existing_blocks)
+
+                system_prompt = _PROCESSING_SYSTEM_PROMPT.format(
+                    existing_context=existing_context,
+                    project_context=project_context,
+                    data_types=", ".join(domains),
+                    custom_prompt_section=custom_section,
                 )
 
-                file_tool_calls: list[str] = []
+                processing_tools = _build_processing_tools(domains)
+
                 result_text = await _run_agent_with_tools(
                     system_prompt=system_prompt,
-                    user_message=user_message,
+                    user_message=(
+                        f"Process this file and extract relevant information.\n\n"
+                        f"File: {file_path}\n\nContent:\n{file_content}"
+                    ),
                     tools=processing_tools,
                     max_steps=_MAX_PROCESSING_STEPS,
-                    user_id=user_id,
-                    session_id=run_id,
-                    langfuse_prompt=prompt_obj,
-                    agent_name="unified-processor",
-                    _tool_calls_out=file_tool_calls,
                 )
-
-                file_created = sum(
-                    1 for name in file_tool_calls if name.startswith("create_")
-                )
-                items_created += file_created
-
-                # Refresh project list when a project was created so
-                # subsequent files see it in the prompt context.
-                if "create_project" in file_tool_calls:
-                    projects = await _fetch_projects()
-                    projects_block = _format_projects(projects)
-
                 logger.info(
-                    "agent_runner: run=%s file=%r created=%d result=%s",
+                    "agent_runner: run=%s file=%r result=%s",
-                    run_id, file_path, file_created, result_text[:200],
+                    run_id,
+                    file_path,
+                    result_text[:200],
                 )
 
             except Exception as exc:
@@ -740,11 +767,10 @@ async def run_local_agent(
         errors=errors,
     )
     logger.info(
-        "agent_runner: run=%s done status=%s processed=%d created=%d errors=%d",
+        "agent_runner: run=%s done status=%s processed=%d errors=%d",
         run_id,
         final_status,
         items_processed,
-        items_created,
         len(errors),
     )
 
@@ -902,12 +928,7 @@ async def run_cloud_agent(
                 continue
             items_processed += 1
 
-            cloud_template, cloud_prompt_obj = get_prompt_or_fallback(
+            processing_prompt = _CLOUD_PROCESSING_PROMPT.format(
-                "batch_cloud_processing", _BATCH_CLOUD_PROCESSING_PROMPT
-            )
-            processing_prompt = compile_prompt(
-                cloud_template,
-                cloud_prompt_obj,
                 data_types=", ".join(config.data_types),
                 project_context="Determine the appropriate project from the message context.",
                 file_list=f"Message from {config.provider} (id: {msg.id})",
@@ -920,10 +941,6 @@ async def run_cloud_agent(
                     user_message=f"Process this message content:\n\n{content_text[:8000]}",
                     tools=processing_tools,
                     max_steps=_MAX_PROCESSING_STEPS,
-                    user_id=user_id,
-                    session_id=run_id,
-                    langfuse_prompt=cloud_prompt_obj,
-                    agent_name="cloud-processor",
                 )
             except Exception as exc:
                 errors.append(f"LLM processing error for message {msg.id!r}: {exc}")

app/core/deep_agent.py

@@ -16,8 +16,7 @@ from app.agents.note_agent import NOTE_TOOLS
 from app.agents.project_agent import PROJECT_TOOLS
 from app.agents.task_agent import TASK_TOOLS
 from app.agents.timeline_agent import TIMELINE_TOOLS
-from app.core.langfuse_client import extract_usage, get_langfuse, get_prompt_or_fallback, langfuse_context
+from app.core.llm import get_llm
-from app.core.llm import get_agent_llm, model_for_agent
 from app.core.memory_middleware import MemoryMiddleware
 from app.core.ws_context import clear_tool_result_collector, execute_on_client, set_tool_result_collector
 from app.db import async_session
@@ -27,35 +26,7 @@ logger = logging.getLogger(__name__)
 FloatingDomainType = Literal["task", "timeline", "project", "node"]
 FloatingDomainSection = Literal["task", "timeline", "note"]
 
-# Mapping of core-memory language values to natural-language names for prompts.
+_HOME_SINGLE_AGENT_SYSTEM = (
-_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",
-}
-
-
-def _language_instruction(context: dict[str, Any]) -> str:
-    """Return a system-prompt suffix that tells the LLM to respond in the user's language.
-
-    Returns an empty string when the language is English or unknown — saves tokens.
-    """
-    core = context.get("core_memory") or {}
-    raw = (core.get("language") or "").strip().lower()
-    if not raw:
-        return ""
-    lang = _LANGUAGE_NAMES.get(raw, raw.title())  # best-effort capitalisation
-    if lang.lower() == "english":
-        return ""
-    return (
-        f"\n\nIMPORTANT: Always respond in {lang}. "
-        f"All your output text must be written in {lang}."
-    )
-
-_HOME_SYSTEM_PROMPT = (
     "You are the home assistant with direct access to all tools: tasks, projects, notes, timelines, and memory tools. "
     "Always use tools for factual data retrieval before answering. "
     "When the user asks to remember, forget, or update what you know about them, use memory tools. "
@@ -68,7 +39,7 @@ _HOME_SYSTEM_PROMPT = (
     "For upcoming tasks, after tag lines add a short recommendation based on due date and priority."
 )
 
-_FLOATING_SYSTEM_PROMPT = (
+_FLOATING_SINGLE_AGENT_SYSTEM = (
     "You are the floating assistant with direct access to all tools: tasks, projects, notes, timelines, and memory tools. "
     "Stay focused on the floating scope in context.scope and answer concisely. "
     "Return plain text only. Do not output XML/HTML-like tags such as <task>, <project>, <note>, <timeline>, or any bracketed id tag wrappers. "
@@ -77,7 +48,7 @@ _FLOATING_SYSTEM_PROMPT = (
     "If context.context.resolved_project_id exists, use it as project_id for scoped list calls. "
 )
 
-_FLOATING_DOMAIN_CLASSIFIER_PROMPT = (
+_FLOATING_DOMAIN_CLASSIFIER_SYSTEM = (
     "You are a strict domain classifier for websocket floating requests. "
     "Return ONLY a JSON object with keys: type, id, section. "
     "Allowed type values: task, timeline, project, node. "
@@ -176,15 +147,6 @@ def _trace_id_from_context(context: dict[str, Any]) -> str | None:
     return None
 
 
-def _session_id_from_context(context: dict[str, Any]) -> str | None:
-    debug = context.get("_debug")
-    if isinstance(debug, dict):
-        session_id = debug.get("session_id")
-        if isinstance(session_id, str) and session_id:
-            return session_id
-    return None
-
-
 def _context_for_model(context: dict[str, Any]) -> dict[str, Any]:
     sanitized = dict(context)
     sanitized.pop("_debug", None)
@@ -573,9 +535,10 @@ async def _infer_floating_domain(message: str, context: dict[str, Any]) -> dict[
     }
 
     try:
-        llm = get_agent_llm("classifier")
+        llm = get_llm()
-        classifier_messages = [
+        response = await llm.ainvoke(
-            SystemMessage(content=_FLOATING_DOMAIN_CLASSIFIER_PROMPT),
+            [
+                SystemMessage(content=_FLOATING_DOMAIN_CLASSIFIER_SYSTEM),
                 HumanMessage(
                     content=(
                         f"Message:\n{message}\n\n"
@@ -583,29 +546,7 @@ async def _infer_floating_domain(message: str, context: dict[str, Any]) -> dict[
                     )
                 ),
             ]
-        lf = get_langfuse()
-        _, classifier_prompt_obj = get_prompt_or_fallback(
-            "floating_domain_classifier", _FLOATING_DOMAIN_CLASSIFIER_PROMPT
         )
-
-        # Extract user/session from context for Langfuse attribution
-        _debug = context.get("_debug") if isinstance(context, dict) else None
-        _lf_user = (_debug or {}).get("user_id") if isinstance(_debug, dict) else None
-        _lf_session = (_debug or {}).get("session_id") if isinstance(_debug, dict) else None
-
-        with langfuse_context(user_id=_lf_user, session_id=_lf_session):
-            if lf:
-                with lf.start_as_current_observation(
-                    as_type="generation",
-                    name="floating-classifier",
-                    model=model_for_agent("classifier"),
-                    prompt=classifier_prompt_obj,
-                    input=classifier_messages,
-                ) as gen:
-                    response = await llm.ainvoke(classifier_messages)
-                    gen.update(output=_as_text(response.content), usage_details=extract_usage(response))
-            else:
-                response = await llm.ainvoke(classifier_messages)
         parsed = _parse_json_object(_as_text(response.content))
         if parsed is not None:
             domain = _normalize_domain_payload(parsed, project_id)
@@ -630,13 +571,9 @@ async def _run_single_agent(
     message: str,
     context: dict[str, Any],
     max_steps: int = 6,
-    langfuse_prompt: Any = None,
-    agent_name: str = "agent",
 ) -> str:
     trace_id = _trace_id_from_context(context)
-    session_id = _session_id_from_context(context)
+    llm = get_llm()
-    lf = get_langfuse()
-    llm = get_agent_llm(agent_name)
     tools = _all_tools_for_user(user_id, trace_id)
     model_context = _context_for_model(context)
     logger.info("deep_agent: run_single_agent_start trace=%s user=%s", trace_id or "-", user_id)
@@ -654,39 +591,9 @@ async def _run_single_agent(
     tool_calls_count = 0
     collected: list[dict[str, Any]] = []
     set_tool_result_collector(collected)
-
-    _lf_ctx = langfuse_context(user_id=user_id, session_id=session_id)
-    _lf_ctx.__enter__()
-
-    _span_ctx = (
-        lf.start_as_current_observation(
-            as_type="span",
-            name=agent_name,
-            metadata={"user_id": user_id, "session_id": trace_id},
-            input=message,
-        )
-        if lf else None
-    )
-    _span = _span_ctx.__enter__() if _span_ctx else None
-
     try:
         for _ in range(max_steps):
-            _gen_ctx = (
-                lf.start_as_current_observation(
-                    as_type="generation",
-                    name=f"{agent_name}-llm",
-                    model=model_for_agent(agent_name),
-                    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)
-
             messages.append(response)
 
             if not response.tool_calls:
@@ -698,8 +605,6 @@ async def _run_single_agent(
                     tool_calls_count,
                     len(final_text),
                 )
-                if _span:
-                    _span.update(output=final_text)
                 return final_text
 
             tool_map = {tool_def.name: tool_def for tool_def in tools}
@@ -739,16 +644,9 @@ async def _run_single_agent(
             tool_calls_count,
             len(final_text),
         )
-        if _span:
-            _span.update(output=final_text)
         return final_text
     finally:
         clear_tool_result_collector()
-        if _span_ctx:
-            _span_ctx.__exit__(None, None, None)
-        _lf_ctx.__exit__(None, None, None)
-        if lf:
-            lf.flush()
 
 
 async def _run_single_agent_stream(
@@ -758,13 +656,9 @@ async def _run_single_agent_stream(
     message: str,
     context: dict[str, Any],
     max_steps: int = 6,
-    langfuse_prompt: Any = None,
-    agent_name: str = "agent",
 ) -> AsyncGenerator[tuple[str, Any], None]:
     trace_id = _trace_id_from_context(context)
-    session_id = _session_id_from_context(context)
+    llm = get_llm()
-    lf = get_langfuse()
-    llm = get_agent_llm(agent_name)
     tools = _all_tools_for_user(user_id, trace_id)
     model_context = _context_for_model(context)
     logger.info("deep_agent: run_single_agent_stream_start trace=%s user=%s", trace_id or "-", user_id)
@@ -783,40 +677,9 @@ async def _run_single_agent_stream(
     streamed_chars = 0
     collected: list[dict[str, Any]] = []
     set_tool_result_collector(collected)
-
-    _lf_ctx = langfuse_context(user_id=user_id, session_id=session_id)
-    _lf_ctx.__enter__()
-
-    _span_ctx = (
-        lf.start_as_current_observation(
-            as_type="span",
-            name=f"{agent_name}-stream",
-            metadata={"user_id": user_id, "session_id": trace_id},
-            input=message,
-        )
-        if lf else None
-    )
-    _span = _span_ctx.__enter__() if _span_ctx else None
-    streamed_text: list[str] = []
-
     try:
         for _ in range(max_steps):
-            _gen_ctx = (
-                lf.start_as_current_observation(
-                    as_type="generation",
-                    name=f"{agent_name}-llm",
-                    model=model_for_agent(agent_name),
-                    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)
-
             messages.append(response)
 
             if not response.tool_calls:
@@ -825,7 +688,6 @@ async def _run_single_agent_stream(
                     token = _as_text(getattr(chunk, "content", ""))
                     if token:
                         streamed_chars += len(token)
-                        streamed_text.append(token)
                         emitted_any = True
                         yield "token", token
 
@@ -834,7 +696,6 @@ async def _run_single_agent_stream(
                     fallback_text = _as_text(response.content)
                     if fallback_text:
                         streamed_chars += len(fallback_text)
-                        streamed_text.append(fallback_text)
                         yield "token", fallback_text
                 logger.info(
                     "deep_agent: run_single_agent_stream_end trace=%s user=%s tool_calls=%d response_chars=%d",
@@ -843,8 +704,6 @@ async def _run_single_agent_stream(
                     tool_calls_count,
                     streamed_chars,
                 )
-                if _span:
-                    _span.update(output="".join(streamed_text))
                 return
 
             tool_map = {tool_def.name: tool_def for tool_def in tools}
@@ -879,7 +738,6 @@ async def _run_single_agent_stream(
             token = _as_text(getattr(chunk, "content", ""))
             if token:
                 streamed_chars += len(token)
-                streamed_text.append(token)
                 yield "token", token
         logger.info(
             "deep_agent: run_single_agent_stream_end trace=%s user=%s tool_calls=%d response_chars=%d fallback=1",
@@ -888,30 +746,17 @@ async def _run_single_agent_stream(
             tool_calls_count,
             streamed_chars,
         )
-        if _span:
-            _span.update(output="".join(streamed_text))
     finally:
         clear_tool_result_collector()
-        if _span_ctx:
-            _span_ctx.__exit__(None, None, None)
-        _lf_ctx.__exit__(None, None, None)
-        if lf:
-            lf.flush()
 
 
 async def run_home(user_id: str, message: str, context: dict[str, Any]) -> str:
     prepared_context = await _prepare_context(message, context)
-    system_prompt, langfuse_prompt = get_prompt_or_fallback(
-        "home_system", _HOME_SYSTEM_PROMPT
-    )
-    system_prompt += _language_instruction(context)
     response = await _run_single_agent(
         user_id=user_id,
-        system_prompt=system_prompt,
+        system_prompt=_HOME_SINGLE_AGENT_SYSTEM,
         message=message,
         context=prepared_context,
-        langfuse_prompt=langfuse_prompt,
-        agent_name="home-agent",
     )
     return _normalize_tagged_list_lines(response, message)
 
@@ -919,17 +764,11 @@ async def run_home(user_id: str, message: str, context: dict[str, Any]) -> str:
 async def run_floating(user_id: str, message: str, context: dict[str, Any]) -> tuple[str, dict[str, str | None]]:
     prepared_context = await _prepare_context(message, context)
     domain = await _infer_floating_domain(message, prepared_context)
-    system_prompt, langfuse_prompt = get_prompt_or_fallback(
-        "floating_system", _FLOATING_SYSTEM_PROMPT
-    )
-    system_prompt += _language_instruction(context)
     response = await _run_single_agent(
         user_id=user_id,
-        system_prompt=system_prompt,
+        system_prompt=_FLOATING_SINGLE_AGENT_SYSTEM,
         message=message,
         context=prepared_context,
-        langfuse_prompt=langfuse_prompt,
-        agent_name="floating-agent",
     )
     sanitized = _strip_floating_markup(response)
     if not sanitized and response:
@@ -943,18 +782,12 @@ async def run_home_stream(
     context: dict[str, Any],
 ) -> AsyncGenerator[tuple[str, Any], None]:
     prepared_context = await _prepare_context(message, context)
-    system_prompt, langfuse_prompt = get_prompt_or_fallback(
-        "home_system", _HOME_SYSTEM_PROMPT
-    )
-    system_prompt += _language_instruction(context)
     text_chunks: list[str] = []
     async for event in _run_single_agent_stream(
         user_id=user_id,
-        system_prompt=system_prompt,
+        system_prompt=_HOME_SINGLE_AGENT_SYSTEM,
         message=message,
         context=prepared_context,
-        langfuse_prompt=langfuse_prompt,
-        agent_name="home-agent",
     ):
         event_type, data = event
         if event_type != "token":
@@ -976,20 +809,14 @@ async def run_floating_stream(
     domain = await _infer_floating_domain(message, prepared_context)
     yield "floating_domain", domain
 
-    system_prompt, langfuse_prompt = get_prompt_or_fallback(
-        "floating_system", _FLOATING_SYSTEM_PROMPT
-    )
-    system_prompt += _language_instruction(context)
     sanitizer = _FloatingStreamSanitizer()
     emitted_sanitized = False
     raw_chunks: list[str] = []
     async for event in _run_single_agent_stream(
         user_id=user_id,
-        system_prompt=system_prompt,
+        system_prompt=_FLOATING_SINGLE_AGENT_SYSTEM,
         message=message,
         context=prepared_context,
-        langfuse_prompt=langfuse_prompt,
-        agent_name="floating-agent",
     ):
         event_type, data = event
         if event_type != "token":

app/core/langfuse_client.py

@@ -1,190 +0,0 @@
-"""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

@@ -19,7 +19,6 @@ from __future__ import annotations
 
 import os
 import warnings
-from collections.abc import Callable
 
 from openai import AsyncOpenAI
 import litellm
@@ -51,6 +50,8 @@ def _api_key_for_model(model: str) -> str | None:
         return settings.GOOGLE_API_KEY or None
     if model.startswith("cerebras/"):
         return settings.CEREBRAS_API_KEY or None
+    if model.startswith("github/"):
+        return settings.GITHUB_TOKEN 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.
@@ -84,6 +85,9 @@ def get_llm(
     if settings.GITHUB_COPILOT_TOKEN_DIR:
         os.environ.setdefault("GITHUB_COPILOT_TOKEN_DIR", settings.GITHUB_COPILOT_TOKEN_DIR)
 
+    if settings.GITHUB_TOKEN:
+        os.environ.setdefault("GITHUB_TOKEN", settings.GITHUB_TOKEN)
+
     # 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:
@@ -96,35 +100,6 @@ def get_llm(
     )
 
 
-_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,
-    "floating-agent":      lambda: settings.LLM_MODEL_FLOATING_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,
-    "setup":               lambda: settings.LLM_MODEL_SETUP_AGENT or settings.LLM_MODEL,
-}
-
-
-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*.
 

app/core/preprocessors/__init__.py

@@ -1,104 +0,0 @@
-"""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

@@ -1,25 +0,0 @@
-"""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

@@ -1,111 +0,0 @@
-"""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/integrations/ms_graph.py

@@ -25,7 +25,7 @@ from __future__ import annotations
 
 import logging
 import re
-from datetime import datetime, timezone
+from datetime import datetime, timedelta, timezone
 from typing import Any
 
 import httpx

app/main.py

@@ -30,7 +30,7 @@ async def lifespan(app: FastAPI):
 
 def create_app() -> FastAPI:
     app = FastAPI(
-        title="AdiuvAI Cloud API",
+        title="Adiuva Cloud API",
         version="0.1.0",
         docs_url="/docs" if settings.ENV == "dev" else None,
         redoc_url=None,
@@ -50,10 +50,14 @@ def create_app() -> FastAPI:
     app.add_middleware(SanitizerMiddleware)
     app.add_middleware(TierRateLimitMiddleware)
 
-    from app.api.routes import agents, auth, billing, chat, device_ws
+    from app.api.routes import agents, auth, backup, billing, chat, device_ws, plugins, storage, vectors
 
     app.include_router(auth.router,       prefix="/api/v1")
     app.include_router(chat.router,       prefix="/api/v1")
+    app.include_router(storage.router,    prefix="/api/v1")
+    app.include_router(vectors.router,    prefix="/api/v1")
+    app.include_router(backup.router,     prefix="/api/v1")
+    app.include_router(plugins.router,    prefix="/api/v1")
     app.include_router(billing.router,    prefix="/api/v1")
     app.include_router(agents.router,     prefix="/api/v1")
     app.include_router(device_ws.router,  prefix="/api/v1")

app/marketplace/__init__.py

@@ -0,0 +1,7 @@
+"""Plugin marketplace package.
+
+Three service classes introduced in Step 10:
+  - ``PluginRegistry``  — catalog, submit/approve/reject, install counts
+  - ``ReviewQueue``     — approval workflow + security checklist
+  - ``RevenueShare``    — 70/30 split tracking and Stripe Connect payouts
+"""

app/marketplace/plugin_registry.py

@@ -0,0 +1,212 @@
+"""Plugin catalog registry backed by PostgreSQL.
+
+Maintains the authoritative list of plugins, their review status, and
+aggregate install counts.  All data is persisted in the ``plugins`` table.
+
+Module-level singleton::
+
+    from app.marketplace.plugin_registry import registry
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any, Literal
+
+from sqlalchemy import select, func
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models import Plugin
+from app.schemas import PluginListResponse, PluginManifest
+
+_PAGE_SIZE = 20
+
+
+def _plugin_to_manifest(p: Plugin) -> PluginManifest:
+    """Convert an ORM ``Plugin`` row to a Pydantic ``PluginManifest``."""
+    try:
+        permissions = json.loads(p.permissions) if p.permissions else []
+    except (json.JSONDecodeError, TypeError):
+        permissions = []
+    return PluginManifest(
+        id=p.id,
+        name=p.name,
+        description=p.description,
+        version=p.version,
+        author=p.author_name,
+        permissions=permissions,
+        category=p.category,
+        price_cents=p.price_cents,
+    )
+
+
+class PluginRegistry:
+    """PostgreSQL-backed plugin catalog.
+
+    All methods accept an ``AsyncSession`` parameter so the calling route
+    controls the session lifecycle.
+    """
+
+    # ── Queries ──────────────────────────────────────────────────────
+
+    async def list_plugins(
+        self,
+        db: AsyncSession,
+        category: str | None = None,
+        query: str | None = None,
+        page: int = 1,
+        sort: Literal["rating", "installs", "newest"] = "newest",
+    ) -> PluginListResponse:
+        """Return a page of approved plugins, optionally filtered and sorted."""
+        base = select(Plugin).where(Plugin.status == "approved")
+
+        if category:
+            base = base.where(Plugin.category == category)
+        if query:
+            pattern = f"%{query}%"
+            base = base.where(
+                Plugin.name.ilike(pattern) | Plugin.description.ilike(pattern)
+            )
+
+        # Count
+        count_q = select(func.count()).select_from(base.subquery())
+        total = (await db.execute(count_q)).scalar_one()
+
+        # Sort
+        if sort == "installs":
+            base = base.order_by(Plugin.install_count.desc())
+        elif sort == "rating":
+            base = base.order_by(Plugin.avg_rating.desc())
+        else:  # newest
+            base = base.order_by(Plugin.created_at.desc())
+
+        base = base.offset((page - 1) * _PAGE_SIZE).limit(_PAGE_SIZE)
+        rows = (await db.execute(base)).scalars().all()
+
+        return PluginListResponse(
+            plugins=[_plugin_to_manifest(r) for r in rows],
+            total=total,
+            page=page,
+        )
+
+    async def get_plugin(self, db: AsyncSession, plugin_id: str) -> dict[str, Any] | None:
+        """Return ``{manifest, status, install_count, avg_rating}`` or ``None``."""
+        result = await db.execute(select(Plugin).where(Plugin.id == plugin_id))
+        p = result.scalar_one_or_none()
+        if p is None:
+            return None
+        return {
+            "manifest": _plugin_to_manifest(p),
+            "status": p.status,
+            "install_count": p.install_count,
+            "avg_rating": p.avg_rating,
+        }
+
+    # ── Mutations ────────────────────────────────────────────────────
+
+    async def submit_plugin(
+        self,
+        db: AsyncSession,
+        manifest: PluginManifest,
+        package_s3_key: str,
+    ) -> str:
+        """Add *manifest* to the catalog with ``status='pending_review'``.
+
+        Returns the plugin_id.  If a plugin with the same id already exists
+        it is overwritten (re-submission after rejection).
+        """
+        plugin_id = manifest.id
+        existing = await db.execute(select(Plugin).where(Plugin.id == plugin_id))
+        row = existing.scalar_one_or_none()
+
+        if row is not None:
+            row.name = manifest.name
+            row.description = manifest.description
+            row.version = manifest.version
+            row.author_name = manifest.author
+            row.category = manifest.category
+            row.price_cents = manifest.price_cents
+            row.permissions = json.dumps(manifest.permissions)
+            row.status = "pending_review"
+            row.s3_package_key = package_s3_key
+            row.rejection_reason = None
+        else:
+            row = Plugin(
+                id=plugin_id,
+                name=manifest.name,
+                description=manifest.description,
+                version=manifest.version,
+                author_name=manifest.author,
+                category=manifest.category,
+                price_cents=manifest.price_cents,
+                permissions=json.dumps(manifest.permissions),
+                status="pending_review",
+                s3_package_key=package_s3_key,
+                install_count=0,
+                avg_rating=0.0,
+            )
+            db.add(row)
+        await db.commit()
+        return plugin_id
+
+    async def approve_plugin(self, db: AsyncSession, plugin_id: str) -> None:
+        """Set *plugin_id* status to ``'approved'``.
+
+        Raises ``KeyError`` if the plugin is not found.
+        """
+        result = await db.execute(select(Plugin).where(Plugin.id == plugin_id))
+        row = result.scalar_one_or_none()
+        if row is None:
+            raise KeyError(f"Plugin not found: {plugin_id}")
+        row.status = "approved"
+        row.rejection_reason = None
+        await db.commit()
+
+    async def reject_plugin(self, db: AsyncSession, plugin_id: str, reason: str) -> None:
+        """Set *plugin_id* status to ``'rejected'`` and record the reason.
+
+        Raises ``KeyError`` if the plugin is not found.
+        """
+        result = await db.execute(select(Plugin).where(Plugin.id == plugin_id))
+        row = result.scalar_one_or_none()
+        if row is None:
+            raise KeyError(f"Plugin not found: {plugin_id}")
+        row.status = "rejected"
+        row.rejection_reason = reason
+        await db.commit()
+
+    async def record_install(self, db: AsyncSession, plugin_id: str) -> None:
+        """Increment the install count for *plugin_id* (no-op if not found)."""
+        result = await db.execute(select(Plugin).where(Plugin.id == plugin_id))
+        row = result.scalar_one_or_none()
+        if row is not None:
+            row.install_count = row.install_count + 1
+            await db.commit()
+
+    async def record_uninstall(self, db: AsyncSession, plugin_id: str) -> None:
+        """Decrement the install count for *plugin_id*, floored at 0."""
+        result = await db.execute(select(Plugin).where(Plugin.id == plugin_id))
+        row = result.scalar_one_or_none()
+        if row is not None:
+            row.install_count = max(0, row.install_count - 1)
+            await db.commit()
+
+    # ── Internal helpers used by ReviewQueue ─────────────────────────
+
+    async def get_pending_entries(self, db: AsyncSession) -> list[dict[str, Any]]:
+        """Return all entries with status='pending_review'."""
+        result = await db.execute(
+            select(Plugin).where(Plugin.status == "pending_review")
+        )
+        rows = result.scalars().all()
+        return [
+            {
+                "manifest": _plugin_to_manifest(r),
+                "submitted_at": int(r.submitted_at.timestamp()) if r.submitted_at else 0,
+            }
+            for r in rows
+        ]
+
+
+# Module-level singleton
+registry = PluginRegistry()

app/marketplace/plugin_review.py

@@ -0,0 +1,125 @@
+"""Plugin review workflow backed by PostgreSQL.
+
+Manages the approval queue for newly submitted plugins and enforces a
+security checklist before any plugin is made visible in the marketplace.
+
+Module-level singleton::
+
+    from app.marketplace.plugin_review import review_queue
+"""
+
+from __future__ import annotations
+
+import re
+from typing import Any, Literal
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.marketplace.plugin_registry import registry
+from app.models import PluginReview as PluginReviewModel
+from app.schemas import PluginManifest
+
+# ── Security policy ───────────────────────────────────────────────────
+
+ALLOWED_PERMISSIONS: frozenset[str] = frozenset(
+    {
+        "read:tasks",
+        "write:tasks",
+        "read:projects",
+        "write:projects",
+        "read:notes",
+        "write:notes",
+        "read:timelines",
+        "write:timelines",
+        "read:calendar",
+        "write:calendar",
+    }
+)
+
+_PLUGIN_ID_RE = re.compile(r"^[a-z0-9-]+$")
+
+
+def validate_manifest(manifest: PluginManifest) -> None:
+    """Enforce the plugin security checklist.
+
+    Raises:
+        ``ValueError`` on the first violation found.  Callers should catch
+        this and return HTTP 422 / reject the submission.
+
+    Checks:
+      1. Plugin id matches ``^[a-z0-9-]+$``
+      2. All declared permissions are in ``ALLOWED_PERMISSIONS``
+      3. No manifest field contains raw binary data
+    """
+    if not _PLUGIN_ID_RE.match(manifest.id):
+        raise ValueError(
+            f"Invalid plugin id format: '{manifest.id}'. "
+            "Only lowercase letters, digits, and hyphens are allowed."
+        )
+
+    for perm in manifest.permissions:
+        if perm not in ALLOWED_PERMISSIONS:
+            raise ValueError(
+                f"Unknown permission: '{perm}'. "
+                f"Allowed permissions: {sorted(ALLOWED_PERMISSIONS)}"
+            )
+
+    for field_name, value in manifest.model_dump().items():
+        if isinstance(value, (bytes, bytearray)):
+            raise ValueError(
+                f"Binary content is not allowed in manifest field '{field_name}'."
+            )
+
+
+class ReviewQueue:
+    """Approval queue for pending plugin submissions.
+
+    Delegates status changes to the shared ``PluginRegistry`` singleton.
+    Review records are persisted in the ``plugin_reviews`` table.
+    """
+
+    async def get_pending(self, db: AsyncSession) -> list[dict[str, Any]]:
+        """Return all plugins currently awaiting review.
+
+        Each item is ``{plugin_id, manifest, submitted_at}``.
+        """
+        entries = await registry.get_pending_entries(db)
+        return [
+            {
+                "plugin_id": e["manifest"].id,
+                "manifest": e["manifest"],
+                "submitted_at": e["submitted_at"],
+            }
+            for e in entries
+        ]
+
+    async def submit_review(
+        self,
+        db: AsyncSession,
+        plugin_id: str,
+        reviewer_id: str,
+        decision: Literal["approved", "rejected"],
+        notes: str = "",
+    ) -> None:
+        """Record a review decision and update the plugin's status.
+
+        Raises:
+            ``KeyError`` if *plugin_id* is not found in the registry.
+        """
+        if decision == "approved":
+            await registry.approve_plugin(db, plugin_id)
+        else:
+            await registry.reject_plugin(db, plugin_id, reason=notes)
+
+        review = PluginReviewModel(
+            plugin_id=plugin_id,
+            reviewer_id=reviewer_id,
+            decision=decision,
+            notes=notes,
+        )
+        db.add(review)
+        await db.commit()
+
+
+# Module-level singleton
+review_queue = ReviewQueue()

app/marketplace/revenue_share.py

@@ -0,0 +1,233 @@
+"""Revenue share tracking and Stripe Connect payouts backed by PostgreSQL.
+
+Records every plugin installation as a revenue event and facilitates
+70 % / 30 % payouts to developers via Stripe Connect.  Data is persisted
+in the ``revenue_events`` table.
+
+Module-level singleton::
+
+    from app.marketplace.revenue_share import revenue_share
+"""
+
+from __future__ import annotations
+
+import logging
+from datetime import datetime, timezone
+from typing import Any
+
+import stripe as stripe_lib
+from sqlalchemy import extract, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.config.settings import settings
+from app.marketplace.plugin_registry import registry
+from app.models import Plugin, RevenueEvent
+
+logger = logging.getLogger(__name__)
+
+# ── Revenue split constants ───────────────────────────────────────────
+
+DEVELOPER_SHARE: float = 0.70
+PLATFORM_SHARE: float = 0.30
+
+
+class RevenueShare:
+    """Records installation revenue events and coordinates developer payouts.
+
+    Stripe Connect calls are gracefully stubbed when ``STRIPE_SECRET_KEY``
+    is not configured, consistent with the rest of the billing layer.
+    """
+
+    # ── Helpers ──────────────────────────────────────────────────────
+
+    @staticmethod
+    def _stripe_configured() -> bool:
+        return bool(settings.STRIPE_SECRET_KEY)
+
+    @staticmethod
+    def _stripe() -> Any:
+        stripe_lib.api_key = settings.STRIPE_SECRET_KEY
+        return stripe_lib
+
+    # ── Core operations ──────────────────────────────────────────────
+
+    async def record_install(
+        self,
+        db: AsyncSession,
+        plugin_id: str,
+        user_id: str,
+        amount_cents: int,
+    ) -> None:
+        """Record a plugin installation and trigger a Stripe Connect charge if paid.
+
+        For free plugins (``amount_cents == 0``) no payment is initiated but
+        the event is still recorded for analytics.
+
+        For paid plugins the developer receives 70 % via a Stripe Connect
+        destination charge.  If Stripe is not configured or the charge fails
+        the installation still succeeds (the event is recorded and the install
+        count is incremented) — a warning is logged for monitoring.
+        """
+        developer_share_cents = int(amount_cents * DEVELOPER_SHARE)
+        stripe_transfer_id: str | None = None
+
+        if amount_cents > 0 and self._stripe_configured():
+            # Look up the plugin's author Stripe account from the DB
+            result = await db.execute(select(Plugin).where(Plugin.id == plugin_id))
+            plugin_row = result.scalar_one_or_none()
+            developer_stripe_account: str | None = None
+            if plugin_row and plugin_row.author_id:
+                # Future: look up user.stripe_connect_account_id
+                developer_stripe_account = None  # no real account yet
+
+            if developer_stripe_account:
+                try:
+                    s = self._stripe()
+                    transfer = s.Transfer.create(
+                        amount=developer_share_cents,
+                        currency="eur",
+                        destination=developer_stripe_account,
+                        description=f"Revenue share for plugin {plugin_id}",
+                        metadata={"plugin_id": plugin_id, "user_id": user_id},
+                    )
+                    stripe_transfer_id = transfer["id"]
+                except Exception as exc:
+                    logger.warning(
+                        "Stripe Connect transfer failed for plugin %s: %s",
+                        plugin_id,
+                        exc,
+                    )
+            else:
+                logger.debug(
+                    "No Stripe account on file for plugin %s developer; "
+                    "skipping transfer.",
+                    plugin_id,
+                )
+
+        event = RevenueEvent(
+            plugin_id=plugin_id,
+            user_id=user_id,
+            amount_cents=amount_cents,
+            developer_share_cents=developer_share_cents,
+            stripe_transfer_id=stripe_transfer_id,
+        )
+        db.add(event)
+        await db.commit()
+
+        await registry.record_install(db, plugin_id)
+
+    async def get_earnings(
+        self,
+        db: AsyncSession,
+        developer_id: str,
+        period: str | None = None,
+    ) -> dict[str, Any]:
+        """Return aggregated earnings for *developer_id*.
+
+        ``period`` is an optional ``YYYY-MM`` string to restrict the window.
+
+        Returns::
+
+            {
+                "developer_id": str,
+                "period": str | None,
+                "total_installs": int,
+                "total_revenue_cents": int,
+                "developer_share_cents": int,
+            }
+        """
+        # Find plugin ids belonging to this developer (by author_name match)
+        plugin_q = select(Plugin.id).where(Plugin.author_name == developer_id)
+        plugin_result = await db.execute(plugin_q)
+        developer_plugin_ids = [row[0] for row in plugin_result.all()]
+
+        if not developer_plugin_ids:
+            return {
+                "developer_id": developer_id,
+                "period": period,
+                "total_installs": 0,
+                "total_revenue_cents": 0,
+                "developer_share_cents": 0,
+            }
+
+        query = select(
+            func.count().label("total_installs"),
+            func.coalesce(func.sum(RevenueEvent.amount_cents), 0).label("total_revenue"),
+            func.coalesce(func.sum(RevenueEvent.developer_share_cents), 0).label("dev_share"),
+        ).where(RevenueEvent.plugin_id.in_(developer_plugin_ids))
+
+        if period:
+            # Filter by YYYY-MM: extract year and month from created_at
+            try:
+                year, month = period.split("-")
+                query = query.where(
+                    extract("year", RevenueEvent.created_at) == int(year),
+                    extract("month", RevenueEvent.created_at) == int(month),
+                )
+            except ValueError:
+                pass  # invalid period format — return all
+
+        result = await db.execute(query)
+        row = result.one()
+
+        return {
+            "developer_id": developer_id,
+            "period": period,
+            "total_installs": row.total_installs,
+            "total_revenue_cents": row.total_revenue,
+            "developer_share_cents": row.dev_share,
+        }
+
+    async def payout_developer(self, db: AsyncSession, plugin_id: str, period: str) -> None:
+        """Aggregate unpaid revenue for *period* and issue a Stripe Transfer.
+
+        Marks processed events with ``paid_at`` timestamp.
+        Stubs gracefully when Stripe is not configured.
+        """
+        try:
+            year, month = period.split("-")
+            year_int, month_int = int(year), int(month)
+        except ValueError:
+            logger.warning("Invalid period format: %s", period)
+            return
+
+        result = await db.execute(
+            select(RevenueEvent).where(
+                RevenueEvent.plugin_id == plugin_id,
+                RevenueEvent.paid_at.is_(None),
+                extract("year", RevenueEvent.created_at) == year_int,
+                extract("month", RevenueEvent.created_at) == month_int,
+            )
+        )
+        unpaid = list(result.scalars().all())
+
+        total_dev_share = sum(e.developer_share_cents for e in unpaid)
+        if total_dev_share <= 0 or not unpaid:
+            logger.debug("Nothing to pay out for plugin %s in period %s", plugin_id, period)
+            return
+
+        if self._stripe_configured():
+            plugin_result = await db.execute(select(Plugin).where(Plugin.id == plugin_id))
+            plugin_row = plugin_result.scalar_one_or_none()
+            developer_stripe_account: str | None = None  # Future: fetch from DB
+            if plugin_row and developer_stripe_account:
+                try:
+                    s = self._stripe()
+                    s.Transfer.create(
+                        amount=total_dev_share,
+                        currency="eur",
+                        destination=developer_stripe_account,
+                        description=f"Payout for plugin {plugin_id} period {period}",
+                    )
+                except Exception as exc:
+                    logger.warning("Payout transfer failed for plugin %s: %s", plugin_id, exc)
+                    return
+
+        paid_ts = datetime.now(timezone.utc)
+        for event in unpaid:
+            event.paid_at = paid_ts
+        await db.commit()
+
+
+# Module-level singleton
+revenue_share = RevenueShare()

app/models.py

@@ -1,15 +1,19 @@
 """SQLAlchemy ORM models for all persistent tables.
 
-Only auth, billing, agent config, and memory data live here.
+Only auth, billing, storage metadata, and marketplace data live here.
-User content (notes, tasks, etc.) lives exclusively on the client.
+User content (notes, tasks, etc.) is NEVER persisted server-side —
+it lives in E2E-encrypted blobs in S3, referenced by storage_records.
 
 Table inventory:
   users               — account credentials + tier
   refresh_tokens      — hashed refresh token store
   subscriptions       — Stripe subscription records
-  local_agent_configs — per-device batch agent configs
+  storage_records     — S3 blob metadata (no plaintext)
-  cloud_agent_configs — OAuth-backed cloud agent configs
+  backup_metadata     — encrypted backup manifests
-  agent_run_logs      — execution history for all agents
+  plugins             — marketplace plugin catalog
+  plugin_installations — per-user install records
+  plugin_reviews      — admin review decisions
+  revenue_events      — Stripe Connect 70/30 split ledger
   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)
@@ -22,6 +26,7 @@ import uuid
 from datetime import datetime, timezone
 
 from sqlalchemy import (
+    BigInteger,
     Boolean,
     DateTime,
     Enum,
@@ -31,6 +36,7 @@ from sqlalchemy import (
     JSON,
     String,
     Text,
+    UniqueConstraint,
     Uuid,
     func,
 )
@@ -52,6 +58,8 @@ 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")
@@ -69,8 +77,7 @@ 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 | None] = mapped_column(String(255), nullable=True)
+    password_hash: Mapped[str] = mapped_column(String(255), nullable=False)
-    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.
@@ -79,9 +86,6 @@ class User(Base):
     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()
     )
@@ -92,9 +96,6 @@ 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):
@@ -115,25 +116,6 @@ 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"
 
@@ -155,6 +137,151 @@ class Subscription(Base):
     user: Mapped[User] = relationship(back_populates="subscription")
 
 
+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()
+    )
+
+
+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")
+    # nullable until developer account system is built
+    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="[]")  # JSON list
+    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")
+
+
 class LocalAgentConfig(Base):
     __tablename__ = "local_agent_configs"
 
@@ -169,7 +296,6 @@ 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="")
-    agent_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)

app/schemas.py

@@ -30,16 +30,6 @@ 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 ─────────────────────────────────────────────────────────────
@@ -60,6 +50,88 @@ 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   # encrypted vector + metadata — backend never decrypts
+    checksum: str
+
+
+class VectorUpsertRequest(BaseModel):
+    vectors: list[VectorItem]
+
+
+class VectorSearchRequest(BaseModel):
+    query_blob: bytes   # encrypted query — backend never decrypts
+    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):
@@ -201,27 +273,6 @@ class WsFloatingDomain(BaseModel):
     domain: WsDomain
 
 
-# ── Agent Config V2 ───────────────────────────────────────────────────
-
-
-class ContentTypeConfig(BaseModel):
-    """Per-type extraction config produced by the journey chatbot."""
-
-    id: str
-    label: str = ""
-    detection_hint: str = ""
-    preprocessing: str = "generic"  # handler name: "email_html", "plain_text", ...
-    extraction_prompt: str
-
-
-class AgentConfig(BaseModel):
-    """Structured agent configuration (replaces freeform prompt_template)."""
-
-    content_types: list[ContentTypeConfig] = []
-    global_rules: list[str] = []
-    data_types: list[str] = []
-
-
 # ── Agent Catalog ─────────────────────────────────────────────────────
 
 class AgentCatalogItem(BaseModel):
@@ -246,11 +297,10 @@ class AgentTriggerRequest(BaseModel):
     device_id: str = Field(default="")
     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 | None = None
+    custom_agent_prompt: str = Field(min_length=1)
-    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 ─────────────────────────────────────────────────────

app/storage/__init__.py

@@ -0,0 +1 @@
+"""Cloud storage layer — E2E encrypted blobs and vectors."""

app/storage/blob_store.py

@@ -0,0 +1,106 @@
+"""S3-backed store for E2E-encrypted blobs.
+
+Keys are structured as ``{user_id}/{table}/{record_id}``.
+The backend never inspects blob content — it stores and retrieves opaque bytes.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import boto3
+
+from app.config.settings import settings
+
+
+class BlobStore:
+    """Thin wrapper around boto3 S3.
+
+    All blobs must be E2E encrypted by the client before upload.
+    The backend adds SSE-S3 as an extra layer of at-rest encryption
+    but cannot decrypt the inner client-side payload.
+    """
+
+    def _client(self) -> Any:
+        kwargs: dict[str, Any] = {
+            "region_name": settings.S3_REGION,
+            "aws_access_key_id": settings.AWS_ACCESS_KEY_ID,
+            "aws_secret_access_key": settings.AWS_SECRET_ACCESS_KEY,
+        }
+        if settings.S3_ENDPOINT_URL and isinstance(settings.S3_ENDPOINT_URL, str):
+            kwargs["endpoint_url"] = settings.S3_ENDPOINT_URL
+        return boto3.client("s3", **kwargs)
+
+    @staticmethod
+    def _key(user_id: str, table: str, record_id: str) -> str:
+        return f"{user_id}/{table}/{record_id}"
+
+    async def upload(
+        self,
+        user_id: str,
+        table: str,
+        record_id: str,
+        blob: bytes,
+        checksum: str,
+    ) -> str:
+        """Store *blob* in S3 and return the S3 key.
+
+        Args:
+            user_id:   Owner of the blob (used as key prefix).
+            table:     Logical table name (e.g. ``"tasks"``).
+            record_id: Record UUID.
+            blob:      Raw bytes (pre-encrypted by client).
+            checksum:  SHA-256 hex digest supplied by the client; stored as
+                       object metadata for download-time verification.
+
+        Returns:
+            The S3 key under which the blob was stored.
+        """
+        key = self._key(user_id, table, record_id)
+        self._client().put_object(
+            Bucket=settings.S3_BUCKET,
+            Key=key,
+            Body=blob,
+            ServerSideEncryption="AES256",  # SSE-S3 at rest
+            Metadata={"checksum": checksum},
+        )
+        return key
+
+    async def download(self, user_id: str, s3_key: str) -> bytes:
+        """Retrieve the blob stored at *s3_key*.
+
+        *user_id* is retained in the signature so higher-level code can
+        enforce ownership without re-parsing the key.
+
+        Raises:
+            ``botocore.exceptions.ClientError`` with code ``NoSuchKey`` if the
+            object does not exist.
+        """
+        response = self._client().get_object(
+            Bucket=settings.S3_BUCKET,
+            Key=s3_key,
+        )
+        return response["Body"].read()
+
+    async def delete(self, user_id: str, s3_key: str) -> None:
+        """Delete the object at *s3_key*.
+
+        S3 ``delete_object`` is idempotent — it succeeds even if the key does
+        not exist.
+        """
+        self._client().delete_object(
+            Bucket=settings.S3_BUCKET,
+            Key=s3_key,
+        )
+
+    async def list_keys(self, user_id: str, table: str) -> list[str]:
+        """Return all S3 keys for a given user + table combination.
+
+        Uses the prefix ``{user_id}/{table}/`` to scope the listing.
+        """
+        prefix = f"{user_id}/{table}/"
+        response = self._client().list_objects_v2(
+            Bucket=settings.S3_BUCKET,
+            Prefix=prefix,
+        )
+        return [obj["Key"] for obj in response.get("Contents", [])]

app/storage/encryption.py

@@ -0,0 +1,32 @@
+"""Integrity verification only — the backend NEVER decrypts user data."""
+
+from __future__ import annotations
+
+import hashlib
+import hmac
+
+from fastapi import HTTPException
+
+
+def verify_checksum(blob: bytes, checksum: str) -> bool:
+    """Return ``True`` if SHA-256(blob) matches *checksum*.
+
+    Uses ``hmac.compare_digest`` for constant-time comparison to prevent
+    timing-based side-channel attacks.
+    """
+    computed = hashlib.sha256(blob).hexdigest()
+    return hmac.compare_digest(computed, checksum)
+
+
+def reject_if_tampered(blob: bytes, checksum: str) -> None:
+    """Raise ``HTTP 400`` if the blob does not match its checksum.
+
+    Call this before storing or forwarding any client-provided blob.
+    The backend never holds decryption keys — this check only verifies
+    that the opaque bytes arrived intact.
+    """
+    if not verify_checksum(blob, checksum):
+        raise HTTPException(
+            status_code=400,
+            detail="Checksum mismatch: blob integrity check failed",
+        )

app/storage/vector_store.py

@@ -0,0 +1,205 @@
+"""Cloud vector store — wraps Pinecone (default) or Qdrant.
+
+Vectors are pre-encrypted blobs from the client.  The backend stores them
+alongside a deterministic 32-dim float representation derived from the blob's
+SHA-256 hash.  Semantic ANN search is not meaningful on encrypted data — this
+is a known trade-off documented in the backend plan.
+
+Isolation: Pinecone uses ``namespace=user_id``; Qdrant filters by
+``user_id`` payload field on a shared collection.
+"""
+
+from __future__ import annotations
+
+import base64
+import hashlib
+from typing import Any
+
+from pinecone import Pinecone
+from qdrant_client import QdrantClient
+from qdrant_client.models import FieldCondition, Filter, MatchValue, PointIdsList, PointStruct
+
+from app.config.settings import settings
+from app.schemas import VectorItem, VectorSearchResult
+
+_QDRANT_COLLECTION = "adiuva_vectors"
+
+
+def _blob_to_vector(blob: bytes) -> list[float]:
+    """Derive a 32-dim float vector from *blob* for storage purposes only.
+
+    Uses SHA-256 to produce a deterministic 32-byte fingerprint, then
+    normalises each byte to the range [-1.0, 1.0].  This vector carries no
+    semantic meaning on encrypted data.
+    """
+    return [(b - 128) / 128.0 for b in hashlib.sha256(blob).digest()]
+
+
+class VectorStore:
+    """Thin wrapper around Pinecone or Qdrant.
+
+    The backend to use is selected at runtime:
+    - Pinecone: when ``settings.PINECONE_API_KEY`` is non-empty.
+    - Qdrant: otherwise (requires ``settings.QDRANT_URL``).
+    """
+
+    def _use_pinecone(self) -> bool:
+        return bool(settings.PINECONE_API_KEY)
+
+    # ── Pinecone helpers ──────────────────────────────────────────────
+
+    def _pinecone_index(self) -> Any:
+        pc = Pinecone(api_key=settings.PINECONE_API_KEY)
+        return pc.Index(settings.PINECONE_INDEX)
+
+    # ── Qdrant helpers ────────────────────────────────────────────────
+
+    def _qdrant_client(self) -> Any:
+        return QdrantClient(
+            url=settings.QDRANT_URL,
+            api_key=settings.QDRANT_API_KEY or None,
+        )
+
+    # ── Public API ────────────────────────────────────────────────────
+
+    async def upsert(self, user_id: str, vectors: list[VectorItem]) -> None:
+        """Store encrypted vectors in the backend.
+
+        Each ``VectorItem.blob`` is base64-encoded and kept in metadata/payload
+        so it can be returned verbatim during search.
+
+        Args:
+            user_id: Used as Pinecone namespace or Qdrant payload field.
+            vectors: List of encrypted vector items from the client.
+        """
+        if self._use_pinecone():
+            await self._pinecone_upsert(user_id, vectors)
+        else:
+            await self._qdrant_upsert(user_id, vectors)
+
+    async def search(
+        self,
+        user_id: str,
+        query_blob: bytes,
+        top_k: int,
+    ) -> list[VectorSearchResult]:
+        """Query the vector store and return encrypted result blobs.
+
+        The query vector is derived from *query_blob* using the same
+        deterministic mapping as upsert.
+
+        Args:
+            user_id:    Scopes the search to this user's namespace.
+            query_blob: Encrypted query from the client.
+            top_k:      Maximum number of results to return.
+
+        Returns:
+            List of ``VectorSearchResult`` with ``id``, ``score``, and ``blob``.
+        """
+        if self._use_pinecone():
+            return await self._pinecone_search(user_id, query_blob, top_k)
+        return await self._qdrant_search(user_id, query_blob, top_k)
+
+    async def delete(self, user_id: str, vector_ids: list[str]) -> None:
+        """Remove vectors by ID, scoped to *user_id*.
+
+        Args:
+            user_id:    Namespace / payload filter to prevent cross-user deletion.
+            vector_ids: List of vector IDs to remove.
+        """
+        if self._use_pinecone():
+            await self._pinecone_delete(user_id, vector_ids)
+        else:
+            await self._qdrant_delete(user_id, vector_ids)
+
+    # ── Pinecone implementation ───────────────────────────────────────
+
+    async def _pinecone_upsert(self, user_id: str, vectors: list[VectorItem]) -> None:
+        index = self._pinecone_index()
+        records = [
+            {
+                "id": v.id,
+                "values": _blob_to_vector(v.blob),
+                "metadata": {
+                    "blob": base64.b64encode(v.blob).decode(),
+                    "checksum": v.checksum,
+                    "user_id": user_id,
+                },
+            }
+            for v in vectors
+        ]
+        index.upsert(vectors=records, namespace=user_id)
+
+    async def _pinecone_search(
+        self, user_id: str, query_blob: bytes, top_k: int
+    ) -> list[VectorSearchResult]:
+        index = self._pinecone_index()
+        query_vector = _blob_to_vector(query_blob)
+        response = index.query(
+            vector=query_vector,
+            top_k=top_k,
+            namespace=user_id,
+            include_metadata=True,
+        )
+        results: list[VectorSearchResult] = []
+        for match in response.get("matches", []):
+            blob_bytes = base64.b64decode(match["metadata"]["blob"])
+            results.append(
+                VectorSearchResult(
+                    id=match["id"],
+                    score=match["score"],
+                    blob=blob_bytes,
+                )
+            )
+        return results
+
+    async def _pinecone_delete(self, user_id: str, vector_ids: list[str]) -> None:
+        index = self._pinecone_index()
+        index.delete(ids=vector_ids, namespace=user_id)
+
+    # ── Qdrant implementation ─────────────────────────────────────────
+
+    async def _qdrant_upsert(self, user_id: str, vectors: list[VectorItem]) -> None:
+        client = self._qdrant_client()
+        points = [
+            PointStruct(
+                id=v.id,
+                vector=_blob_to_vector(v.blob),
+                payload={
+                    "blob": base64.b64encode(v.blob).decode(),
+                    "checksum": v.checksum,
+                    "user_id": user_id,
+                },
+            )
+            for v in vectors
+        ]
+        client.upsert(collection_name=_QDRANT_COLLECTION, points=points)
+
+    async def _qdrant_search(
+        self, user_id: str, query_blob: bytes, top_k: int
+    ) -> list[VectorSearchResult]:
+        client = self._qdrant_client()
+        query_vector = _blob_to_vector(query_blob)
+        hits = client.search(
+            collection_name=_QDRANT_COLLECTION,
+            query_vector=query_vector,
+            query_filter=Filter(
+                must=[FieldCondition(key="user_id", match=MatchValue(value=user_id))]
+            ),
+            limit=top_k,
+        )
+        return [
+            VectorSearchResult(
+                id=str(hit.id),
+                score=hit.score,
+                blob=base64.b64decode(hit.payload["blob"]),
+            )
+            for hit in hits
+        ]
+
+    async def _qdrant_delete(self, user_id: str, vector_ids: list[str]) -> None:
+        client = self._qdrant_client()
+        client.delete(
+            collection_name=_QDRANT_COLLECTION,
+            points_selector=PointIdsList(points=vector_ids),
+        )

docker-compose.yml

@@ -7,7 +7,7 @@ services:
       - path: .env
         required: false
     environment:
-      DATABASE_URL: postgresql+asyncpg://postgres:postgres@db:5432/adiuvai
+      DATABASE_URL: postgresql+asyncpg://postgres:postgres@db:5432/adiuva
       GITHUB_COPILOT_TOKEN_DIR: /root/.config/litellm/github_copilot
     volumes:
       - copilot_tokens:/root/.config/litellm/github_copilot
@@ -21,7 +21,7 @@ services:
     environment:
       POSTGRES_USER: postgres
       POSTGRES_PASSWORD: postgres
-      POSTGRES_DB: adiuvai
+      POSTGRES_DB: adiuva
     volumes:
       - postgres_data:/var/lib/postgresql/data
     healthcheck:
@@ -36,6 +36,37 @@ services:
   #   image: redis:7-alpine
   #   restart: unless-stopped
 
+  # ── Local S3-compatible storage (MinIO) ──
+  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
+
+  # ── Local vector store (Qdrant) ──
+  qdrant:
+    image: qdrant/qdrant:latest
+    ports:
+      - "6333:6333"
+      - "6334:6334"
+    volumes:
+      - qdrant_data:/qdrant/storage
+    restart: unless-stopped
+
 volumes:
   postgres_data:
+  minio_data:
+  qdrant_data:
   copilot_tokens:

docs/MICROSERVICES_ARCHITECTURE.md

@@ -0,0 +1,941 @@
+# 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

@@ -32,8 +32,6 @@ google-auth-oauthlib>=1.2.0
 google-auth-httplib2>=0.2.0
 msal>=1.28.0
 cryptography>=42.0.0
-langfuse>=2.0.0
+redis>=5.0.0
-beautifulsoup4>=4.12.0
+langfuse>=3.0.0
-lxml>=5.0.0
-PyYAML>=6.0.0
 ruff>=0.8.0

services/auth/.env.example

@@ -0,0 +1,19 @@
+# ── 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

@@ -0,0 +1,36 @@
+# ── 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

@@ -0,0 +1,16 @@
+# 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

@@ -0,0 +1,34 @@
+"""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

@@ -0,0 +1,69 @@
+"""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

@@ -0,0 +1,62 @@
+"""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

@@ -0,0 +1,249 @@
+"""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

@@ -0,0 +1,66 @@
+"""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

@@ -0,0 +1,11 @@
+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

services/batch-agent/Dockerfile

@@ -0,0 +1,36 @@
+# ── builder ──────────────────────────────────────────────────────────────────
+FROM python:3.12-slim AS builder
+
+WORKDIR /build
+
+COPY services/batch-agent/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
+
+# Shared module
+COPY shared/ shared/
+
+# Service source
+COPY services/batch-agent/app/ app/
+
+RUN chown -R appuser:appgroup /app
+
+USER appuser
+
+EXPOSE 8000
+
+# Batch runs are long-lived — use a longer timeout than chat (300s vs 120s)
+CMD ["gunicorn", "app.main:app", \
+     "-k", "uvicorn.workers.UvicornWorker", \
+     "--bind", "0.0.0.0:8000", \
+     "--workers", "2", \
+     "--timeout", "300"]

services/batch-agent/README.md

@@ -0,0 +1,23 @@
+# Batch Agent Service
+
+Owns: agent_runner, journey builder, filesystem_agent, integrations (Gmail, MS Graph).
+
+## Tables owned
+- `local_agent_configs`
+- `cloud_agent_configs`
+- `agent_run_logs`
+
+## Endpoints
+- `GET /agents/catalog`
+- `POST /agents/can-create`
+- `POST /agents/trigger`
+- `GET /agents/{id}/history`
+
+## Redis channels
+- Subscribe: `batch:request:{user_id}`
+- Publish: `ws:out:{user_id}` (journey replies + tool calls)
+- BRPOP: `tool:result:{call_id}` (30s timeout)
+- SET+EX: `journey:{user_id}` (session state, TTL 1800s)
+
+## TODO
+- [ ] Integrate Langfuse tracing (reuse `services/chat/app/tracing.py` pattern — `trace_span()`, `get_langfuse_callback()`, prompt management). Each batch agent run should create a trace with input/output, link prompts, and pass the LangChain `CallbackHandler` to LLM calls.

services/batch-agent/app/agent_runner.py

@@ -0,0 +1,910 @@
+"""Agent run orchestrator — adapted for Batch Agent Service.
+
+Key changes from monolith app/core/agent_runner.py:
+  - No DeviceConnectionManager — tool calls go through Redis ws_context.
+  - set_current_user / clear_current_user replace set_client_executor.
+  - run_local_agent accepts a serialized dict (from Redis / REST) instead
+    of SQLAlchemy model objects.
+  - _finalize_run writes to PostgreSQL via shared.db.async_session.
+  - Cloud agent import path changed to app.integrations.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import uuid
+from datetime import datetime, timedelta, timezone
+from typing import Any
+
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
+from sqlalchemy import select
+
+from app.agents.filesystem_agent import FILESYSTEM_TOOLS
+from shared.agents.note_agent import NOTE_TOOLS
+from shared.agents.project_agent import PROJECT_TOOLS
+from shared.agents.task_agent import TASK_TOOLS
+from shared.agents.timeline_agent import TIMELINE_TOOLS
+from shared.llm import get_llm
+from shared.ws_context import execute_on_client, set_current_user, clear_current_user
+import app.tracing as tracing
+from shared.db import async_session
+from shared.models import AgentRunLog, CloudAgentConfig, LocalAgentConfig
+from shared.redis import redis_client, ws_out_channel
+
+logger = logging.getLogger(__name__)
+
+# ── Concurrency guard ─────────────────────────────────────────────────────
+_running_agents: set[str] = set()
+
+
+def is_agent_running(agent_id: str) -> bool:
+    return agent_id in _running_agents
+
+
+# ── Timeouts ───────────────────────────────────────────────────────────────
+_TOOL_CALL_TIMEOUT: int = 30
+_MAX_PROCESSING_STEPS: int = 12
+_MAX_SCAN_DEPTH: int = 5
+
+# ── Data-type to tool mapping ─────────────────────────────────────────────
+_DATA_TYPE_TOOLS: dict[str, list[Any]] = {
+    "tasks": TASK_TOOLS,
+    "notes": NOTE_TOOLS,
+    "timelines": TIMELINE_TOOLS,
+}
+
+# ── Step 1: Classification prompt ─────────────────────────────────────────
+
+_DOMAIN_DESCRIPTIONS: dict[str, str] = {
+    "tasks": (
+        "Action items, to-dos, deliverables — anything that describes work to be done, "
+        "assigned to someone, or tracked with a due date or status."
+    ),
+    "notes": (
+        "Documentation, meeting notes, summaries, reference material — "
+        "written content meant to be read and referenced rather than acted on."
+    ),
+    "timelines": (
+        "Project milestones, deadlines, scheduled events — "
+        "specific dates that mark a point in the progress of a project."
+    ),
+    "projects": (
+        "High-level project entities — only relevant if the file clearly introduces "
+        "a new project or updates the scope of an existing one."
+    ),
+}
+
+_STEP1_SYSTEM_PROMPT = """\
+You are a file classifier for a freelance project management tool.
+
+Your job is to match a file to an existing project and identify which data domains to extract.
+
+## Project matching rules (STRICT — follow in order)
+
+1. Search the file content for any mention of a project name, client name, acronym, or topic
+   that overlaps with the existing projects listed below.
+2. The match does NOT need to be exact — partial name, abbreviation, or topic similarity is enough.
+3. STRONGLY PREFER matching an existing project. Only return "new" as an absolute last resort
+   when the file has zero meaningful connection to any listed project.
+4. When in doubt, pick the closest match from the list.
+
+## Response format
+
+Respond ONLY with a JSON object — no markdown, no explanation:
+
+{{"project_id": "<exact id from the list below, or new>", "new_project_name": "<concise 2-5 word name, only when project_id is new>", "domains": ["tasks", "notes"]}}
+
+## Domain definitions (only consider domains in the allowed list)
+
+{domain_definitions}
+
+## Existing projects
+
+{projects_list}
+"""
+
+# ── Step 2: Processing prompt ─────────────────────────────────────────────
+
+_PROCESSING_SYSTEM_PROMPT = """\
+You are a data extraction assistant for a freelance project management tool.
+
+Your task: extract structured data from the file content and persist it using the available tools.
+
+## Mandatory process — follow this order for EVERY item you extract
+
+1. READ the existing records listed below for the relevant domain.
+2. SEARCH for a match by title, topic, or semantic similarity.
+3. If a match exists → call the update_* tool with the existing record's id.
+4. If no match exists → call the create_* tool and set isAiSuggested=1.
+
+NEVER call create_* without first checking the existing records.
+NEVER duplicate a record that already exists under a different wording.
+
+## Existing records (source of truth)
+
+{existing_context}
+
+## Context
+
+Project: {project_context}
+Domains to extract: {data_types}
+
+{custom_prompt_section}
+"""
+
+# ── Cloud processing prompt ───────────────────────────────────────────────
+
+_CLOUD_PROCESSING_PROMPT = """\
+You are a data extraction and management assistant for a freelance project
+management tool.
+
+Available tools:
+  Filesystem : read_file_content, list_directory, get_file_metadata
+  Tasks      : list_tasks, create_task, update_task, add_task_comment
+  Notes      : list_notes, get_note, create_note, update_note
+  Timelines  : list_timelines, create_timeline, update_timeline
+  Projects   : list_all_projects, get_project, create_project, update_project
+
+Your task:
+1. Read the full content of each file below using read_file_content.
+2. For each piece of information found, ALWAYS try to match and update an
+   existing record before creating a new one.
+3. ONLY act on these entity types: {data_types}.
+4. Do NOT invent data. Only extract what is clearly present in the files.
+5. If a file contains no relevant data for the target entity types, skip it.
+
+{project_context}
+
+Files to process:
+{file_list}
+
+{custom_prompt_section}
+
+After processing all files, respond with a brief summary of what you updated
+and what you created.
+"""
+
+
+# ── LLM tool-calling loop ─────────────────────────────────────────────────
+
+
+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 _run_agent_with_tools(
+    *,
+    system_prompt: str,
+    user_message: str,
+    tools: list[Any],
+    max_steps: int,
+    langfuse_handler: Any | None = None,
+) -> str:
+    """Run an LLM agent with tool-calling, returning the final text response."""
+    callbacks = [langfuse_handler] if langfuse_handler else None
+    llm = get_llm(callbacks=callbacks)
+    llm_with_tools = llm.bind_tools(tools)
+    messages: list[Any] = [
+        SystemMessage(content=system_prompt),
+        HumanMessage(content=user_message),
+    ]
+
+    tool_map = {tool_def.name: tool_def for tool_def in tools}
+
+    for _ in range(max_steps):
+        response: AIMessage = await llm_with_tools.ainvoke(messages)
+        messages.append(response)
+
+        if not response.tool_calls:
+            return _as_text(response.content)
+
+        for call in response.tool_calls:
+            call_id = str(call.get("id", ""))
+            call_name = str(call.get("name", ""))
+            call_args = call.get("args", {})
+            logger.info(
+                "agent_runner: tool_call name=%s args=%s",
+                call_name,
+                json.dumps(call_args, ensure_ascii=True)[:800],
+            )
+
+            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_runner: tool_result name=%s output=%s",
+                call_name,
+                str(tool_output)[:200],
+            )
+            messages.append(ToolMessage(content=str(tool_output), tool_call_id=call["id"]))
+
+    final = await llm.ainvoke(messages)
+    return _as_text(final.content)
+
+
+# ── Tool list builder ─────────────────────────────────────────────────────
+
+
+def _build_processing_tools(data_types: list[str]) -> list[Any]:
+    tools: list[Any] = list(FILESYSTEM_TOOLS)
+    for dt in data_types:
+        dt_tools = _DATA_TYPE_TOOLS.get(dt)
+        if dt_tools:
+            tools.extend(dt_tools)
+    return tools
+
+
+# ── Code-based directory scanner ─────────────────────────────────────────
+
+
+async def _scan_directories(
+    paths: list[str],
+    extensions: list[str],
+    last_run_at: datetime | None,
+) -> list[str]:
+    all_files: list[str] = []
+    ext_set = {e.lstrip(".").lower() for e in extensions} if extensions else set()
+
+    async def _walk(path: str, depth: int) -> None:
+        if depth > _MAX_SCAN_DEPTH:
+            return
+        try:
+            result = await execute_on_client(action="list_directory", data={"path": path})
+        except Exception as exc:
+            logger.warning("agent_runner: list_directory failed %r: %s", path, exc)
+            return
+        for entry in result.get("entries", []):
+            entry_path = entry.get("path", "")
+            if not entry_path:
+                continue
+            if entry.get("type") == "directory":
+                await _walk(entry_path, depth + 1)
+            elif entry.get("type") == "file":
+                if ext_set:
+                    dot_pos = entry_path.rfind(".")
+                    file_ext = entry_path[dot_pos + 1:].lower() if dot_pos != -1 else ""
+                    if file_ext not in ext_set:
+                        continue
+                all_files.append(entry_path)
+
+    for root in paths:
+        await _walk(root, depth=0)
+
+    if last_run_at is None:
+        return all_files
+
+    last_run_ms = int(last_run_at.timestamp() * 1000)
+    filtered: list[str] = []
+    for file_path in all_files:
+        try:
+            meta = await execute_on_client(action="get_file_metadata", data={"path": file_path})
+            modified_at = meta.get("modifiedAt")
+            if modified_at is None:
+                filtered.append(file_path)
+                continue
+            if isinstance(modified_at, (int, float)):
+                mod_ms = int(modified_at)
+            else:
+                mod_ms = int(datetime.fromisoformat(str(modified_at)).timestamp() * 1000)
+            if mod_ms > last_run_ms:
+                filtered.append(file_path)
+        except Exception:
+            filtered.append(file_path)
+
+    return filtered
+
+
+# ── Code-based entity fetchers ────────────────────────────────────────────
+
+
+async def _fetch_projects() -> list[dict]:
+    try:
+        result = await execute_on_client(action="select", table="projects")
+        return result.get("rows", [])
+    except Exception as exc:
+        logger.warning("agent_runner: failed to fetch projects: %s", exc)
+        return []
+
+
+_DOMAIN_TABLE: dict[str, str] = {
+    "tasks": "tasks",
+    "notes": "notes",
+    "timelines": "timelines",
+    "projects": "projects",
+}
+
+
+async def _fetch_domain_entities(domain: str, project_id: str) -> list[dict]:
+    table = _DOMAIN_TABLE.get(domain)
+    if not table:
+        return []
+    filters: dict[str, Any] = {}
+    if project_id != "standalone" and domain != "projects":
+        filters["projectId"] = project_id
+    try:
+        result = await execute_on_client(
+            action="select",
+            table=table,
+            filters=filters if filters else None,
+        )
+        return result.get("rows", [])
+    except Exception as exc:
+        logger.warning("agent_runner: failed to fetch %s: %s", domain, exc)
+        return []
+
+
+def _format_entities_for_context(domain: str, rows: list[dict]) -> str:
+    if not rows:
+        return f"No existing {domain}."
+    lines: list[str] = []
+    for r in rows:
+        if domain == "tasks":
+            desc = r.get("description") or ""
+            desc_part = f" — {desc[:120]}" if desc else ""
+            assignee = r.get("assignee") or r.get("assignees") or ""
+            due = r.get("dueDate") or r.get("due_date") or ""
+            meta = ", ".join(filter(None, [
+                f"priority: {r.get('priority', '')}" if r.get("priority") else "",
+                f"assignee: {assignee}" if assignee else "",
+                f"due: {due}" if due else "",
+            ]))
+            lines.append(
+                f"  - [{r.get('status', '?')}] {r.get('title', '')}{desc_part}"
+                f" ({meta}, id: {r['id']})"
+            )
+        elif domain == "notes":
+            snippet = (r.get("content") or "")[:200].replace("\n", " ")
+            snippet_part = f"\n      Preview: {snippet}" if snippet else ""
+            lines.append(
+                f"  - {r.get('title', '')} (id: {r['id']}){snippet_part}"
+            )
+        elif domain == "timelines":
+            lines.append(
+                f"  - {r.get('title', '')} date={r.get('date', '')} (id: {r['id']})"
+            )
+        elif domain == "projects":
+            summary = (r.get("aiSummary") or r.get("ai_summary") or "")[:120]
+            summary_part = f" — {summary}" if summary else ""
+            lines.append(
+                f"  - {r.get('name', '')} [{r.get('status', '')}]{summary_part}"
+                f" (id: {r['id']})"
+            )
+    return f"Existing {domain}:\n" + "\n".join(lines)
+
+
+# ── Step 1: LLM file classifier ───────────────────────────────────────────
+
+
+async def _classify_file(
+    file_path: str,
+    file_content: str,
+    projects: list[dict],
+    config_data_types: list[str],
+    langfuse_handler: Any | None = None,
+    custom_system_prompt: str | None = None,
+) -> tuple[str, list[str], str | None]:
+    fallback: tuple[str, list[str], str | None] = ("new", list(config_data_types), None)
+
+    if not file_content.strip():
+        return fallback
+
+    valid_project_ids = {p["id"] for p in projects}
+
+    def _fmt_project(p: dict) -> str:
+        summary = (p.get("aiSummary") or p.get("ai_summary") or "").strip()
+        summary_part = f" — {summary[:100]}" if summary else ""
+        return f"  - id={p['id']} | name={p.get('name', '')} | status={p.get('status', '')}{summary_part}"
+
+    projects_list = "\n".join(_fmt_project(p) for p in projects) or "  (none yet)"
+
+    domain_definitions = "\n".join(
+        f"  - {d}: {_DOMAIN_DESCRIPTIONS[d]}"
+        for d in config_data_types
+        if d in _DOMAIN_DESCRIPTIONS
+    )
+
+    if custom_system_prompt:
+        # Fixture-provided prompt takes absolute priority
+        system = custom_system_prompt.format_map(
+            {"domain_definitions": domain_definitions, "projects_list": projects_list}
+        )
+    else:
+        system = tracing.compile_prompt(
+            "batch_file_classifier",
+            fallback=_STEP1_SYSTEM_PROMPT,
+            variables={
+                "domain_definitions": domain_definitions,
+                "projects_list": projects_list,
+            },
+        )
+
+    llm = get_llm(callbacks=[langfuse_handler] if langfuse_handler else None)
+    try:
+        response = await llm.ainvoke([
+            SystemMessage(content=system),
+            HumanMessage(content=f"File: {file_path}\n\nContent:\n{file_content[:4000]}"),
+        ])
+        raw = _as_text(response.content).strip()
+        if raw.startswith("```"):
+            raw = raw.split("```")[1]
+            if raw.startswith("json"):
+                raw = raw[4:]
+        parsed = json.loads(raw.strip())
+        raw_project_id: str = str(parsed.get("project_id") or "new")
+        project_id = raw_project_id if raw_project_id in valid_project_ids else "new"
+        new_project_name: str | None = (
+            str(parsed["new_project_name"]).strip() or None
+            if project_id == "new" and parsed.get("new_project_name")
+            else None
+        )
+        domains: list[str] = [
+            d for d in parsed.get("domains", [])
+            if d in config_data_types
+        ]
+        if not domains:
+            domains = list(config_data_types)
+        return project_id, domains, new_project_name
+    except Exception as exc:
+        logger.warning(
+            "agent_runner: step1 classification failed for %r: %s", file_path, exc
+        )
+        return fallback
+
+
+# ── Local agent runner (two-step per file) ────────────────────────────────
+
+
+async def run_local_agent(user_id: str, trigger_data: dict[str, Any], *, langfuse_handler: Any | None = None) -> None:
+    """Execute a local directory agent run.
+
+    In the microservice world, trigger_data is a serialized dict from
+    the REST route (forwarded via Redis), containing the agent config
+    fields and run_context.
+
+    set_current_user() must be called BEFORE this function.
+    """
+    run_context: dict = trigger_data.get("run_context", {})
+    agent_id = run_context.get("agent_id", str(uuid.uuid4()))
+    run_id = run_context.get("run_id")
+
+    _running_agents.add(agent_id)
+
+    # Extract config from trigger payload
+    directory_paths: list[str] = trigger_data.get("directory_paths", [])
+    if not directory_paths:
+        directory = trigger_data.get("directory", "")
+        if directory:
+            directory_paths = [directory]
+
+    data_types: list[str] = trigger_data.get("data_types", [])
+    file_extensions: list[str] = trigger_data.get("file_extensions", [])
+    prompt_template: str = trigger_data.get("prompt_template", "")
+    last_run_at_raw = trigger_data.get("last_run_at")
+    last_run_at: datetime | None = None
+    if last_run_at_raw:
+        if isinstance(last_run_at_raw, str):
+            last_run_at = datetime.fromisoformat(last_run_at_raw)
+        elif isinstance(last_run_at_raw, (int, float)):
+            last_run_at = datetime.fromtimestamp(last_run_at_raw / 1000, tz=timezone.utc)
+
+    errors: list[str] = []
+    items_processed = 0
+    items_created = 0
+
+    custom_section = (
+        f"User instructions:\n{prompt_template}"
+        if prompt_template
+        else ""
+    )
+
+    # Create or load run log
+    run_log_id = run_id
+    if not run_log_id:
+        async with async_session() as db:
+            run_log = AgentRunLog(
+                agent_id=agent_id,
+                agent_type="local",
+                user_id=user_id,
+                status="running",
+            )
+            db.add(run_log)
+            await db.commit()
+            await db.refresh(run_log)
+            run_log_id = run_log.id
+
+    try:
+        # ── Scan directories ─────────────────────────────────────────
+        logger.info("agent_runner: run=%s scanning directories user=%s", run_log_id, user_id)
+        file_paths = await _scan_directories(
+            paths=directory_paths,
+            extensions=file_extensions,
+            last_run_at=last_run_at,
+        )
+        logger.info(
+            "agent_runner: run=%s found %d file(s) after filtering", run_log_id, len(file_paths)
+        )
+
+        if not file_paths:
+            await _finalize_run(run_log_id, status="success", items_processed=0, items_created=0)
+            return
+
+        # ── Fetch all projects once ──────────────────────────────────
+        projects = await _fetch_projects()
+
+        for file_path in file_paths:
+            try:
+                file_result = await execute_on_client(
+                    action="read_file_content", data={"path": file_path}
+                )
+                file_content: str = file_result.get("content", "")
+                if not file_content:
+                    continue
+
+                items_processed += 1
+
+                # Step 1 — classify file
+                project_id, domains, new_project_name = await _classify_file(
+                    file_path=file_path,
+                    file_content=file_content,
+                    projects=projects,
+                    config_data_types=data_types,
+                    langfuse_handler=langfuse_handler,
+                )
+
+                # Step 2 — resolve project_id, fetch entities, process
+                if project_id == "new":
+                    proj_name = new_project_name or "Untitled Project"
+                    try:
+                        proj_result = await execute_on_client(
+                            action="insert",
+                            table="projects",
+                            data={"name": proj_name, "clientId": None},
+                        )
+                        created = proj_result.get("row", {})
+                        effective_project_id = created.get("id", "standalone")
+                        if "id" in created:
+                            projects.append(created)
+                    except Exception as exc:
+                        logger.warning("agent_runner: run=%s create project failed: %s", run_log_id, exc)
+                        effective_project_id = "standalone"
+                        proj_name = "unknown"
+                    project_context = (
+                        f"Project: {proj_name} (id: {effective_project_id}). "
+                        "Always set projectId to this id on every record you create."
+                    )
+                else:
+                    effective_project_id = project_id
+                    proj = next((p for p in projects if p["id"] == project_id), None)
+                    proj_name = proj.get("name", project_id) if proj else project_id
+                    project_context = (
+                        f"Project: {proj_name} (id: {project_id}). "
+                        "Always set projectId to this id on every record you create."
+                    )
+
+                domains = [d for d in domains if d != "projects"]
+
+                existing_blocks: list[str] = []
+                for domain in domains:
+                    rows = await _fetch_domain_entities(domain, effective_project_id)
+                    existing_blocks.append(_format_entities_for_context(domain, rows))
+
+                existing_context = "\n\n".join(existing_blocks)
+
+                system_prompt = tracing.compile_prompt(
+                    "batch_processing",
+                    fallback=_PROCESSING_SYSTEM_PROMPT,
+                    variables={
+                        "existing_context": existing_context,
+                        "project_context": project_context,
+                        "data_types": ", ".join(domains),
+                        "custom_prompt_section": custom_section,
+                    },
+                )
+
+                processing_tools = _build_processing_tools(domains)
+
+                result_text = await _run_agent_with_tools(
+                    system_prompt=system_prompt,
+                    user_message=(
+                        f"Process this file and extract relevant information.\n\n"
+                        f"File: {file_path}\n\nContent:\n{file_content}"
+                    ),
+                    tools=processing_tools,
+                    max_steps=_MAX_PROCESSING_STEPS,
+                    langfuse_handler=langfuse_handler,
+                )
+                logger.info(
+                    "agent_runner: run=%s file=%r result=%s",
+                    run_log_id, file_path, result_text[:200],
+                )
+
+            except Exception as exc:
+                errors.append(f"Error processing '{file_path}': {exc}")
+                logger.error("agent_runner: run=%s file=%r failed: %s", run_log_id, file_path, exc)
+
+    except Exception as exc:
+        errors.append(f"Agent run failed: {exc}")
+        logger.error("agent_runner: run=%s failed: %s", run_log_id, exc)
+    finally:
+        _running_agents.discard(agent_id)
+
+    # ── Finalise ────────────────────────────────────────────────────
+    if errors and items_processed == 0:
+        final_status = "error"
+    elif errors:
+        final_status = "partial"
+    else:
+        final_status = "success"
+
+    await _finalize_run(
+        run_log_id,
+        status=final_status,
+        items_processed=items_processed,
+        items_created=items_created,
+        errors=errors,
+    )
+
+    # Notify Electron that the run is complete via Redis
+    if run_context:
+        try:
+            channel = ws_out_channel(user_id)
+            await redis_client.publish(channel, json.dumps({
+                "type": "run_complete",
+                "run_context": run_context,
+                "status": final_status,
+            }))
+        except Exception as exc:
+            logger.warning("agent_runner: run=%s failed to send run_complete: %s", run_log_id, exc)
+
+
+# ── Cloud agent runner ─────────────────────────────────────────────────────
+
+_CLOUD_DEFAULT_LOOKBACK_DAYS: int = 7
+
+
+async def run_cloud_agent(user_id: str, config_id: str, *, langfuse_handler: Any | None = None) -> None:
+    """Execute a cloud connector agent run.
+
+    Loads the CloudAgentConfig from DB, decrypts OAuth tokens, fetches
+    messages from the provider, and runs LLM extraction.
+
+    set_current_user() must be called BEFORE this function.
+    """
+    from app.integrations import decrypt_token, encrypt_token, get_provider
+
+    async with async_session() as db:
+        result = await db.execute(
+            select(CloudAgentConfig).where(CloudAgentConfig.id == config_id)
+        )
+        config = result.scalar_one_or_none()
+        if config is None:
+            logger.error("agent_runner: cloud config %s not found", config_id)
+            return
+
+        # Create run log
+        run_log = AgentRunLog(
+            agent_id=config.id,
+            agent_type="cloud",
+            user_id=user_id,
+            status="running",
+        )
+        db.add(run_log)
+        await db.commit()
+        await db.refresh(run_log)
+        run_log_id = run_log.id
+
+    # ── Decrypt OAuth token ────────────────────────────────────────
+    if not config.oauth_token_encrypted:
+        await _finalize_run(
+            run_log_id,
+            status="error",
+            errors=[f"No OAuth token stored for cloud agent '{config.name}'"],
+        )
+        return
+
+    try:
+        credentials_info = decrypt_token(config.oauth_token_encrypted)
+    except ValueError as exc:
+        await _finalize_run(
+            run_log_id,
+            status="error",
+            errors=[f"Failed to decrypt OAuth token: {exc}"],
+        )
+        return
+
+    # ── Instantiate provider ──────────────────────────────────────
+    try:
+        provider = get_provider(config.provider, credentials_info)
+    except ValueError as exc:
+        await _finalize_run(run_log_id, status="error", errors=[str(exc)])
+        return
+
+    # ── Fetch messages ────────────────────────────────────────────
+    since: datetime | None = config.last_run_at
+    if since is None:
+        since = datetime.now(timezone.utc) - timedelta(days=_CLOUD_DEFAULT_LOOKBACK_DAYS)
+    if since.tzinfo is None:
+        since = since.replace(tzinfo=timezone.utc)
+
+    errors: list[str] = []
+    items_processed = 0
+
+    try:
+        if config.provider == "gmail":
+            raw_messages = await provider.fetch_messages(
+                filter_config=config.filter_config,
+                since=since,
+            )
+        elif config.provider == "outlook":
+            raw_messages = await provider.fetch_emails(
+                filter_config=config.filter_config,
+                since=since,
+            )
+        elif config.provider == "teams":
+            raw_messages = await provider.fetch_messages(
+                filter_config=config.filter_config,
+                since=since,
+            )
+        else:
+            raw_messages = []
+    except RuntimeError as exc:
+        await _finalize_run(
+            run_log_id,
+            status="error",
+            errors=[f"Provider fetch failed: {exc}"],
+            update_config_last_run=True,
+            config_id=config.id,
+            config_type="cloud",
+        )
+        return
+
+    logger.info(
+        "agent_runner: cloud agent %s fetched %d item(s) from %s",
+        config.id, len(raw_messages), config.provider,
+    )
+
+    # ── Extract + insert via LLM ─────────────────────────────────
+    try:
+        processing_tools = _build_processing_tools(config.data_types)
+        custom_section = (
+            f"User instructions:\n{config.prompt_template}"
+            if config.prompt_template
+            else ""
+        )
+
+        for msg in raw_messages:
+            content_text = msg.as_text
+            if not content_text:
+                continue
+            items_processed += 1
+
+            processing_prompt = tracing.compile_prompt(
+                "batch_cloud_processing",
+                fallback=_CLOUD_PROCESSING_PROMPT,
+                variables={
+                    "data_types": ", ".join(config.data_types),
+                    "project_context": "Determine the appropriate project from the message context.",
+                    "file_list": f"Message from {config.provider} (id: {msg.id})",
+                    "custom_prompt_section": custom_section,
+                },
+            )
+
+            try:
+                await _run_agent_with_tools(
+                    system_prompt=processing_prompt,
+                    user_message=f"Process this message content:\n\n{content_text[:8000]}",
+                    tools=processing_tools,
+                    max_steps=_MAX_PROCESSING_STEPS,
+                    langfuse_handler=langfuse_handler,
+                )
+            except Exception as exc:
+                errors.append(f"LLM processing error for message {msg.id!r}: {exc}")
+    except Exception as exc:
+        errors.append(f"Agent run failed: {exc}")
+
+    # ── Persist refreshed token ───────────────────────────────────
+    refreshed = getattr(provider, "refreshed_credentials", None)
+    if refreshed:
+        try:
+            new_encrypted = encrypt_token(refreshed)
+            async with async_session() as db:
+                cfg_result = await db.execute(
+                    select(CloudAgentConfig).where(CloudAgentConfig.id == config.id)
+                )
+                cfg_row = cfg_result.scalar_one_or_none()
+                if cfg_row:
+                    cfg_row.oauth_token_encrypted = new_encrypted
+                    await db.commit()
+        except Exception as exc:
+            logger.warning("agent_runner: failed to persist refreshed token: %s", exc)
+
+    # ── Finalise ──────────────────────────────────────────────────
+    if errors and items_processed == 0:
+        final_status = "error"
+    elif errors:
+        final_status = "partial"
+    else:
+        final_status = "success"
+
+    await _finalize_run(
+        run_log_id,
+        status=final_status,
+        items_processed=items_processed,
+        items_created=0,
+        errors=errors,
+        update_config_last_run=True,
+        config_id=config.id,
+        config_type="cloud",
+    )
+
+
+# ── Internal helper ─────────────────────────────────────────────────────────
+
+
+async def _finalize_run(
+    run_log_id: int | str,
+    *,
+    status: str,
+    items_processed: int = 0,
+    items_created: int = 0,
+    errors: list[str] | None = None,
+    update_config_last_run: bool = False,
+    config_id: str | None = None,
+    config_type: str | None = None,
+) -> None:
+    """Persist the run outcome and optionally update last_run_at on the config."""
+    now = datetime.now(timezone.utc)
+    try:
+        async with async_session() as db:
+            result = await db.execute(
+                select(AgentRunLog).where(AgentRunLog.id == run_log_id)
+            )
+            managed = result.scalar_one_or_none()
+            if managed is None:
+                logger.warning("agent_runner: run_log %s not found for finalization", run_log_id)
+                return
+
+            managed.status = status
+            managed.items_processed = items_processed
+            managed.items_created = items_created
+            managed.errors = errors or []
+            managed.completed_at = now
+
+            if update_config_last_run and config_id:
+                if config_type == "local":
+                    cfg_result = await db.execute(
+                        select(LocalAgentConfig).where(LocalAgentConfig.id == config_id)
+                    )
+                    cfg = cfg_result.scalar_one_or_none()
+                    if cfg:
+                        cfg.last_run_at = now
+                elif config_type == "cloud":
+                    cfg_result = await db.execute(
+                        select(CloudAgentConfig).where(CloudAgentConfig.id == config_id)
+                    )
+                    cfg = cfg_result.scalar_one_or_none()
+                    if cfg:
+                        cfg.last_run_at = now
+
+            await db.commit()
+    except Exception as exc:
+        logger.error("agent_runner: failed to finalize run_log=%s: %s", run_log_id, exc)

services/batch-agent/app/agents/__init__.py

@@ -0,0 +1 @@
+"""Batch Agent Service domain agents and filesystem tools."""

services/batch-agent/app/agents/filesystem_agent.py

@@ -0,0 +1,83 @@
+"""Filesystem agent — tools for reading local directories and files on Electron.
+
+Adapted for Batch Agent Service: import from app.ws_context.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from langchain_core.tools import tool
+
+from shared.ws_context import execute_on_client
+
+
+@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,
+]

services/batch-agent/app/integrations/__init__.py

@@ -0,0 +1,108 @@
+"""Cloud provider integration utilities.
+
+Adapted for Batch Agent Service: import from shared.config instead of app.config.
+
+Provides:
+  * Shared message dataclasses (EmailMessage, ChatMessage)
+  * get_provider() — factory for Gmail/MS Graph clients
+  * encrypt_token() / decrypt_token() — Fernet-based OAuth token encryption
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import dataclass, field
+from datetime import datetime
+from typing import TYPE_CHECKING
+
+from cryptography.fernet import Fernet, InvalidToken
+
+from shared.config import settings
+
+if TYPE_CHECKING:
+    from app.integrations.gmail import GmailClient
+    from app.integrations.ms_graph import MSGraphClient
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class EmailMessage:
+    id: str
+    subject: str
+    sender: str
+    body_text: str
+    date: datetime
+    labels: list[str] = field(default_factory=list)
+
+    @property
+    def as_text(self) -> str:
+        date_str = self.date.strftime("%Y-%m-%d %H:%M")
+        labels_str = f" [{', '.join(self.labels)}]" if self.labels else ""
+        return (
+            f"From: {self.sender}\n"
+            f"Date: {date_str}{labels_str}\n"
+            f"Subject: {self.subject}\n\n"
+            f"{self.body_text}"
+        )
+
+
+@dataclass
+class ChatMessage:
+    id: str
+    content: str
+    sender: str
+    channel: str | None
+    date: datetime
+
+    @property
+    def as_text(self) -> str:
+        date_str = self.date.strftime("%Y-%m-%d %H:%M")
+        channel_str = f" [channel: {self.channel}]" if self.channel else ""
+        return (
+            f"From: {self.sender}\n"
+            f"Date: {date_str}{channel_str}\n\n"
+            f"{self.content}"
+        )
+
+
+def _get_fernet() -> Fernet:
+    key = settings.OAUTH_ENCRYPTION_KEY
+    if not key:
+        raise RuntimeError(
+            "OAUTH_ENCRYPTION_KEY is not set. "
+            "Generate one with: python -c \"from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())\""
+        )
+    return Fernet(key.encode() if isinstance(key, str) else key)
+
+
+def encrypt_token(token_info: dict) -> str:
+    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")
+    return _get_fernet().encrypt(plaintext).decode("utf-8")
+
+
+def decrypt_token(encrypted: str) -> dict:
+    try:
+        plaintext = _get_fernet().decrypt(encrypted.encode("utf-8"))
+        return json.loads(plaintext)
+    except (InvalidToken, json.JSONDecodeError) as exc:
+        raise ValueError(f"Failed to decrypt OAuth token: {exc}") from exc
+
+
+def get_provider(
+    provider: str,
+    credentials_info: dict,
+) -> "GmailClient | MSGraphClient":
+    if provider == "gmail":
+        from app.integrations.gmail import GmailClient
+        return GmailClient(credentials_info)
+    if provider in {"outlook", "teams"}:
+        from app.integrations.ms_graph import MSGraphClient
+        return MSGraphClient(credentials_info)
+    raise ValueError(
+        f"Unknown cloud provider {provider!r}. "
+        "Supported: 'gmail', 'outlook', 'teams'."
+    )

services/batch-agent/app/integrations/gmail.py

@@ -0,0 +1,252 @@
+"""Gmail API client for cloud agent integration.
+
+Adapted for Batch Agent Service: import from app.integrations instead of
+app.integrations (same relative path within the service).
+"""
+
+from __future__ import annotations
+
+import asyncio
+import base64
+import email
+import html
+import logging
+import re
+from datetime import datetime, timezone
+from typing import Any
+
+from app.integrations import EmailMessage
+
+logger = logging.getLogger(__name__)
+
+_GMAIL_DATE_FMT = "%Y/%m/%d"
+_BODY_TRUNCATE = 8_000
+_MAX_MESSAGES = 200
+
+
+def _build_gmail_query(
+    filter_config: dict[str, Any] | None,
+    since: datetime | None,
+) -> str:
+    parts: list[str] = []
+    cfg = filter_config or {}
+
+    labels: list[str] = cfg.get("labels", [])
+    if labels:
+        if len(labels) == 1:
+            parts.append(f"label:{labels[0]}")
+        else:
+            label_expr = " OR ".join(f"label:{lbl}" for lbl in labels)
+            parts.append(f"({label_expr})")
+
+    senders: list[str] = cfg.get("senders", [])
+    for sender in senders:
+        parts.append(f"from:{sender}")
+
+    date_range: dict = cfg.get("date_range", {})
+    from_str: str | None = date_range.get("from")
+    to_str: str | None = date_range.get("to")
+
+    effective_since: datetime | None = since
+    if from_str:
+        try:
+            cfg_since = datetime.fromisoformat(from_str.replace("Z", "+00:00"))
+            if cfg_since.tzinfo is None:
+                cfg_since = cfg_since.replace(tzinfo=timezone.utc)
+            if effective_since is None or cfg_since > effective_since:
+                effective_since = cfg_since
+        except ValueError:
+            logger.warning("gmail: invalid date_range.from %r — ignoring", from_str)
+
+    if effective_since:
+        parts.append(f"after:{effective_since.strftime(_GMAIL_DATE_FMT)}")
+
+    if to_str:
+        try:
+            to_dt = datetime.fromisoformat(to_str.replace("Z", "+00:00"))
+            parts.append(f"before:{to_dt.strftime(_GMAIL_DATE_FMT)}")
+        except ValueError:
+            logger.warning("gmail: invalid date_range.to %r — ignoring", to_str)
+
+    return " ".join(parts)
+
+
+def _strip_html(raw_html: str) -> str:
+    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:
+    mime_type: str = payload.get("mimeType", "")
+    body: dict = payload.get("body", {})
+    parts: list[dict] = payload.get("parts", [])
+
+    if mime_type == "text/plain":
+        data = body.get("data", "")
+        if data:
+            return base64.urlsafe_b64decode(data + "==").decode("utf-8", errors="replace")
+        return ""
+
+    if mime_type == "text/html":
+        data = body.get("data", "")
+        if data:
+            raw = base64.urlsafe_b64decode(data + "==").decode("utf-8", errors="replace")
+            return _strip_html(raw)
+        return ""
+
+    plain_fallback = ""
+    for part in parts:
+        part_mime = part.get("mimeType", "")
+        if part_mime == "text/plain":
+            return _parse_body(part)
+        if part_mime == "text/html" and not plain_fallback:
+            plain_fallback = _parse_body(part)
+        if part_mime.startswith("multipart/"):
+            nested = _parse_body(part)
+            if nested:
+                return nested
+    return plain_fallback
+
+
+def _parse_date(raw: str) -> datetime:
+    try:
+        parsed = email.utils.parsedate_to_datetime(raw)
+        if parsed.tzinfo is None:
+            parsed = parsed.replace(tzinfo=timezone.utc)
+        return parsed.astimezone(timezone.utc)
+    except Exception:
+        return datetime.now(timezone.utc)
+
+
+class GmailClient:
+    def __init__(self, credentials_info: dict[str, Any]) -> None:
+        from google.oauth2.credentials import Credentials
+
+        self._credentials_info = credentials_info
+        expiry_str: str | None = credentials_info.get("expiry")
+        expiry: datetime | None = None
+        if expiry_str:
+            try:
+                expiry = datetime.fromisoformat(
+                    expiry_str.replace("Z", "+00:00")
+                ).replace(tzinfo=timezone.utc)
+            except ValueError:
+                pass
+
+        self._credentials = Credentials(
+            token=credentials_info.get("token"),
+            refresh_token=credentials_info.get("refresh_token"),
+            token_uri=credentials_info.get("token_uri", "https://oauth2.googleapis.com/token"),
+            client_id=credentials_info.get("client_id"),
+            client_secret=credentials_info.get("client_secret"),
+            scopes=credentials_info.get("scopes"),
+            expiry=expiry,
+        )
+
+    async def fetch_messages(
+        self,
+        filter_config: dict[str, Any] | None = None,
+        since: datetime | None = None,
+    ) -> list[EmailMessage]:
+        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:
+        creds = self._credentials
+        if not creds.valid and creds.expired:
+            return None
+        if creds.token != self._credentials_info.get("token"):
+            result = {
+                "token": creds.token,
+                "refresh_token": creds.refresh_token,
+                "token_uri": creds.token_uri,
+                "client_id": creds.client_id,
+                "client_secret": creds.client_secret,
+                "scopes": list(creds.scopes or []),
+            }
+            if creds.expiry:
+                result["expiry"] = creds.expiry.isoformat()
+            return result
+        return None
+
+    def _fetch_sync(self, query: str) -> list[EmailMessage]:
+        import googleapiclient.discovery
+        import googleapiclient.errors
+        from google.auth.transport.requests import Request
+
+        if self._credentials.expired and self._credentials.refresh_token:
+            try:
+                self._credentials.refresh(Request())
+            except Exception as exc:
+                raise RuntimeError(f"Gmail token refresh failed: {exc}") from exc
+
+        service = googleapiclient.discovery.build(
+            "gmail", "v1", credentials=self._credentials, cache_discovery=False
+        )
+        user_api = service.users()
+
+        ids: list[str] = []
+        page_token: str | None = None
+        while len(ids) < _MAX_MESSAGES:
+            batch_size = min(100, _MAX_MESSAGES - len(ids))
+            kwargs: dict[str, Any] = {
+                "userId": "me",
+                "maxResults": batch_size,
+            }
+            if query:
+                kwargs["q"] = query
+            if page_token:
+                kwargs["pageToken"] = page_token
+
+            try:
+                resp = user_api.messages().list(**kwargs).execute()
+            except googleapiclient.errors.HttpError as exc:
+                raise RuntimeError(f"Gmail messages.list failed: {exc}") from exc
+
+            for msg in resp.get("messages", []):
+                ids.append(msg["id"])
+
+            page_token = resp.get("nextPageToken")
+            if not page_token:
+                break
+
+        if not ids:
+            return []
+
+        logger.info("gmail: fetching %d message(s)", len(ids))
+
+        messages: list[EmailMessage] = []
+        for msg_id in ids:
+            try:
+                msg = user_api.messages().get(
+                    userId="me", id=msg_id, format="full"
+                ).execute()
+
+                headers: dict[str, str] = {
+                    h["name"].lower(): h["value"]
+                    for h in msg.get("payload", {}).get("headers", [])
+                }
+                subject = headers.get("subject", "(no subject)")
+                sender = headers.get("from", "unknown")
+                date_raw = headers.get("date", "")
+                date = _parse_date(date_raw) if date_raw else datetime.now(timezone.utc)
+
+                body_text = _parse_body(msg.get("payload", {}))[:_BODY_TRUNCATE]
+                labels = msg.get("labelIds", [])
+
+                messages.append(EmailMessage(
+                    id=msg_id,
+                    subject=subject,
+                    sender=sender,
+                    body_text=body_text,
+                    date=date,
+                    labels=labels,
+                ))
+            except Exception as exc:
+                logger.warning("gmail: skipping message %s: %s", msg_id, exc)
+
+        logger.info("gmail: returned %d message(s)", len(messages))
+        return messages

services/batch-agent/app/integrations/ms_graph.py

@@ -0,0 +1,266 @@
+"""Microsoft Graph API client for Outlook and Teams.
+
+Adapted for Batch Agent Service: import settings from shared.config.
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+from datetime import datetime, timezone
+from typing import Any
+
+import httpx
+
+from shared.config import settings
+from app.integrations import ChatMessage, EmailMessage
+
+logger = logging.getLogger(__name__)
+
+_GRAPH_BASE = "https://graph.microsoft.com/v1.0"
+
+_MAX_EMAILS = 200
+_MAX_MESSAGES = 200
+_BODY_TRUNCATE = 8_000
+
+
+def _strip_html(raw: str) -> str:
+    no_tags = re.sub(r"<[^>]+>", " ", raw)
+    import html as _html
+    decoded = _html.unescape(no_tags)
+    return re.sub(r"\s+", " ", decoded).strip()
+
+
+def _odata_datetime(dt: datetime) -> str:
+    utc = dt.astimezone(timezone.utc)
+    return utc.strftime("%Y-%m-%dT%H:%M:%SZ")
+
+
+def _build_email_filter(
+    filter_config: dict[str, Any] | None,
+    since: datetime | None,
+) -> str:
+    clauses: list[str] = []
+    cfg = filter_config or {}
+
+    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: dict = cfg.get("date_range", {})
+    from_str: str | None = date_range.get("from")
+
+    effective_since: datetime | None = since
+    if from_str:
+        try:
+            cfg_since = datetime.fromisoformat(from_str.replace("Z", "+00:00"))
+            if cfg_since.tzinfo is None:
+                cfg_since = cfg_since.replace(tzinfo=timezone.utc)
+            if effective_since is None or cfg_since > effective_since:
+                effective_since = cfg_since
+        except ValueError:
+            logger.warning("ms_graph: invalid date_range.from %r — ignoring", from_str)
+
+    if effective_since:
+        clauses.append(f"receivedDateTime ge {_odata_datetime(effective_since)}")
+
+    to_str: str | None = date_range.get("to")
+    if to_str:
+        try:
+            to_dt = datetime.fromisoformat(to_str.replace("Z", "+00:00"))
+            if to_dt.tzinfo is None:
+                to_dt = to_dt.replace(tzinfo=timezone.utc)
+            clauses.append(f"receivedDateTime le {_odata_datetime(to_dt)}")
+        except ValueError:
+            logger.warning("ms_graph: invalid date_range.to %r — ignoring", to_str)
+
+    return " and ".join(clauses)
+
+
+class MSGraphClient:
+    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")
+
+    def _auth_headers(self) -> dict[str, str]:
+        return {"Authorization": f"Bearer {self._access_token}"}
+
+    async def _refresh_access_token(self) -> None:
+        import msal
+
+        app = msal.ConfidentialClientApplication(
+            client_id=settings.MS_CLIENT_ID,
+            client_credential=settings.MS_CLIENT_SECRET,
+            authority=f"https://login.microsoftonline.com/{settings.MS_TENANT_ID}",
+        )
+        scopes: list[str] = self._credentials_info.get("scope", "").split()
+        if not scopes:
+            scopes = ["https://graph.microsoft.com/.default"]
+
+        result = app.acquire_token_by_refresh_token(
+            self._refresh_token,
+            scopes=scopes,
+        )
+        if "access_token" not in result:
+            error = result.get("error_description", result.get("error", "unknown"))
+            raise RuntimeError(f"MS Graph token refresh failed: {error}")
+
+        self._access_token = result["access_token"]
+        if "refresh_token" in result:
+            self._refresh_token = result["refresh_token"]
+            self._credentials_info["refresh_token"] = result["refresh_token"]
+        self._credentials_info["access_token"] = self._access_token
+
+    @property
+    def refreshed_credentials(self) -> dict[str, Any] | None:
+        if self._access_token != self._original_access_token:
+            return {**self._credentials_info, "access_token": self._access_token}
+        return None
+
+    async def _get(
+        self,
+        client: httpx.AsyncClient,
+        url: str,
+        params: dict[str, Any] | None = None,
+        *,
+        retry_on_401: bool = True,
+    ) -> dict[str, Any]:
+        resp = await client.get(url, params=params, headers=self._auth_headers())
+        if resp.status_code == 401 and retry_on_401 and self._refresh_token:
+            await self._refresh_access_token()
+            resp = await client.get(url, params=params, headers=self._auth_headers())
+        if resp.status_code == 429:
+            raise RuntimeError("MS Graph rate limit hit (429). Try again later.")
+        resp.raise_for_status()
+        return resp.json()
+
+    async def fetch_emails(
+        self,
+        filter_config: dict[str, Any] | None = None,
+        since: datetime | None = None,
+    ) -> list[EmailMessage]:
+        odata_filter = _build_email_filter(filter_config, since)
+        params: dict[str, Any] = {
+            "$top": 50,
+            "$select": "id,subject,from,receivedDateTime,body,bodyPreview",
+            "$orderby": "receivedDateTime desc",
+        }
+        if odata_filter:
+            params["$filter"] = odata_filter
+
+        emails: list[EmailMessage] = []
+        url = f"{_GRAPH_BASE}/me/messages"
+
+        async with httpx.AsyncClient(timeout=30.0) as client:
+            while url and len(emails) < _MAX_EMAILS:
+                data = await self._get(client, url, params if url.startswith(_GRAPH_BASE) else None)
+                for item in data.get("value", []):
+                    emails.append(self._parse_email(item))
+                    if len(emails) >= _MAX_EMAILS:
+                        break
+                url = data.get("@odata.nextLink", "")
+                params = {}
+
+        logger.info("ms_graph: fetched %d Outlook email(s)", len(emails))
+        return emails
+
+    async def fetch_messages(
+        self,
+        filter_config: dict[str, Any] | None = None,
+        since: datetime | None = None,
+    ) -> list[ChatMessage]:
+        cfg = filter_config or {}
+        channel_filter: list[str] = [c.lower() for c in cfg.get("channels", [])]
+        params: dict[str, Any] = {"$top": 50}
+        if since:
+            params["$filter"] = f"createdDateTime ge {_odata_datetime(since)}"
+
+        messages: list[ChatMessage] = []
+        url = f"{_GRAPH_BASE}/me/chats/getAllMessages"
+
+        async with httpx.AsyncClient(timeout=30.0) as client:
+            while url and len(messages) < _MAX_MESSAGES:
+                try:
+                    data = await self._get(client, url, params if url.startswith(_GRAPH_BASE) else None)
+                except httpx.HTTPStatusError as exc:
+                    if exc.response.status_code in (403, 404):
+                        logger.warning(
+                            "ms_graph: /me/chats/getAllMessages not available (%d)",
+                            exc.response.status_code,
+                        )
+                        break
+                    raise
+
+                for item in data.get("value", []):
+                    msg = self._parse_teams_message(item)
+                    if channel_filter and msg.channel:
+                        if not any(c in msg.channel.lower() for c in channel_filter):
+                            continue
+                    messages.append(msg)
+                    if len(messages) >= _MAX_MESSAGES:
+                        break
+                url = data.get("@odata.nextLink", "")
+                params = {}
+
+        logger.info("ms_graph: fetched %d Teams message(s)", len(messages))
+        return messages
+
+    @staticmethod
+    def _parse_email(item: dict[str, Any]) -> EmailMessage:
+        subject: str = item.get("subject", "(no subject)") or "(no subject)"
+        sender_block = item.get("from", {}) or {}
+        sender_addr = (
+            (sender_block.get("emailAddress") or {}).get("address", "unknown")
+        )
+        date_str: str = item.get("receivedDateTime", "")
+        try:
+            date = datetime.fromisoformat(date_str.replace("Z", "+00:00"))
+        except Exception:
+            date = datetime.now(timezone.utc)
+
+        body_block = item.get("body", {}) or {}
+        content_type: str = body_block.get("contentType", "text")
+        raw_body: str = body_block.get("content", "")
+        if content_type == "html":
+            body_text = _strip_html(raw_body)
+        else:
+            body_text = raw_body or item.get("bodyPreview", "")
+        body_text = body_text[:_BODY_TRUNCATE]
+
+        return EmailMessage(
+            id=item.get("id", ""),
+            subject=subject,
+            sender=sender_addr,
+            body_text=body_text,
+            date=date,
+        )
+
+    @staticmethod
+    def _parse_teams_message(item: dict[str, Any]) -> ChatMessage:
+        msg_id: str = item.get("id", "")
+        sender_block = (item.get("from") or {}).get("user") or {}
+        sender: str = sender_block.get("displayName", "unknown")
+        channel: str | None = (item.get("channelIdentity") or {}).get("channelId")
+
+        date_str: str = item.get("createdDateTime", "")
+        try:
+            date = datetime.fromisoformat(date_str.replace("Z", "+00:00"))
+        except Exception:
+            date = datetime.now(timezone.utc)
+
+        body_block = item.get("body", {}) or {}
+        content_type: str = body_block.get("contentType", "text")
+        raw_content: str = body_block.get("content", "")
+        content = _strip_html(raw_content) if content_type == "html" else raw_content
+        content = content[:_BODY_TRUNCATE]
+
+        return ChatMessage(
+            id=msg_id,
+            content=content,
+            sender=sender,
+            channel=channel,
+            date=date,
+        )

services/batch-agent/app/journey.py

@@ -0,0 +1,395 @@
+"""Chatbot Journey — guided conversation to build an agent prompt_template.
+
+Adapted for Batch Agent Service: imports from app.agents.filesystem_agent
+and app.llm instead of monolith paths.  Session state is in-memory (could
+be moved to Redis for horizontal scaling in the future).
+
+Journey flow:
+  1. Redis consumer dispatches ``journey_start`` with basic agent config.
+  2. Server creates an in-memory session, runs the setup LLM with
+     file-system tools to explore the directory, returns first question.
+  3. ``journey_message`` frames drive the conversation.
+  4. After 3-5 turns the LLM emits PROMPT_TEMPLATE_START / _END block.
+  5. Server parses the block and returns ``journey_reply`` with ``done=True``.
+"""
+
+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 FILESYSTEM_TOOLS
+from shared.llm import get_llm
+import app.tracing as tracing
+
+logger = logging.getLogger(__name__)
+
+# ── Session TTL ───────────────────────────────────────────────────────────
+
+_SESSION_TTL_SECONDS: int = 1800  # 30 minutes
+
+# Sentinel strings used to delimit the LLM-produced prompt_template.
+_TEMPLATE_START = "PROMPT_TEMPLATE_START"
+_TEMPLATE_END = "PROMPT_TEMPLATE_END"
+
+_MIN_TURNS_BEFORE_NUDGE: int = 3
+_MAX_TURNS: int = 15
+_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 = ""
+    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 builder ─────────────────────────────────────────────────
+
+_SYSTEM_PROMPT_TEMPLATE = """\
+You are a friendly assistant helping a freelancer configure a data-extraction agent.
+Your job is to understand exactly what data the user wants to extract from their
+local directory and produce a concise prompt_template that a separate AI will use
+as its instruction set.
+
+You have access to file-system tools to explore the user's directory:
+- list_directory: to see folder structure
+- read_file_content: to peek at file contents
+- get_file_metadata: to check file info
+
+The user's configured directory is: {directory}
+Target data types: {data_types}
+
+IMPORTANT — project assignment is handled automatically.  You MUST NOT ask the user
+about projects, projectId, or how to link records to projects.  Never include
+projectId logic or project creation instructions in the generated prompt_template.
+
+Start by exploring the directory to understand its structure.  Then ask concise,
+focused questions one at a time.  Cover only the topics relevant to the target
+data types listed above:
+
+  1. Content type and format — confirmed by your exploration.
+  2. For TASKS (if in scope): field mapping for title, status, priority, content,
+       dueDate (where is the date found? what's the fallback when absent?),
+       and assignee (is there a person name to assign?).
+  3. For NOTES when TASKS are also in scope: note vs task distinction —
+       what makes something a note rather than a task?
+  4. For TIMELINES (if in scope): the date source — what marks a milestone or event?
+  5. Exclusions and special handling applicable to the target data types.
+
+Keep asking focused questions until you are at least 90% confident.  Then stop and
+output the final prompt_template immediately, wrapped between these exact markers
+on their own lines:
+
+{template_start}
+<the complete extraction prompt here>
+{template_end}
+
+The prompt_template must be concise (bullet points, ~15–25 lines maximum).
+Specify only:
+  - Scope: what files/content qualify and what entity types to create.
+  - Field mapping rules per entity type (camelCase fields: title, status, priority,
+    dueDate, content, assignee, etc.).
+  - dueDate rule (if tasks in scope): source and fallback behaviour.
+  - Note vs task rule (if both in scope): the criterion that separates them.
+  - Timeline date rule (if timelines in scope): what constitutes a timeline event.
+  - Exclusion/filtering rules.
+  - 2–3 concrete mapping examples based on what you discovered.
+
+{existing_section}Begin by exploring the directory, then ask your first question.\
+"""
+
+
+def _build_system_prompt(
+    directory: str,
+    data_types: list[str],
+    existing_template: str | None = None,
+) -> str:
+    existing_section = (
+        f"\nThe user already has the following prompt_template — refine it based on their answers:\n"
+        f"---\n{existing_template}\n---\n"
+        if existing_template
+        else ""
+    )
+    # Use Langfuse compile_prompt ({{variable}} syntax) with Python .format() fallback
+    return tracing.compile_prompt(
+        "journey_system",
+        fallback=_SYSTEM_PROMPT_TEMPLATE,
+        variables={
+            "directory": directory,
+            "data_types": ", ".join(data_types),
+            "existing_section": existing_section,
+        },
+    )
+
+
+# ── Template extraction ───────────────────────────────────────────────────
+
+
+def _extract_template(text: str) -> str | None:
+    """Return the text between PROMPT_TEMPLATE_START and PROMPT_TEMPLATE_END, or None."""
+    if _TEMPLATE_START not in text or _TEMPLATE_END not in text:
+        return None
+    start_idx = text.index(_TEMPLATE_START) + len(_TEMPLATE_START)
+    end_idx = text.index(_TEMPLATE_END)
+    return text[start_idx:end_idx].strip() or 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],
+    langfuse_handler: Any | None = 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.
+    """
+    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"]))
+
+    callbacks = [langfuse_handler] if langfuse_handler else None
+    llm = get_llm(model=None, temperature=0.4, callbacks=callbacks)
+    llm_with_tools = llm.bind_tools(tools)
+    tool_map = {tool_def.name: tool_def for tool_def in tools}
+
+    for _ in range(_MAX_TOOL_STEPS):
+        response: AIMessage = await llm_with_tools.ainvoke(messages)
+        messages.append(response)
+
+        if not response.tool_calls:
+            return _as_text(response.content)
+
+        for call in response.tool_calls:
+            call_name = str(call.get("name", ""))
+            call_args = call.get("args", {})
+            logger.info(
+                "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(
+                "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 tool steps.
+    final = await llm.ainvoke(messages)
+    return _as_text(final.content)
+
+
+# ── Journey handlers (called from redis_consumer) ────────────────────────
+
+
+async def handle_journey_start(
+    user_id: str,
+    frame: dict[str, Any],
+    *,
+    langfuse_handler: Any | None = None,
+) -> dict[str, Any]:
+    """Handle a ``journey_start`` request.
+
+    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_template = frame.get("existing_template")
+
+    session_id = frame.get("session_id") or str(uuid.uuid4())
+    system_prompt = _build_system_prompt(directory, data_types, existing_template)
+
+    session = JourneySession(
+        session_id=session_id,
+        user_id=user_id,
+        agent_type=agent_type,
+        directory=directory,
+        data_types=data_types,
+        system_prompt=system_prompt,
+    )
+
+    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=list(FILESYSTEM_TOOLS),
+        langfuse_handler=langfuse_handler,
+    )
+
+    session.history.extend(seed_history)
+    session.history.append({"role": "assistant", "content": ai_reply})
+    _sessions[session_id] = session
+
+    logger.info(
+        "journey: session %s started for user %s (directory=%s)",
+        session_id,
+        user_id,
+        directory,
+    )
+
+    prompt_template = _extract_template(ai_reply)
+    done = prompt_template is not None
+
+    display_message = ai_reply
+    if done:
+        display_message = (
+            ai_reply[: ai_reply.index(_TEMPLATE_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,
+        "prompt_template": prompt_template,
+    }
+
+
+async def handle_journey_message(
+    user_id: str,
+    frame: dict[str, Any],
+    *,
+    langfuse_handler: Any | None = None,
+) -> dict[str, Any]:
+    """Handle a ``journey_message`` request.
+
+    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,
+            "prompt_template": None,
+        }
+
+    session.history.append({"role": "user", "content": message})
+
+    ai_reply = await _call_llm_with_tools(
+        system_prompt=session.system_prompt,
+        history=session.history,
+        tools=list(FILESYSTEM_TOOLS),
+        langfuse_handler=langfuse_handler,
+    )
+
+    session.history.append({"role": "assistant", "content": ai_reply})
+
+    prompt_template = _extract_template(ai_reply)
+    done = prompt_template is not None
+
+    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"prompt_template now, wrapped in {_TEMPLATE_START} / {_TEMPLATE_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=list(FILESYSTEM_TOOLS),
+                langfuse_handler=langfuse_handler,
+            )
+            session.history.append({"role": "assistant", "content": nudge_reply})
+
+            prompt_template = _extract_template(nudge_reply)
+            if prompt_template is not None:
+                done = True
+                ai_reply = nudge_reply
+
+    display_message = ai_reply
+    if done:
+        display_message = (
+            ai_reply[: ai_reply.index(_TEMPLATE_START)].strip()
+            if _TEMPLATE_START in ai_reply
+            else "Here is your agent configuration. You can save it or continue refining."
+        )
+        _sessions.pop(session_id, None)
+        logger.info("journey: session %s completed for user %s", session_id, user_id)
+
+    return {
+        "type": "journey_reply",
+        "session_id": session_id,
+        "message": display_message,
+        "done": done,
+        "prompt_template": prompt_template,
+    }

services/batch-agent/app/llm.py

@@ -0,0 +1,76 @@
+"""LLM factory — centralised model instantiation via LiteLLM.
+
+Identical to services/chat/app/llm.py. Uses shared.config.settings.
+"""
+
+from __future__ import annotations
+
+import os
+import warnings
+
+from openai import AsyncOpenAI
+import litellm
+
+from langchain_openai import ChatOpenAI
+from langchain_litellm import ChatLiteLLM
+
+from shared.config import settings
+
+litellm.drop_params = True
+
+warnings.filterwarnings(
+    "ignore",
+    message=r"PydanticSerializationUnexpectedValue\(Expected `ResponseAPIUsage`",
+    category=UserWarning,
+)
+
+
+def _api_key_for_model(model: str) -> str | None:
+    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("github/"):
+        return settings.GITHUB_TOKEN or None
+    if model.startswith("github_copilot/"):
+        return None
+    return settings.OPENAI_API_KEY or None
+
+
+def get_llm(
+    *,
+    model: str | None = None,
+    temperature: float = 0,
+    callbacks: list | None = None,
+) -> ChatOpenAI | ChatLiteLLM:
+    model = model or settings.LLM_MODEL
+
+    if settings.GITHUB_COPILOT_TOKEN_DIR:
+        os.environ.setdefault("GITHUB_COPILOT_TOKEN_DIR", settings.GITHUB_COPILOT_TOKEN_DIR)
+
+    if settings.GITHUB_TOKEN:
+        os.environ.setdefault("GITHUB_TOKEN", settings.GITHUB_TOKEN)
+
+    if "/" in model:
+        return ChatLiteLLM(model=model, temperature=temperature, callbacks=callbacks)
+
+    return ChatOpenAI(
+        model=model,
+        temperature=temperature,
+        api_key=_api_key_for_model(model),
+        callbacks=callbacks,
+    )
+
+
+async def embed(text: str) -> list[float]:
+    model = settings.LLM_EMBED_MODEL
+
+    if model.startswith("github_copilot/") or "/" in model:
+        response = await litellm.aembedding(model=model, input=[text])
+        return response.data[0]["embedding"]
+
+    client = AsyncOpenAI(api_key=settings.OPENAI_API_KEY)
+    response = await client.embeddings.create(model=model, input=text)
+    return response.data[0].embedding

services/batch-agent/app/main.py

@@ -0,0 +1,79 @@
+"""Batch Agent Service — FastAPI application.
+
+Owns: agent_runner (local directory + cloud connectors), journey builder,
+filesystem_agent, integrations (Gmail, MS Graph).
+
+Communicates with WS Gateway via Redis:
+  - Subscribes to  batch:request:{user_id}  (journey_start, journey_message)
+  - Publishes to   ws:out:{user_id}         (journey replies + tool calls)
+  - BRPOP on       tool:result:{call_id}    (tool-call round-trip, 30s timeout)
+  - SET+EX on      journey:{user_id}        (journey session state, TTL 1800s)
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import sys
+from pathlib import Path
+
+# Ensure the repo root is on sys.path so ``shared`` is importable when
+# running locally (in Docker the COPY already places it at /app/shared/).
+_repo_root = str(Path(__file__).resolve().parents[3])
+if _repo_root not in sys.path:
+    sys.path.insert(0, _repo_root)
+
+from contextlib import asynccontextmanager
+from typing import AsyncGenerator
+
+from fastapi import FastAPI
+from fastapi.middleware.cors import CORSMiddleware
+
+from app.redis_consumer import start_consumer
+from app.routes import router
+
+logger = logging.getLogger(__name__)
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
+    # Initialise Langfuse tracing (no-op if keys are missing)
+    from app.tracing import init_langfuse
+    init_langfuse()
+
+    logger.info("batch-agent: starting Redis consumer")
+    task = asyncio.create_task(start_consumer())
+    yield
+    task.cancel()
+    try:
+        await task
+    except asyncio.CancelledError:
+        pass
+
+    from app.tracing import shutdown as shutdown_langfuse
+    shutdown_langfuse()
+
+    from shared.db import engine
+    await engine.dispose()
+
+    from shared.redis import redis_client
+    await redis_client.aclose()
+
+    logger.info("batch-agent: Redis consumer stopped")
+
+
+app = FastAPI(title="Adiuva Batch Agent Service", lifespan=lifespan)
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_methods=["GET", "POST"],
+    allow_headers=["*"],
+)
+
+app.include_router(router)
+
+
+@app.get("/health")
+async def health() -> dict[str, str]:
+    return {"status": "ok", "service": "batch-agent"}

services/batch-agent/app/redis_consumer.py

@@ -0,0 +1,183 @@
+"""Redis consumer for the Batch Agent Service.
+
+Subscribes to  batch:request:*  (pattern) and dispatches:
+  - journey_start   → handle_journey_start
+  - journey_message → handle_journey_message
+  - agent_trigger   → run_local_agent / run_cloud_agent
+
+Results are published back to  ws:out:{user_id}  via Redis.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+from typing import Any
+
+from shared.redis import redis_client, batch_request_channel, ws_out_channel
+
+import app.tracing as tracing
+from shared.ws_context import set_current_user, clear_current_user
+
+logger = logging.getLogger(__name__)
+
+
+async def _publish_to_user(user_id: str, payload: dict[str, Any]) -> None:
+    """Publish a frame to the user's WS outbound channel."""
+    channel = ws_out_channel(user_id)
+    await redis_client.publish(channel, json.dumps(payload))
+
+
+async def _handle_journey_start(user_id: str, data: dict[str, Any]) -> None:
+    """Handle a journey_start request from WS Gateway."""
+    from app.journey import handle_journey_start
+
+    session_id = data.get("session_id", "")
+    set_current_user(user_id)
+    try:
+        with tracing.trace_span(
+            name="journey_start",
+            user_id=user_id,
+            session_id=session_id,
+            input=data.get("directory", ""),
+            metadata={"data_types": data.get("data_types", [])},
+            tags=["journey"],
+        ) as span:
+            langfuse_handler = tracing.get_langfuse_callback()
+            reply = await handle_journey_start(user_id, data, langfuse_handler=langfuse_handler)
+            tracing.link_prompt_to_trace(span, "journey_system")
+            span.update(output=reply.get("message", "")[:500])
+            await _publish_to_user(user_id, reply)
+        tracing.flush()
+    except Exception as exc:
+        logger.error("batch-agent: journey_start failed user=%s: %s", user_id, exc)
+        await _publish_to_user(user_id, {
+            "type": "journey_reply",
+            "session_id": session_id,
+            "message": f"Journey setup failed: {exc}",
+            "done": True,
+            "prompt_template": None,
+        })
+    finally:
+        clear_current_user()
+
+
+async def _handle_journey_message(user_id: str, data: dict[str, Any]) -> None:
+    """Handle a journey_message from WS Gateway."""
+    from app.journey import handle_journey_message
+
+    session_id = data.get("session_id", "")
+    set_current_user(user_id)
+    try:
+        with tracing.trace_span(
+            name="journey_message",
+            user_id=user_id,
+            session_id=session_id,
+            input=data.get("message", "")[:200],
+            tags=["journey"],
+        ) as span:
+            langfuse_handler = tracing.get_langfuse_callback()
+            reply = await handle_journey_message(user_id, data, langfuse_handler=langfuse_handler)
+            tracing.link_prompt_to_trace(span, "journey_system")
+            span.update(output=reply.get("message", "")[:500])
+            await _publish_to_user(user_id, reply)
+        tracing.flush()
+    except Exception as exc:
+        logger.error("batch-agent: journey_message failed user=%s: %s", user_id, exc)
+        await _publish_to_user(user_id, {
+            "type": "journey_reply",
+            "session_id": session_id,
+            "message": f"Journey processing failed: {exc}",
+            "done": True,
+            "prompt_template": None,
+        })
+    finally:
+        clear_current_user()
+
+
+async def _handle_agent_trigger(user_id: str, data: dict[str, Any]) -> None:
+    """Handle an agent_trigger request from the REST route (forwarded via Redis)."""
+    from app.agent_runner import run_local_agent
+
+    run_context = data.get("run_context", {})
+    agent_id = run_context.get("agent_id", "")
+    set_current_user(user_id)
+    try:
+        with tracing.trace_span(
+            name="agent_trigger",
+            user_id=user_id,
+            trace_id=run_context.get("run_id"),
+            input={"agent_id": agent_id, "directory": data.get("directory", "")},
+            metadata={"data_types": data.get("data_types", [])},
+            tags=["batch", "agent_run"],
+        ) as span:
+            langfuse_handler = tracing.get_langfuse_callback()
+            await run_local_agent(user_id, data, langfuse_handler=langfuse_handler)
+            tracing.link_prompt_to_trace(span, "batch_processing")
+            span.update(output={"status": "completed"})
+        tracing.flush()
+    except Exception as exc:
+        logger.error("batch-agent: agent_trigger failed user=%s: %s", user_id, exc)
+        await _publish_to_user(user_id, {
+            "type": "run_complete",
+            "status": "error",
+            "run_context": run_context,
+        })
+    finally:
+        clear_current_user()
+
+
+async def _dispatch(user_id: str, message_data: dict[str, Any]) -> None:
+    """Route a batch request to the correct handler."""
+    msg_type = message_data.get("type", "")
+
+    if msg_type == "journey_start":
+        await _handle_journey_start(user_id, message_data)
+    elif msg_type == "journey_message":
+        await _handle_journey_message(user_id, message_data)
+    elif msg_type == "agent_trigger":
+        await _handle_agent_trigger(user_id, message_data)
+    elif msg_type == "device_online":
+        logger.info("batch-agent: device_online user=%s device=%s", user_id, message_data.get("device_id", "?"))
+    else:
+        logger.warning("batch-agent: unknown message type %r from user=%s", msg_type, user_id)
+
+
+async def start_consumer() -> None:
+    """Subscribe to batch:request:* and dispatch incoming frames."""
+    pubsub = redis_client.pubsub()
+    await pubsub.psubscribe("batch:request:*")
+    logger.info("batch-agent: subscribed to batch:request:*")
+
+    try:
+        async for message in pubsub.listen():
+            if message["type"] != "pmessage":
+                continue
+
+            channel: str = message["channel"]
+            if isinstance(channel, bytes):
+                channel = channel.decode()
+
+            # Extract user_id from channel: batch:request:{user_id}
+            parts = channel.split(":", 2)
+            if len(parts) < 3:
+                continue
+            user_id = parts[2]
+
+            raw = message["data"]
+            if isinstance(raw, bytes):
+                raw = raw.decode()
+
+            try:
+                data = json.loads(raw)
+            except json.JSONDecodeError:
+                logger.warning("batch-agent: invalid JSON on channel %s", channel)
+                continue
+
+            # Dispatch in a separate task to avoid blocking the consumer
+            asyncio.create_task(_dispatch(user_id, data))
+    except asyncio.CancelledError:
+        logger.info("batch-agent: consumer shutting down")
+    finally:
+        await pubsub.punsubscribe("batch:request:*")

services/batch-agent/app/routes.py

@@ -0,0 +1,208 @@
+"""Agent REST routes — catalog, billing checks, trigger.
+
+Adapted for Batch Agent Service: uses shared.db, shared.models, shared.schemas.
+Agent trigger dispatches via Redis to the consumer instead of spawning
+an in-process background task.
+"""
+
+from __future__ import annotations
+
+import json
+import uuid
+from datetime import datetime, timezone
+
+from fastapi import APIRouter, Header, HTTPException, status
+from sqlalchemy import func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from shared.db import async_session
+from shared.models import AgentRunLog
+from shared.redis import redis_client, batch_request_channel
+
+from app.agent_runner import is_agent_running
+
+router = APIRouter(prefix="/agents", tags=["agents"])
+
+# ── Tier feature limits ───────────────────────────────────────────────
+# Mirrors app/billing/tier_manager.py FEATURES dict.
+FEATURES: dict[str, dict] = {
+    "free":  {"batch_active": 1,  "batch_runs_per_day": 3},
+    "pro":   {"batch_active": 5,  "batch_runs_per_day": 20},
+    "power": {"batch_active": 20, "batch_runs_per_day": 100},
+    "team":  {"batch_active": -1, "batch_runs_per_day": -1},
+}
+
+
+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 _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) -> None:
+    limit: int = FEATURES.get(tier, FEATURES["free"])["batch_runs_per_day"]
+    if limit == -1:
+        return
+    today_start = datetime.now(timezone.utc).replace(
+        hour=0, minute=0, second=0, microsecond=0
+    )
+    async with async_session() as db:
+        result = await db.execute(
+            select(func.count(AgentRunLog.id)).where(
+                AgentRunLog.user_id == user_id,
+                AgentRunLog.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.",
+        )
+
+
+# ── Catalog ───────────────────────────────────────────────────────────
+
+@router.get("/catalog")
+async def get_agent_catalog(
+    x_user_id: str = Header(..., alias="X-User-Id"),
+) -> list[dict]:
+    return [
+        {
+            "type": "local_directory",
+            "name": "Local Directory Monitor",
+            "description": "Watches local directories, extracts data from files using AI",
+        },
+        {
+            "type": "gmail",
+            "name": "Gmail Connector",
+            "description": "Scans Gmail inbox, extracts tasks/notes from emails",
+        },
+        {
+            "type": "teams",
+            "name": "Microsoft Teams Connector",
+            "description": "Monitors Teams messages, extracts action items",
+        },
+        {
+            "type": "outlook",
+            "name": "Outlook Connector",
+            "description": "Scans Outlook inbox, extracts tasks/notes",
+        },
+    ]
+
+
+# ── Can-create check ─────────────────────────────────────────────────
+
+@router.post("/can-create")
+async def can_create_agent(
+    body: dict,
+    x_user_id: str = Header(..., alias="X-User-Id"),
+    x_user_tier: str = Header("free", alias="X-User-Tier"),
+) -> dict:
+    active_agents = body.get("active_agents", 0)
+    limit: int = FEATURES.get(x_user_tier, FEATURES["free"])["batch_active"]
+    allowed = limit == -1 or active_agents < limit
+    return {
+        "allowed": allowed,
+        "tier": x_user_tier,
+        "active_agents": active_agents,
+        "limit": limit,
+    }
+
+
+# ── Trigger ──────────────────────────────────────────────────────────
+
+@router.post("/trigger", status_code=status.HTTP_202_ACCEPTED)
+async def trigger_agent_run(
+    body: dict,
+    x_user_id: str = Header(..., alias="X-User-Id"),
+    x_user_tier: str = Header("free", alias="X-User-Tier"),
+) -> dict:
+    """Trigger a local agent run — creates run log and dispatches via Redis."""
+    active_agents = body.get("active_agents", 0)
+    _enforce_agent_limit(x_user_tier, active_agents)
+    await _enforce_run_frequency(x_user_tier, x_user_id)
+
+    stable_agent_id = body.get("agent_id") or str(uuid.uuid4())
+
+    if is_agent_running(stable_agent_id):
+        raise HTTPException(
+            status_code=status.HTTP_409_CONFLICT,
+            detail="Agent is already running.",
+        )
+
+    # Create run log in DB
+    async with async_session() as db:
+        run_log = AgentRunLog(
+            agent_id=stable_agent_id,
+            agent_type="local",
+            user_id=x_user_id,
+            status="running",
+        )
+        db.add(run_log)
+        await db.commit()
+        await db.refresh(run_log)
+        run_log_id = run_log.id
+
+    run_context = {
+        "type": "agent_batch",
+        "run_id": run_log_id,
+        "agent_id": stable_agent_id,
+    }
+
+    # Dispatch to the Redis consumer for processing
+    trigger_data = {
+        "type": "agent_trigger",
+        "directory": body.get("directory", ""),
+        "directory_paths": [body.get("directory", "")] if body.get("directory") else [],
+        "data_types": _to_data_types(body.get("what_to_extract", [])),
+        "file_extensions": body.get("file_extensions", []),
+        "prompt_template": body.get("custom_agent_prompt", ""),
+        "device_id": body.get("device_id", ""),
+        "run_context": run_context,
+    }
+
+    channel = batch_request_channel(x_user_id)
+    await redis_client.publish(channel, json.dumps(trigger_data))
+
+    return {
+        "id": run_log_id,
+        "agent_id": stable_agent_id,
+        "agent_type": "local",
+        "status": "running",
+        "items_processed": 0,
+        "items_created": 0,
+        "errors": [],
+        "started_at": _dt_ms(run_log.started_at),
+        "completed_at": None,
+    }

services/batch-agent/app/tracing.py

@@ -0,0 +1,336 @@
+"""Langfuse tracing & prompt management for the Batch Agent Service (v4 SDK).
+
+Provides:
+- ``init_langfuse()`` — initialise the singleton client at startup
+- ``trace_span()`` — context manager that creates a trace + span
+- ``get_langfuse_callback()`` — LangChain callback handler (auto-inherits trace)
+- ``get_prompt()`` — fetch a managed prompt from Langfuse by name
+- ``flush()`` / ``shutdown()`` — lifecycle management
+
+All functions gracefully degrade to no-ops when Langfuse is not configured,
+so the service works identically with or without observability keys.
+
+Requires ``langfuse >= 3.0.0`` (v4 / "Fast Preview" SDK).
+"""
+
+from __future__ import annotations
+
+import logging
+from contextlib import contextmanager
+from typing import Any
+
+from shared.config import settings
+
+logger = logging.getLogger(__name__)
+
+# ── State ────────────────────────────────────────────────────────────────
+
+_initialised: bool = False
+_disabled: bool = False
+
+
+def _is_configured() -> bool:
+    return bool(settings.LANGFUSE_SECRET_KEY and settings.LANGFUSE_PUBLIC_KEY)
+
+
+def init_langfuse() -> None:
+    """Initialise the Langfuse singleton. Call once at startup."""
+    global _initialised, _disabled
+
+    if _initialised or _disabled:
+        return
+
+    if not _is_configured():
+        _disabled = True
+        logger.info("tracing: Langfuse keys not set — tracing disabled")
+        return
+
+    try:
+        from langfuse import Langfuse
+
+        Langfuse(
+            secret_key=settings.LANGFUSE_SECRET_KEY,
+            public_key=settings.LANGFUSE_PUBLIC_KEY,
+            host=settings.LANGFUSE_HOST,
+        )
+        _initialised = True
+        logger.info("tracing: Langfuse client initialised (host=%s)", settings.LANGFUSE_HOST)
+    except Exception as exc:
+        _disabled = True
+        logger.warning("tracing: failed to initialise Langfuse: %s", exc)
+
+
+def _get_client() -> Any | None:
+    """Return the singleton Langfuse client, or *None* if disabled."""
+    if _disabled:
+        return None
+    if not _initialised:
+        init_langfuse()
+    if _disabled:
+        return None
+    try:
+        from langfuse import get_client
+        return get_client()
+    except Exception:
+        return None
+
+
+# ── Null span (no-op when Langfuse is disabled) ─────────────────────────
+
+
+class _NullSpan:
+    """Drop-in replacement when Langfuse is disabled."""
+
+    def update(self, **_: Any) -> None: ...
+    def set_trace_io(self, **_: Any) -> None: ...
+    def score_trace(self, **_: Any) -> None: ...
+
+
+# ── Trace context manager ───────────────────────────────────────────────
+
+
+@contextmanager
+def trace_span(
+    *,
+    name: str,
+    user_id: str,
+    session_id: str | None = None,
+    trace_id: str | None = None,
+    input: Any = None,
+    metadata: dict[str, Any] | None = None,
+    tags: list[str] | None = None,
+):
+    """Context manager that creates a Langfuse trace/span.
+
+    Yields the span object (or a ``_NullSpan`` if Langfuse is disabled).
+    A ``CallbackHandler`` created inside this block auto-inherits the trace
+    context, so there is no need to pass trace IDs manually.
+    """
+    lf = _get_client()
+    if lf is None:
+        yield _NullSpan()
+        return
+
+    try:
+        from langfuse import Langfuse, propagate_attributes
+
+        trace_ctx: dict[str, str] = {}
+        if trace_id is not None:
+            trace_ctx["trace_id"] = Langfuse.create_trace_id(seed=trace_id)
+
+        with lf.start_as_current_observation(
+            as_type="span",
+            name=name,
+            input=input,
+            metadata=metadata or {},
+            **({"trace_context": trace_ctx} if trace_ctx else {}),
+        ) as span:
+            with propagate_attributes(
+                user_id=user_id,
+                session_id=session_id,
+                tags=tags or [],
+            ):
+                yield span
+    except Exception as exc:
+        logger.warning("tracing: trace_span(%s) failed: %s", name, exc)
+        yield _NullSpan()
+
+
+# ── LangChain callback handler ──────────────────────────────────────────
+
+
+def get_langfuse_callback() -> Any | None:
+    """Return a LangChain ``CallbackHandler`` that auto-inherits the current trace.
+
+    Must be called inside a ``trace_span()`` block for proper linking.
+    Returns *None* when Langfuse is disabled.
+    """
+    if _disabled and not _initialised:
+        return None
+
+    try:
+        from langfuse.langchain import CallbackHandler
+        return CallbackHandler()
+    except Exception as exc:
+        logger.warning("tracing: get_langfuse_callback failed: %s", exc)
+        return None
+
+
+# ── Prompt management ────────────────────────────────────────────────────
+
+
+def get_prompt(
+    name: str,
+    *,
+    version: int | None = None,
+    label: str | None = None,
+    fallback: str | None = None,
+    cache_ttl_seconds: int = 300,
+) -> str | None:
+    """Fetch a managed prompt from Langfuse by name (without variable compilation).
+
+    Returns the raw prompt string, or *fallback* if the prompt is not
+    found or Langfuse is disabled.
+    """
+    lf = _get_client()
+    if lf is None:
+        return fallback
+
+    try:
+        kwargs: dict[str, Any] = {
+            "name": name,
+            "cache_ttl_seconds": cache_ttl_seconds,
+        }
+        if version is not None:
+            kwargs["version"] = version
+        if label is not None:
+            kwargs["label"] = label
+        prompt = lf.get_prompt(**kwargs)
+        return prompt.prompt
+    except Exception as exc:
+        logger.warning("tracing: get_prompt(%s) failed: %s", name, exc)
+        return fallback
+
+
+def compile_prompt(
+    name: str,
+    *,
+    fallback: str,
+    variables: dict[str, str],
+    version: int | None = None,
+    label: str | None = None,
+    cache_ttl_seconds: int = 300,
+) -> str:
+    """Fetch a managed prompt from Langfuse and compile it with ``{{variables}}``.
+
+    If the prompt exists in Langfuse, uses the SDK's ``.compile(**variables)``
+    which replaces ``{{key}}`` placeholders.  If Langfuse is disabled or the
+    prompt is not found, falls back to ``fallback.format(**variables)`` (Python
+    ``{key}`` placeholders).
+
+    This means:
+      - Langfuse prompts use ``{{variable}}`` syntax.
+      - Hardcoded fallback strings use Python ``{variable}`` syntax.
+    """
+    lf = _get_client()
+    if lf is None:
+        return fallback.format(**variables)
+
+    try:
+        kwargs: dict[str, Any] = {
+            "name": name,
+            "cache_ttl_seconds": cache_ttl_seconds,
+        }
+        if version is not None:
+            kwargs["version"] = version
+        if label is not None:
+            kwargs["label"] = label
+        prompt = lf.get_prompt(**kwargs)
+        return prompt.compile(**variables)
+    except Exception as exc:
+        logger.warning("tracing: compile_prompt(%s) failed, using fallback: %s", name, exc)
+        return fallback.format(**variables)
+
+
+def get_prompt_object(
+    name: str,
+    *,
+    version: int | None = None,
+    label: str | None = None,
+    cache_ttl_seconds: int = 300,
+) -> Any | None:
+    """Fetch the raw Langfuse prompt *object* (not the compiled string).
+
+    Returns ``None`` when Langfuse is disabled or the prompt is not found.
+    Use this when you need to pass the prompt to ``start_observation(prompt=...)``
+    for linking the prompt to a trace in the Langfuse UI.
+    """
+    lf = _get_client()
+    if lf is None:
+        return None
+
+    try:
+        kwargs: dict[str, Any] = {
+            "name": name,
+            "cache_ttl_seconds": cache_ttl_seconds,
+        }
+        if version is not None:
+            kwargs["version"] = version
+        if label is not None:
+            kwargs["label"] = label
+        return lf.get_prompt(**kwargs)
+    except Exception as exc:
+        logger.warning("tracing: get_prompt_object(%s) failed: %s", name, exc)
+        return None
+
+
+def link_prompt_to_trace(
+    span: Any,
+    prompt_name: str,
+    *,
+    version: int | None = None,
+    label: str | None = None,
+) -> None:
+    """Link a Langfuse managed prompt to a span/observation.
+
+    Uses the SDK v4 ``prompt=`` parameter so that the prompt version
+    appears linked in the Langfuse UI with metrics tracking.
+    """
+    lf = _get_client()
+    if lf is None or isinstance(span, _NullSpan):
+        return
+
+    try:
+        prompt = get_prompt_object(prompt_name, version=version, label=label)
+        if prompt is not None:
+            span.update(prompt=prompt)
+    except Exception as exc:
+        logger.warning("tracing: link_prompt_to_trace(%s) failed: %s", prompt_name, exc)
+
+
+# ── Scoring helper ───────────────────────────────────────────────────────
+
+
+def score_trace(
+    trace_id: str,
+    name: str,
+    value: float,
+    *,
+    comment: str | None = None,
+) -> None:
+    """Post a score to a trace (e.g. user feedback, latency, quality)."""
+    lf = _get_client()
+    if lf is None:
+        return
+
+    try:
+        lf.create_score(trace_id=trace_id, name=name, value=value, comment=comment)
+    except Exception as exc:
+        logger.warning("tracing: score_trace failed: %s", exc)
+
+
+# ── Shutdown ─────────────────────────────────────────────────────────────
+
+
+def flush() -> None:
+    """Flush pending Langfuse events."""
+    lf = _get_client()
+    if lf is not None:
+        try:
+            lf.flush()
+        except Exception as exc:
+            logger.warning("tracing: flush failed: %s", exc)
+
+
+def shutdown() -> None:
+    """Flush and close the Langfuse client."""
+    global _initialised, _disabled
+    lf = _get_client()
+    if lf is not None:
+        try:
+            lf.flush()
+            lf.shutdown()
+        except Exception as exc:
+            logger.warning("tracing: shutdown failed: %s", exc)
+    _initialised = False
+    _disabled = False

services/batch-agent/eval/__init__.py

@@ -0,0 +1 @@
+"""Batch Agent E2E evaluation harness."""

services/batch-agent/eval/__main__.py

@@ -0,0 +1,5 @@
+"""Allow running the eval package as ``python -m eval``."""
+
+from eval.cli import main
+
+main()

services/batch-agent/eval/cli.py

@@ -0,0 +1,285 @@
+"""CLI entry point for the batch agent evaluation harness.
+
+Usage::
+
+    # From services/batch-agent/:
+    python -m eval run                                # all agent fixtures, default model
+    python -m eval run --fixture=classify-invoices    # single fixture
+    python -m eval run --models=gpt-4o,gpt-5.3-codex  # multiple models
+    python -m eval run --mode=step1                   # only step1 fixtures
+    python -m eval run --no-judge                     # skip LLM judge scoring
+
+    python -m eval interactive                        # interactive journey session
+    python -m eval interactive --fixture=journey-invoice-setup
+    python -m eval interactive --model=gpt-4o
+    python -m eval interactive --judge-model=github_copilot/gpt-4o-mini
+
+    python -m eval list                               # list all fixtures
+    python -m eval sync                               # sync fixtures to Langfuse datasets
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import logging
+import sys
+from pathlib import Path
+
+# Ensure the service root and repo root are in sys.path.
+# Service root must come BEFORE repo root so its ``app/`` package
+# shadows the monolith ``app/`` in the repo root.
+_SERVICE_ROOT = Path(__file__).resolve().parent.parent
+_REPO_ROOT = _SERVICE_ROOT.parent.parent
+_sr = str(_SERVICE_ROOT)
+_rr = str(_REPO_ROOT)
+if _rr not in sys.path:
+    sys.path.insert(0, _rr)
+# Always force service root to position 0 (python -m may have already
+# added CWD further down the list, which loses to repo root).
+if _sr in sys.path:
+    sys.path.remove(_sr)
+sys.path.insert(0, _sr)
+
+from eval.config import discover_fixtures, discover_journey_fixtures
+from eval.runner import run_fixture_eval, print_results
+from eval.interactive import run_interactive
+from eval import langfuse_eval
+
+
+def _setup_logging(verbose: bool) -> None:
+    level = logging.DEBUG if verbose else logging.INFO
+    logging.basicConfig(
+        level=level,
+        format="%(asctime)s %(name)-20s %(levelname)-5s %(message)s",
+        datefmt="%H:%M:%S",
+    )
+    # Quiet noisy libraries
+    for name in ("httpx", "httpcore", "openai", "litellm", "urllib3"):
+        logging.getLogger(name).setLevel(logging.WARNING)
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Batch Agent E2E evaluation harness",
+        prog="python -m eval",
+    )
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    # ── run ───────────────────────────────────────────────────────
+    run_cmd = sub.add_parser("run", help="Run evaluations")
+    run_cmd.add_argument(
+        "--fixture", "-f",
+        help="Run only the named fixture (default: all)",
+    )
+    run_cmd.add_argument(
+        "--models", "-m",
+        default="github_copilot/gpt-5.3-codex",
+        help="Comma-separated list of models to test (default: github_copilot/gpt-5.3-codex)",
+    )
+    run_cmd.add_argument(
+        "--mode",
+        default=None,
+        choices=["step1", "step2", "full"],
+        help="Only run fixtures with this mode (default: all)",
+    )
+    run_cmd.add_argument(
+        "--no-judge",
+        action="store_true",
+        help="Skip LLM-as-judge scoring",
+    )
+    run_cmd.add_argument(
+        "--judge-model",
+        default="gpt-4o",
+        help="Model for LLM judge (default: gpt-4o)",
+    )
+    run_cmd.add_argument(
+        "--fixtures-dir",
+        default=None,
+        help="Path to fixtures directory (default: eval/fixtures/)",
+    )
+    run_cmd.add_argument("-v", "--verbose", action="store_true")
+
+    # ── list ──────────────────────────────────────────────────────
+    list_cmd = sub.add_parser("list", help="List available fixtures")
+    list_cmd.add_argument("--fixtures-dir", default=None)
+    list_cmd.add_argument("-v", "--verbose", action="store_true")
+
+    # ── sync ──────────────────────────────────────────────────────
+    sync_cmd = sub.add_parser("sync", help="Sync fixtures to Langfuse datasets")
+    sync_cmd.add_argument("--fixture", "-f", default=None, help="Sync only the named fixture")
+    sync_cmd.add_argument("--fixtures-dir", default=None)
+    sync_cmd.add_argument("-v", "--verbose", action="store_true")
+
+    # ── interactive ───────────────────────────────────────────────
+    inter_cmd = sub.add_parser("interactive", help="Interactive journey session (human-in-the-loop)")
+    inter_cmd.add_argument(
+        "--fixture", "-f",
+        help="Journey fixture to use (default: pick interactively)",
+    )
+    inter_cmd.add_argument(
+        "--model", "-m",
+        default="github_copilot/gpt-5.3-codex",
+        help="Model for the journey AI (default: github_copilot/gpt-5.3-codex)",
+    )
+    inter_cmd.add_argument(
+        "--judge-model",
+        default="gpt-4o",
+        help="Model for LLM judge (default: gpt-4o)",
+    )
+    inter_cmd.add_argument(
+        "--fixtures-dir",
+        default=None,
+        help="Path to fixtures directory (default: eval/fixtures/)",
+    )
+    inter_cmd.add_argument(
+        "--data-dir",
+        default=None,
+        help="Override sample data directory (e.g. path to private test files not in git)",
+    )
+    inter_cmd.add_argument("-v", "--verbose", action="store_true")
+
+    return parser.parse_args()
+
+
+def _fixtures_dir(arg: str | None) -> Path | None:
+    if arg:
+        return Path(arg)
+    return None
+
+
+async def _cmd_run(args: argparse.Namespace) -> None:
+    fixtures = discover_fixtures(_fixtures_dir(args.fixtures_dir))
+    if not fixtures:
+        print("No fixtures found. Create YAML files in eval/fixtures/.")
+        return
+
+    if args.fixture:
+        fixtures = [f for f in fixtures if f.name == args.fixture]
+        if not fixtures:
+            print(f"Fixture '{args.fixture}' not found.")
+            return
+
+    models = [m.strip() for m in args.models.split(",")]
+
+    all_results = []
+    for fixture in fixtures:
+        if args.mode and fixture.mode != args.mode:
+            continue
+        results = await run_fixture_eval(
+            fixture,
+            models=models,
+            use_llm_judge=not args.no_judge,
+            judge_model=args.judge_model,
+        )
+        all_results.extend(results)
+
+    print_results(all_results)
+
+
+def _cmd_list(args: argparse.Namespace) -> None:
+    fixtures = discover_fixtures(_fixtures_dir(args.fixtures_dir))
+    journey_fixtures = discover_journey_fixtures(_fixtures_dir(args.fixtures_dir))
+
+    if not fixtures and not journey_fixtures:
+        print("No fixtures found.")
+        return
+
+    if fixtures:
+        print(f"\n{'[Agent Fixtures]'}")
+        print(f"{'Name':<30} {'Mode':<6} {'Types':<25} {'Expected'}")
+        print("-" * 90)
+        for f in fixtures:
+            types = ", ".join(f.data_types)
+            n_expected = len(f.expected) + len(f.expected_classification)
+            print(f"{f.name:<30} {f.mode:<6} {types:<25} {n_expected}")
+
+    if journey_fixtures:
+        print(f"\n{'[Journey Fixtures]'}")
+        print(f"{'Name':<30} {'Types':<25} {'Messages':<10} {'Criteria'}")
+        print("-" * 90)
+        for f in journey_fixtures:
+            types = ", ".join(f.data_types)
+            print(f"{f.name:<30} {types:<25} {len(f.user_messages):<10} {len(f.expected_template_criteria)}")
+
+    print()
+
+
+def _cmd_sync(args: argparse.Namespace) -> None:
+    fixtures = discover_fixtures(_fixtures_dir(args.fixtures_dir))
+    journey_fixtures = discover_journey_fixtures(_fixtures_dir(args.fixtures_dir))
+
+    if args.fixture:
+        fixtures = [f for f in fixtures if f.name == args.fixture]
+        journey_fixtures = [f for f in journey_fixtures if f.name == args.fixture]
+
+    if not fixtures and not journey_fixtures:
+        print("No fixtures to sync.")
+        return
+
+    for fixture in fixtures:
+        name = langfuse_eval.sync_fixture_to_dataset(fixture)
+        if name:
+            print(f"Synced: {fixture.name} → {name}")
+        else:
+            print(f"Skipped: {fixture.name} (Langfuse not configured)")
+
+    for fixture in journey_fixtures:
+        name = langfuse_eval.sync_journey_fixture_to_dataset(fixture)
+        if name:
+            print(f"Synced: {fixture.name} → {name}")
+        else:
+            print(f"Skipped: {fixture.name} (Langfuse not configured)")
+
+
+async def _cmd_interactive(args: argparse.Namespace) -> None:
+    journey_fixtures = discover_journey_fixtures(_fixtures_dir(args.fixtures_dir))
+    if not journey_fixtures:
+        print("No journey fixtures found. Create YAML files with type: journey in eval/fixtures/.")
+        return
+
+    if args.fixture:
+        fixtures = [f for f in journey_fixtures if f.name == args.fixture]
+        if not fixtures:
+            print(f"Journey fixture '{args.fixture}' not found.")
+            return
+        fixture = fixtures[0]
+    elif len(journey_fixtures) == 1:
+        fixture = journey_fixtures[0]
+    else:
+        # Let user pick
+        print("\nAvailable journey fixtures:")
+        for i, f in enumerate(journey_fixtures, 1):
+            print(f"  {i}. {f.name} — {f.description[:60]}")
+        print()
+        try:
+            choice = int(input("Pick a fixture number: ").strip()) - 1
+            fixture = journey_fixtures[choice]
+        except (ValueError, IndexError, EOFError, KeyboardInterrupt):
+            print("Invalid choice.")
+            return
+
+    await run_interactive(
+        fixture,
+        model=args.model,
+        judge_model=args.judge_model,
+        data_dir=Path(args.data_dir).resolve() if args.data_dir else None,
+    )
+
+
+def main() -> None:
+    args = _parse_args()
+    _setup_logging(args.verbose)
+
+    if args.command == "run":
+        asyncio.run(_cmd_run(args))
+    elif args.command == "interactive":
+        asyncio.run(_cmd_interactive(args))
+    elif args.command == "list":
+        _cmd_list(args)
+    elif args.command == "sync":
+        _cmd_sync(args)
+
+
+if __name__ == "__main__":
+    main()

services/batch-agent/eval/config.py

@@ -0,0 +1,220 @@
+"""Eval configuration — YAML fixture loader and dataclasses.
+
+Fixtures come in two families:
+
+1. **Agent fixtures** — test the batch agent pipeline.
+   Three modes controlled by ``mode``:
+
+   ``step1``  — classification prompt only.
+   ``step2``  — processing prompt only.
+   ``full``   — both steps in sequence.
+
+2. **Journey fixtures** — test the prompt-template builder conversation
+   (unchanged).
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Literal
+
+import yaml
+
+logger = logging.getLogger(__name__)
+
+EvalMode = Literal["step1", "step2", "full"]
+
+
+@dataclass
+class ExpectedRecord:
+    """A single expected extraction result.
+
+    Only the fields specified are checked — unspecified fields are ignored.
+    """
+
+    table: str  # tasks | notes | timelines | projects
+    fields: dict[str, Any]  # field_name → expected_value
+
+
+@dataclass
+class ExpectedClassification:
+    """Expected output of step-1 classification for one file."""
+
+    file: str  # relative path to the sample file
+    project_id: str  # expected matched project id, or "new"
+    domains: list[str]  # expected domain list
+    new_project_name: str | None = None
+
+
+@dataclass
+class EvalFixture:
+    """A complete test scenario loaded from YAML.
+
+    ``mode`` determines which pipeline steps are exercised:
+
+    - **step1**: only ``_classify_file``
+    - **step2**: only the processing LLM + tool loop
+    - **full**: both steps in sequence (``run_local_agent``)
+    """
+
+    name: str
+    description: str
+    mode: EvalMode
+    directory: str  # relative path to sample files
+    data_types: list[str]
+    file_extensions: list[str]
+    models: list[str]  # if empty, use CLI default
+    fixture_path: Path = field(default_factory=lambda: Path("."))
+
+    # ── Step-1 inputs (classification) ───────────────────────────
+    domain_definitions: str = ""
+    projects_list: list[dict[str, Any]] = field(default_factory=list)
+    custom_step1_prompt: str = ""
+
+    # ── Step-2 inputs (processing) ───────────────────────────────
+    existing_context: str = ""
+    project_context: str = ""
+    custom_prompt_section: str = ""
+
+    # ── Seed records for mock executor ───────────────────────────
+    seed_records: dict[str, list[dict]] = field(default_factory=dict)
+
+    # ── Expected outputs ─────────────────────────────────────────
+    expected_classification: list[ExpectedClassification] = field(default_factory=list)
+    expected: list[ExpectedRecord] = field(default_factory=list)
+
+    @property
+    def fixture_dir(self) -> Path:
+        """Absolute path to the sample files directory."""
+        return self.fixture_path.parent / self.directory
+
+    @classmethod
+    def from_yaml(cls, path: Path) -> "EvalFixture":
+        """Load a fixture from a YAML file."""
+        raw = yaml.safe_load(path.read_text(encoding="utf-8"))
+
+        mode: EvalMode = raw.get("mode", "full")
+
+        # Parse expected records (step2/full)
+        expected: list[ExpectedRecord] = []
+        for table, records in (raw.get("expected") or {}).items():
+            for rec in records:
+                expected.append(ExpectedRecord(table=table, fields=rec))
+
+        # Parse expected classification (step1/full)
+        expected_classification: list[ExpectedClassification] = []
+        for item in raw.get("expected_classification") or []:
+            expected_classification.append(ExpectedClassification(
+                file=item["file"],
+                project_id=item["project_id"],
+                domains=item.get("domains", []),
+                new_project_name=item.get("new_project_name"),
+            ))
+
+        return cls(
+            name=raw["name"],
+            description=raw.get("description", ""),
+            mode=mode,
+            directory=raw.get("directory", "sample_files"),
+            data_types=raw.get("data_types", ["tasks"]),
+            file_extensions=raw.get("file_extensions", []),
+            models=raw.get("models", []),
+            fixture_path=path,
+            # Step-1 inputs
+            domain_definitions=raw.get("domain_definitions", ""),
+            projects_list=raw.get("projects_list", []),
+            # Step-2 inputs
+            existing_context=raw.get("existing_context", ""),
+            project_context=raw.get("project_context", ""),
+            custom_prompt_section=raw.get("custom_prompt_section", ""),
+            # Shared
+            seed_records=raw.get("seed_records", {}),
+            expected_classification=expected_classification,
+            expected=expected,
+        )
+
+
+def discover_fixtures(fixtures_dir: Path | None = None) -> list[EvalFixture]:
+    """Find and load all YAML fixtures in the fixtures directory."""
+    if fixtures_dir is None:
+        fixtures_dir = Path(__file__).parent / "fixtures"
+
+    fixtures: list[EvalFixture] = []
+    if not fixtures_dir.is_dir():
+        logger.warning("eval: fixtures directory not found: %s", fixtures_dir)
+        return fixtures
+
+    for yaml_path in sorted(fixtures_dir.glob("*.yaml")):
+        try:
+            raw = yaml.safe_load(yaml_path.read_text(encoding="utf-8"))
+            if raw.get("type") == "journey":
+                continue  # Skip journey fixtures
+            fixtures.append(EvalFixture.from_yaml(yaml_path))
+            logger.info("eval: loaded fixture %s from %s", fixtures[-1].name, yaml_path.name)
+        except Exception as exc:
+            logger.error("eval: failed to load fixture %s: %s", yaml_path.name, exc)
+
+    return fixtures
+
+
+# ── Journey fixtures ─────────────────────────────────────────────────────
+
+
+@dataclass
+class JourneyFixture:
+    """A journey test scenario — tests the prompt_template builder conversation."""
+
+    name: str
+    description: str
+    directory: str  # relative path to sample files
+    data_types: list[str]
+    expected_template_criteria: list[str]  # what the template should contain/satisfy
+    user_messages: list[str] = field(default_factory=list)  # for automated journey runs (unused in interactive mode)
+    models: list[str] = field(default_factory=list)
+    fixture_path: Path = field(default_factory=lambda: Path("."))
+
+    @property
+    def fixture_dir(self) -> Path:
+        """Absolute path to the sample files directory."""
+        return self.fixture_path.parent / self.directory
+
+    @classmethod
+    def from_yaml(cls, path: Path) -> "JourneyFixture":
+        """Load a journey fixture from a YAML file."""
+        raw = yaml.safe_load(path.read_text(encoding="utf-8"))
+
+        return cls(
+            name=raw["name"],
+            description=raw.get("description", ""),
+            directory=raw.get("directory", "sample_files"),
+            data_types=raw.get("data_types", ["tasks"]),
+            user_messages=raw.get("user_messages", []),
+            expected_template_criteria=raw.get("expected_template_criteria", []),
+            models=raw.get("models", []),
+            fixture_path=path,
+        )
+
+
+def discover_journey_fixtures(fixtures_dir: Path | None = None) -> list[JourneyFixture]:
+    """Find and load all journey YAML fixtures in the fixtures directory."""
+    if fixtures_dir is None:
+        fixtures_dir = Path(__file__).parent / "fixtures"
+
+    fixtures: list[JourneyFixture] = []
+    if not fixtures_dir.is_dir():
+        logger.warning("eval: fixtures directory not found: %s", fixtures_dir)
+        return fixtures
+
+    for yaml_path in sorted(fixtures_dir.glob("*.yaml")):
+        try:
+            raw = yaml.safe_load(yaml_path.read_text(encoding="utf-8"))
+            if raw.get("type") != "journey":
+                continue
+            fixtures.append(JourneyFixture.from_yaml(yaml_path))
+            logger.info("eval: loaded journey fixture %s from %s", fixtures[-1].name, yaml_path.name)
+        except Exception as exc:
+            logger.error("eval: failed to load journey fixture %s: %s", yaml_path.name, exc)
+
+    return fixtures

services/batch-agent/eval/fixtures/classify_invoices.yaml

@@ -0,0 +1,40 @@
+# Fixture: classify-invoices (step1)
+# Tests _STEP1_SYSTEM_PROMPT — file classification and project matching.
+# Verifies that the LLM correctly matches files to existing projects
+# and identifies the right data domains.
+
+name: classify-invoices
+mode: step1
+description: >
+  Test file classification on Italian freelance invoices and meeting notes.
+  Verifies project matching and domain identification.
+
+directory: sample_files/invoices
+data_types: [tasks, notes, timelines]
+file_extensions: [txt, md]
+
+# ── Step-1 prompt variables ──────────────────────────────────────
+domain_definitions: |
+  - tasks: Action items, deliverables, things to do — anything that someone needs to complete.
+  - notes: Meeting summaries, decisions, reference information — permanent knowledge entries.
+  - timelines: Project milestones, deadlines, scheduled events — specific dates that mark a point in the progress of a project.
+
+projects_list:
+  - id: "proj-web-redesign"
+    name: "Redesign Sito Web Corporate"
+    status: "active"
+    aiSummary: "Corporate website redesign for Studio Architettura Bianchi"
+  - id: "proj-ecommerce"
+    name: "E-Commerce FashionStore"
+    status: "active"
+    aiSummary: "Next.js e-commerce platform for FashionStore srl"
+
+# ── Expected classification results ─────────────────────────────
+expected_classification:
+  - file: "sample_files/invoices/fattura_042.txt"
+    project_id: "proj-web-redesign"
+    domains: [tasks, notes, timelines]
+
+  - file: "sample_files/invoices/meeting_ecommerce.md"
+    project_id: "proj-ecommerce"
+    domains: [tasks, notes, timelines]

services/batch-agent/eval/fixtures/full_invoices.yaml

@@ -0,0 +1,108 @@
+# Fixture: full-invoices (full)
+# Tests both _STEP1_SYSTEM_PROMPT and _PROCESSING_SYSTEM_PROMPT in sequence
+# via run_local_agent(). Verifies end-to-end classification + extraction.
+
+name: full-invoices
+mode: full
+description: >
+  End-to-end test: classify Italian invoices/meeting notes into the
+  correct project, then extract tasks, notes, and timeline events.
+
+directory: sample_files/invoices
+data_types: [tasks, notes, timelines]
+file_extensions: [txt, md]
+
+# ── Step-1 prompt variables ──────────────────────────────────────
+domain_definitions: |
+  - tasks: Action items, deliverables, things to do — anything that someone needs to complete.
+  - notes: Meeting summaries, decisions, reference information — permanent knowledge entries.
+  - timelines: Project milestones, deadlines, scheduled events — specific dates that mark a point in the progress of a project.
+
+projects_list:
+  - id: "proj-web-redesign"
+    name: "Redesign Sito Web Corporate"
+    status: "active"
+    aiSummary: "Corporate website redesign for Studio Architettura Bianchi"
+  - id: "proj-ecommerce"
+    name: "E-Commerce FashionStore"
+    status: "active"
+    aiSummary: "Next.js e-commerce platform for FashionStore srl"
+
+# ── Step-2 prompt variables ──────────────────────────────────────
+existing_context: |
+  Existing tasks:
+    (none)
+
+  Existing notes:
+    (none)
+
+  Existing timelines:
+    (none)
+
+project_context: ""
+
+custom_prompt_section: |
+  User instructions:
+  Estrai i dati dai file come segue:
+  - TASK: ogni azione da fare, deliverable, o item con scadenza.
+    Mappa "URGENTE" o "ALTA PRIORITÀ" → priority: high.
+    Mappa "media priorità" → priority: medium.
+    Mappa "bassa priorità" → priority: low.
+    Se un item è marcato come "completato" o [x], impostalo status: done.
+    Altrimenti status: todo.
+  - NOTE: riassunti di meeting, decisioni prese, note tecniche.
+  - TIMELINE: date di scadenza, milestone, meeting futuri.
+  Imposta sempre isAiSuggested=1.
+
+# ── Seed records (pre-existing DB state) ─────────────────────────
+seed_records:
+  projects:
+    - id: "proj-web-redesign"
+      name: "Redesign Sito Web Corporate"
+      status: "active"
+      aiSummary: "Corporate website redesign for Studio Architettura Bianchi"
+    - id: "proj-ecommerce"
+      name: "E-Commerce FashionStore"
+      status: "active"
+      aiSummary: "Next.js e-commerce platform for FashionStore srl"
+  tasks: []
+  notes: []
+  timelines: []
+
+# ── Expected classification (step 1) ─────────────────────────────
+expected_classification:
+  - file: "sample_files/invoices/fattura_042.txt"
+    project_id: "proj-web-redesign"
+    domains: [tasks, notes, timelines]
+
+  - file: "sample_files/invoices/meeting_ecommerce.md"
+    project_id: "proj-ecommerce"
+    domains: [tasks, notes, timelines]
+
+# ── Expected extractions (step 2) ────────────────────────────────
+expected:
+  tasks:
+    - title: "Sviluppo frontend React"
+      priority: "high"
+      status: "todo"
+    - title: "Integrazione API backend"
+      priority: "medium"
+      status: "todo"
+    - title: "Testing cross-browser e fix bug responsive"
+      status: "todo"
+    - title: "Preparare wireframe homepage"
+      priority: "high"
+      status: "todo"
+    - title: "Setup progetto Next.js e configurare CI/CD"
+      priority: "medium"
+      status: "todo"
+    - title: "Ricerca plugin Stripe per gestione abbonamenti"
+      priority: "low"
+      status: "todo"
+
+  notes:
+    - title: "Meeting Kickoff Progetto E-Commerce"
+
+  timelines:
+    - title: "MVP E-Commerce pronto"
+    - title: "Meeting di revisione"

services/batch-agent/eval/fixtures/journey_invoice_setup.yaml

@@ -0,0 +1,28 @@
+# Journey Fixture: journey-invoice-setup
+# Used by `python -m eval interactive` for human-in-the-loop testing
+# of the journey chatbot's prompt-building conversation.
+
+type: journey
+name: journey-invoice-setup
+description: >
+  Interactive test for the journey chatbot — explore a directory of
+  Italian invoices and meeting notes, answer the chatbot's questions,
+  and verify it produces a well-structured prompt_template for data
+  extraction.
+
+directory: sample_files/invoices
+data_types: [tasks, notes, timelines, projects]
+
+# Criteria the generated prompt_template must satisfy
+# Each is scored 0-1 by an LLM judge
+expected_template_criteria:
+  - "Mentions creating tasks from action items and work descriptions"
+  - "Mentions creating notes from meeting summaries"
+  - "Mentions extracting timeline events from deadlines and meeting dates"
+  - "Mentions creating projects from relevant information"
+  - "Sets isAiSuggested=1 on all created records"
+  - "Does NOT include projectId assignment logic"
+  - "Uses camelCase field names (title, status, priority, dueDate, content)"
+
+# Models to test (empty = use CLI --models default)
+models: []

services/batch-agent/eval/fixtures/process_invoices.yaml

@@ -0,0 +1,81 @@
+# Fixture: process-invoices (step2)
+# Tests _PROCESSING_SYSTEM_PROMPT — data extraction & tool calling.
+# The classification step is skipped; prompt variables are injected directly.
+
+name: process-invoices
+mode: step2
+description: >
+  Test data extraction from Italian freelance invoices.
+  Verifies correct record creation via tool calls with the right
+  fields, priorities, and status values.
+
+directory: sample_files/invoices
+data_types: [tasks, notes, timelines]
+file_extensions: [txt, md]
+
+# ── Step-2 prompt variables ──────────────────────────────────────
+existing_context: |
+  Existing tasks:
+    (none)
+
+  Existing notes:
+    (none)
+
+  Existing timelines:
+    (none)
+
+project_context: >
+  Project: Redesign Sito Web Corporate (id: proj-web-redesign).
+  Always set projectId to this id on every record you create.
+
+custom_prompt_section: |
+  User instructions:
+  Estrai i dati dai file come segue:
+  - TASK: ogni azione da fare, deliverable, o item con scadenza.
+    Mappa "URGENTE" o "ALTA PRIORITÀ" → priority: high.
+    Mappa "media priorità" → priority: medium.
+    Mappa "bassa priorità" → priority: low.
+    Se un item è marcato come "completato" o [x], impostalo status: done.
+    Altrimenti status: todo.
+  - NOTE: riassunti di meeting, decisioni prese, note tecniche.
+    Il titolo deve essere descrittivo. Il content deve includere tutti i dettagli.
+  - TIMELINE: date di scadenza, milestone, meeting futuri.
+  Imposta sempre isAiSuggested=1.
+
+# ── Seed records (pre-existing DB state) ─────────────────────────
+seed_records:
+  projects:
+    - id: "proj-web-redesign"
+      name: "Redesign Sito Web Corporate"
+      status: "active"
+  tasks: []
+  notes: []
+  timelines: []
+
+# ── Expected extractions ─────────────────────────────────────────
+expected:
+  tasks:
+    - title: "Sviluppo frontend React"
+      priority: "high"
+      status: "todo"
+    - title: "Integrazione API backend"
+      priority: "medium"
+      status: "todo"
+    - title: "Testing cross-browser e fix bug responsive"
+      status: "todo"
+    - title: "Preparare wireframe homepage"
+      priority: "high"
+      status: "todo"
+    - title: "Setup progetto Next.js e configurare CI/CD"
+      priority: "medium"
+      status: "todo"
+    - title: "Ricerca plugin Stripe per gestione abbonamenti"
+      priority: "low"
+      status: "todo"
+
+  notes:
+    - title: "Meeting Kickoff Progetto E-Commerce"
+
+  timelines:
+    - title: "MVP E-Commerce pronto"
+    - title: "Meeting di revisione"

services/batch-agent/eval/fixtures/sample_files/invoices/fattura_042.txt

@@ -0,0 +1,18 @@
+FATTURA N. 2026-0042
+Data: 15 Marzo 2026
+Cliente: Studio Architettura Bianchi
+
+Progetto: Redesign Sito Web Corporate
+
+Descrizione lavori:
+- Sviluppo frontend React (40 ore) — URGENTE, completare entro 20 marzo
+- Integrazione API backend (20 ore) — priorità media
+- Design UI/UX mockup homepage (8 ore) — completato
+- Testing cross-browser e fix bug responsive (12 ore) — da iniziare
+
+Totale: €4.800,00 + IVA
+
+Note:
+Meeting di revisione previsto per il 18 marzo alle 10:00.
+Il cliente ha richiesto modifiche al layout mobile della sezione contatti.
+Attendere conferma budget aggiuntivo per sezione blog.

services/batch-agent/eval/fixtures/sample_files/invoices/meeting_ecommerce.md

@@ -0,0 +1,25 @@
+# Meeting Notes - Kickoff Progetto E-Commerce
+
+**Data:** 10 Marzo 2026
+**Partecipanti:** Marco R., Giulia T., Cliente (FashionStore srl)
+
+## Decisioni prese
+
+1. **Piattaforma**: Next.js + Stripe per i pagamenti
+2. **Timeline**: MVP pronto entro 30 aprile 2026
+3. **Budget**: €12.000 totale, €4.000 anticipo già ricevuto
+
+## Action items
+
+- [ ] Marco: preparare wireframe homepage entro 14 marzo — ALTA PRIORITÀ
+- [ ] Giulia: setup progetto Next.js e configurare CI/CD — media priorità
+- [ ] Marco: ricerca plugin Stripe per gestione abbonamenti — bassa priorità
+- [x] Giulia: inviare contratto firmato al cliente — COMPLETATO
+
+## Note aggiuntive
+
+Il cliente vuole un design minimalista, ispirato a Zara.com.
+Colori primari: nero, bianco, oro.
+Font: Inter per body, Playfair Display per headings.
+
+Prossimo meeting: 24 marzo 2026 ore 15:00.

services/batch-agent/eval/interactive.py

@@ -0,0 +1,471 @@
+"""Interactive journey session — human-in-the-loop CLI conversation.
+
+Flow:
+1. Show the system prompt used by the journey AI.
+2. Start the journey (AI explores files, asks first question).
+3. User types responses in the terminal — AI replies.
+4. User types `/done` to end the conversation.
+5. User writes a comment about the interaction quality.
+6. LLM judge scores the conversation + generated template.
+7. Results are reported to Langfuse.
+
+Usage::
+
+    python -m eval interactive                        # pick a fixture interactively
+    python -m eval interactive --fixture=journey-invoice-setup
+    python -m eval interactive --model=gpt-4o
+    python -m eval interactive --judge-model=github_copilot/gpt-4o-mini
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import sys
+import time
+import uuid
+from dataclasses import dataclass, field
+from typing import Any
+
+from langchain_core.messages import HumanMessage, SystemMessage
+
+from eval.config import JourneyFixture, discover_journey_fixtures
+from eval.mock_executor import MockExecutor
+from eval import langfuse_eval
+
+logger = logging.getLogger(__name__)
+
+# ── Special commands ─────────────────────────────────────────────────────
+
+_CMD_DONE = "/done"
+_CMD_QUIT = "/quit"
+_CMD_TEMPLATE = "/template"
+_CMD_HELP = "/help"
+
+_HELP_TEXT = f"""\
+  {_CMD_DONE}       — End the conversation and proceed to evaluation
+  {_CMD_QUIT}       — Abort without evaluation
+  {_CMD_TEMPLATE}   — Show the generated template (if any)
+  {_CMD_HELP}       — Show this help"""
+
+# ── Terminal colours (ANSI) ──────────────────────────────────────────────
+
+_C_RESET = "\033[0m"
+_C_BOLD = "\033[1m"
+_C_DIM = "\033[2m"
+_C_CYAN = "\033[36m"
+_C_GREEN = "\033[32m"
+_C_YELLOW = "\033[33m"
+_C_MAGENTA = "\033[35m"
+_C_RED = "\033[31m"
+_C_BLUE = "\033[34m"
+
+
+def _print_header(text: str) -> None:
+    print(f"\n{_C_BOLD}{_C_CYAN}{'═' * 80}")
+    print(f"  {text}")
+    print(f"{'═' * 80}{_C_RESET}\n")
+
+
+def _print_ai(text: str) -> None:
+    print(f"\n{_C_GREEN}{_C_BOLD}AI:{_C_RESET} {text}\n")
+
+
+def _print_system(text: str) -> None:
+    print(f"{_C_DIM}{text}{_C_RESET}")
+
+
+def _print_score(label: str, score: float) -> None:
+    if score >= 0.7:
+        color = _C_GREEN
+        tag = "PASS"
+    elif score >= 0.4:
+        color = _C_YELLOW
+        tag = "PARTIAL"
+    else:
+        color = _C_RED
+        tag = "FAIL"
+    print(f"  {color}{tag:>7}{_C_RESET} ({score:.1f}) {label}")
+
+
+# ── Result type ──────────────────────────────────────────────────────────
+
+
+@dataclass
+class InteractiveResult:
+    fixture_name: str
+    model: str
+    judge_model: str
+    prompt_template: str | None
+    conversation: list[dict[str, str]]
+    user_comment: str
+    done: bool
+    criteria_scores: dict[str, float]
+    overall_score: float
+    judge_reasoning: str
+    elapsed_seconds: float
+
+    def summary(self) -> dict[str, Any]:
+        return {
+            "fixture": self.fixture_name,
+            "model": self.model,
+            "judge_model": self.judge_model,
+            "done": self.done,
+            "turns": len([c for c in self.conversation if c["role"] == "user"]),
+            "overall_score": round(self.overall_score, 3),
+            "user_comment": self.user_comment,
+            "criteria_scores": {k: round(v, 3) for k, v in self.criteria_scores.items()},
+            "elapsed_s": round(self.elapsed_seconds, 1),
+        }
+
+
+# ── LLM judge ────────────────────────────────────────────────────────────
+
+_INTERACTIVE_JUDGE_SYSTEM = """\
+You are an evaluation judge for AI-generated prompt templates produced during
+an interactive conversation between a human and a journey chatbot.
+
+The chatbot explored a directory and through multi-turn conversation with the
+user produced a prompt_template — an instruction set for a data-extraction agent.
+
+You have access to:
+- The full conversation transcript
+- The generated prompt_template (if any)
+- The user's own comment about the interaction
+- A list of quality criteria
+
+Score each criterion from 0 to 1:
+  - 1.0: Fully satisfied
+  - 0.5: Partially satisfied
+  - 0.0: Not satisfied
+
+Also provide an overall_quality score (0-1) evaluating the conversation flow,
+how well the AI understood the user, and the template quality.
+
+Respond with ONLY a JSON object:
+{
+  "criteria_scores": {"criterion_1": 0.8, ...},
+  "overall_quality": 0.85,
+  "reasoning": "Brief explanation covering both conversation quality and template accuracy"
+}
+"""
+
+
+async def _judge_interactive(
+    conversation: list[dict[str, str]],
+    prompt_template: str | None,
+    user_comment: str,
+    criteria: list[str],
+    *,
+    judge_model: str = "gpt-4o-mini",
+) -> tuple[dict[str, float], float, str]:
+    """Score an interactive session. Returns (criteria_scores, overall_quality, reasoning)."""
+    from shared.llm import get_llm
+
+    llm = get_llm(model=judge_model, temperature=0)
+
+    conv_text = "\n".join(
+        f"{'USER' if t['role'] == 'user' else 'AI'}: {t['content']}"
+        for t in conversation
+    )
+    criteria_text = "\n".join(f"  {i+1}. {c}" for i, c in enumerate(criteria))
+
+    user_content = (
+        f"## Conversation transcript\n```\n{conv_text}\n```\n\n"
+        f"## Generated prompt_template\n```\n{prompt_template or '(none — conversation did not complete)'}\n```\n\n"
+        f"## User's comment\n{user_comment}\n\n"
+        f"## Criteria to evaluate\n{criteria_text}"
+    )
+
+    try:
+        response = await llm.ainvoke([
+            SystemMessage(content=_INTERACTIVE_JUDGE_SYSTEM),
+            HumanMessage(content=user_content),
+        ])
+        raw = response.content.strip()
+        if raw.startswith("```"):
+            raw = raw.split("```")[1]
+            if raw.startswith("json"):
+                raw = raw[4:]
+        parsed = json.loads(raw.strip())
+
+        scores_raw = parsed.get("criteria_scores", parsed.get("scores", {}))
+        criteria_scores: dict[str, float] = {}
+        for i, criterion in enumerate(criteria):
+            key_candidates = [f"criterion_{i+1}", criterion, criterion[:50], str(i + 1)]
+            score = 0.0
+            for key in key_candidates:
+                if key in scores_raw:
+                    score = float(scores_raw[key])
+                    break
+            if score == 0.0 and i < len(scores_raw):
+                score = float(list(scores_raw.values())[i])
+            criteria_scores[criterion] = score
+
+        overall = float(parsed.get("overall_quality", 0.0))
+        reasoning = str(parsed.get("reasoning", ""))
+        return criteria_scores, overall, reasoning
+
+    except Exception as exc:
+        logger.warning("interactive judge failed: %s", exc)
+        return {c: 0.0 for c in criteria}, 0.0, f"Judge error: {exc}"
+
+
+# ── Interactive session ──────────────────────────────────────────────────
+
+
+async def run_interactive(
+    fixture: JourneyFixture,
+    *,
+    model: str = "gpt-4o",
+    judge_model: str = "gpt-4o-mini",
+    data_dir: Path | None = None,
+) -> InteractiveResult:
+    """Run an interactive journey session in the terminal.
+
+    Parameters
+    ----------
+    data_dir :
+        If set, overrides the fixture's sample-file directory.  The LLM
+        will explore this folder instead of the default
+        ``fixtures/sample_files/…``.  Useful for private test data that
+        shouldn't be committed to git.
+    """
+    from shared.config import settings
+    from shared.ws_context import set_current_user, clear_current_user
+    from app.journey import (
+        handle_journey_start,
+        handle_journey_message,
+        _build_system_prompt,
+    )
+
+    # When --data-dir is given, the MockExecutor's root becomes
+    # data_dir's parent and the journey directory is data_dir's name.
+    # This way the LLM sees a meaningful directory name (not ".") and
+    # MockExecutor resolves paths correctly.
+    # Otherwise, use the fixture's YAML parent and its relative path.
+    if data_dir:
+        mock_root = data_dir.parent
+        journey_directory = data_dir.name
+    else:
+        mock_root = fixture.fixture_path.parent
+        journey_directory = fixture.directory
+
+    mock = MockExecutor(
+        fixture_dir=mock_root,
+        seed_records={},
+    )
+
+    original_model = settings.LLM_MODEL
+    settings.LLM_MODEL = model
+    eval_user_id = f"interactive-{uuid.uuid4().hex[:8]}"
+
+    # ── Show system prompt ───────────────────────────────────────
+    system_prompt = _build_system_prompt(journey_directory, fixture.data_types)
+
+    _print_header("SYSTEM PROMPT")
+    print(f"{_C_DIM}{system_prompt}{_C_RESET}")
+
+    _print_header(f"INTERACTIVE JOURNEY  |  fixture: {fixture.name}  |  model: {model}")
+    print(f"  Data dir: {mock_root}")
+    print(f"  Type your responses. Commands: {_CMD_DONE}, {_CMD_QUIT}, {_CMD_TEMPLATE}, {_CMD_HELP}")
+    print(f"  Judge model: {judge_model}")
+    print(f"  Criteria: {len(fixture.expected_template_criteria)}")
+    print()
+
+    conversation: list[dict[str, str]] = []
+    prompt_template: str | None = None
+    done = False
+    start_time = time.time()
+
+    try:
+        set_current_user(eval_user_id)
+
+        with mock.patch():
+            # ── Start ────────────────────────────────────────────
+            _print_system("Starting journey... (AI is exploring your files)")
+
+            start_frame: dict[str, Any] = {
+                "agent_type": "local",
+                "directory": journey_directory,
+                "data_types": fixture.data_types,
+                "session_id": f"interactive-{uuid.uuid4().hex[:8]}",
+            }
+
+            reply = await handle_journey_start(eval_user_id, start_frame)
+            session_id = reply["session_id"]
+            conversation.append({"role": "assistant", "content": reply["message"]})
+            _print_ai(reply["message"])
+
+            if reply["done"]:
+                prompt_template = reply.get("prompt_template")
+                done = True
+                _print_system("Journey completed on first reply (template generated).")
+
+            # ── Conversation loop ────────────────────────────────
+            while not done:
+                try:
+                    user_input = input(f"{_C_BOLD}{_C_BLUE}YOU:{_C_RESET} ").strip()
+                except (EOFError, KeyboardInterrupt):
+                    print()
+                    user_input = _CMD_QUIT
+
+                if not user_input:
+                    continue
+
+                # Handle commands
+                if user_input.lower() == _CMD_QUIT:
+                    _print_system("Aborted — no evaluation will be performed.")
+                    settings.LLM_MODEL = original_model
+                    clear_current_user()
+                    return InteractiveResult(
+                        fixture_name=fixture.name, model=model, judge_model=judge_model,
+                        prompt_template=None, conversation=conversation,
+                        user_comment="(aborted)", done=False,
+                        criteria_scores={}, overall_score=0.0,
+                        judge_reasoning="Session aborted by user.",
+                        elapsed_seconds=time.time() - start_time,
+                    )
+
+                if user_input.lower() == _CMD_HELP:
+                    print(_HELP_TEXT)
+                    continue
+
+                if user_input.lower() == _CMD_TEMPLATE:
+                    if prompt_template:
+                        print(f"\n{_C_MAGENTA}{prompt_template}{_C_RESET}\n")
+                    else:
+                        _print_system("No template generated yet.")
+                    continue
+
+                if user_input.lower() == _CMD_DONE:
+                    _print_system("Ending conversation...")
+                    break
+
+                # ── Send message to AI ───────────────────────────
+                conversation.append({"role": "user", "content": user_input})
+                _print_system("AI is thinking...")
+
+                msg_frame: dict[str, Any] = {
+                    "session_id": session_id,
+                    "message": user_input,
+                }
+                reply = await handle_journey_message(eval_user_id, msg_frame)
+                conversation.append({"role": "assistant", "content": reply["message"]})
+                _print_ai(reply["message"])
+
+                if reply["done"]:
+                    prompt_template = reply.get("prompt_template")
+                    done = True
+                    _print_system("Journey completed — template generated!")
+
+    except Exception as exc:
+        logger.error("interactive journey failed: %s", exc)
+        _print_system(f"Error: {exc}")
+    finally:
+        settings.LLM_MODEL = original_model
+        clear_current_user()
+
+    elapsed = time.time() - start_time
+    turns = len([c for c in conversation if c["role"] == "user"])
+
+    # ── Show template if generated ───────────────────────────────
+    if prompt_template:
+        _print_header("GENERATED TEMPLATE")
+        print(f"{_C_MAGENTA}{prompt_template}{_C_RESET}\n")
+    else:
+        _print_system("No template was generated during this session.")
+
+    # ── User comment ─────────────────────────────────────────────
+    _print_header("YOUR EVALUATION")
+    print("  Write your comment about this interaction (press Enter twice to finish):")
+    print()
+    comment_lines: list[str] = []
+    try:
+        while True:
+            line = input()
+            if line == "" and comment_lines and comment_lines[-1] == "":
+                comment_lines.pop()  # remove trailing empty
+                break
+            comment_lines.append(line)
+    except (EOFError, KeyboardInterrupt):
+        pass
+    user_comment = "\n".join(comment_lines).strip() or "(no comment)"
+
+    # ── Judge ────────────────────────────────────────────────────
+    _print_header("LLM JUDGE EVALUATION")
+    _print_system(f"Scoring with {judge_model}...")
+
+    criteria_scores, overall_quality, judge_reasoning = await _judge_interactive(
+        conversation=conversation,
+        prompt_template=prompt_template,
+        user_comment=user_comment,
+        criteria=fixture.expected_template_criteria,
+        judge_model=judge_model,
+    )
+
+    # ── Display scores ───────────────────────────────────────────
+    print()
+    for criterion, score in criteria_scores.items():
+        _print_score(criterion, score)
+
+    overall = (
+        sum(criteria_scores.values()) / len(criteria_scores)
+        if criteria_scores
+        else 0.0
+    )
+
+    print(f"\n  {_C_BOLD}Criteria avg:      {overall:.2f}{_C_RESET}")
+    print(f"  {_C_BOLD}Overall quality:   {overall_quality:.2f}{_C_RESET}")
+    print(f"  {_C_BOLD}Turns:             {turns}{_C_RESET}")
+    print(f"  {_C_BOLD}Time:              {elapsed:.1f}s{_C_RESET}")
+    print(f"\n  {_C_DIM}Judge: {judge_reasoning}{_C_RESET}")
+    print(f"  {_C_DIM}Your comment: {user_comment}{_C_RESET}\n")
+
+    result = InteractiveResult(
+        fixture_name=fixture.name,
+        model=model,
+        judge_model=judge_model,
+        prompt_template=prompt_template,
+        conversation=conversation,
+        user_comment=user_comment,
+        done=done,
+        criteria_scores=criteria_scores,
+        overall_score=overall_quality,
+        judge_reasoning=judge_reasoning,
+        elapsed_seconds=elapsed,
+    )
+
+    # ── Report to Langfuse ───────────────────────────────────────
+    trace_id = langfuse_eval.log_eval_trace(
+        fixture_name=fixture.name,
+        model=model,
+        prompt_variant="interactive",
+        prompt_template=prompt_template or "(not generated)",
+        actual_mutations=[{
+            "conversation": conversation[:30],
+            "user_comment": user_comment,
+        }],
+        scores_summary=result.summary(),
+        langfuse_prompt_names=["journey_system"],
+    )
+
+    if trace_id:
+        from eval.scorer import EvalScores
+        scores_obj = EvalScores(
+            fixture_name=fixture.name,
+            model=model,
+            prompt_variant="interactive",
+            precision=overall,
+            recall=float(done),
+            f1=overall,
+            llm_judge_score=overall_quality,
+            llm_judge_reasoning=judge_reasoning,
+        )
+        langfuse_eval.post_eval_scores(scores_obj, trace_id=trace_id)
+        _print_system(f"Results reported to Langfuse (trace: {trace_id})")
+    else:
+        _print_system("Langfuse not configured — results not reported.")
+
+    return result

services/batch-agent/eval/journey_runner.py

@@ -0,0 +1,385 @@
+"""Journey eval runner — tests the prompt_template builder conversation.
+
+For each (journey_fixture × model) combination:
+1. Build a MockExecutor (for filesystem tools used during journey)
+2. Patch execute_on_client
+3. Override LLM_MODEL
+4. Call handle_journey_start to kick off the conversation
+5. Feed simulated user_messages via handle_journey_message
+6. Collect the generated prompt_template
+7. Score it against expected_template_criteria (via LLM judge)
+8. Report to Langfuse
+"""
+
+from __future__ import annotations
+
+import asyncio
+import copy
+import json
+import logging
+import time
+import uuid
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+from langchain_core.messages import HumanMessage, SystemMessage
+
+from eval.config import JourneyFixture
+from eval.mock_executor import MockExecutor
+from eval import langfuse_eval
+
+logger = logging.getLogger(__name__)
+
+
+# ── Result type ──────────────────────────────────────────────────────────
+
+
+@dataclass
+class JourneyEvalResult:
+    """Result of one journey eval run."""
+
+    fixture_name: str
+    model: str
+    prompt_template: str | None  # the generated template (None if journey failed)
+    conversation_turns: int
+    done: bool  # whether journey reached completion
+    criteria_scores: dict[str, float]  # criterion → 0-1 score
+    overall_score: float  # average of criteria scores
+    judge_reasoning: str
+    elapsed_seconds: float
+
+    def summary(self) -> dict[str, Any]:
+        return {
+            "fixture": self.fixture_name,
+            "model": self.model,
+            "done": self.done,
+            "turns": self.conversation_turns,
+            "overall_score": round(self.overall_score, 3),
+            "criteria_scores": {k: round(v, 3) for k, v in self.criteria_scores.items()},
+            "elapsed_s": round(self.elapsed_seconds, 1),
+        }
+
+
+# ── LLM judge for template quality ──────────────────────────────────────
+
+_JOURNEY_JUDGE_SYSTEM = """\
+You are an evaluation judge for AI-generated prompt templates.
+
+A journey chatbot explored a user's directory structure and through
+conversation produced a prompt_template — an instruction set for a
+data-extraction agent.
+
+Your task: evaluate the generated template against a list of criteria.
+Score each criterion from 0 to 1:
+  - 1.0: Fully satisfied, clearly present in the template
+  - 0.5: Partially satisfied or ambiguously addressed
+  - 0.0: Not satisfied, missing from the template
+
+Respond with ONLY a JSON object:
+{
+  "scores": {"criterion_1": 0.8, "criterion_2": 1.0, ...},
+  "reasoning": "Brief explanation"
+}
+"""
+
+
+async def _judge_template(
+    prompt_template: str,
+    criteria: list[str],
+    *,
+    judge_model: str = "gpt-4o-mini",
+) -> tuple[dict[str, float], str]:
+    """Use an LLM to evaluate a generated prompt_template against criteria.
+
+    Returns (criteria_scores, reasoning).
+    """
+    from shared.llm import get_llm
+
+    llm = get_llm(model=judge_model, temperature=0)
+
+    criteria_text = "\n".join(f"  {i+1}. {c}" for i, c in enumerate(criteria))
+    user_content = (
+        f"## Generated prompt_template\n```\n{prompt_template}\n```\n\n"
+        f"## Criteria to evaluate\n{criteria_text}"
+    )
+
+    try:
+        response = await llm.ainvoke([
+            SystemMessage(content=_JOURNEY_JUDGE_SYSTEM),
+            HumanMessage(content=user_content),
+        ])
+        raw = response.content.strip()
+        if raw.startswith("```"):
+            raw = raw.split("```")[1]
+            if raw.startswith("json"):
+                raw = raw[4:]
+        parsed = json.loads(raw.strip())
+
+        scores_raw = parsed.get("scores", {})
+        # Map criterion keys back to the original criteria text
+        criteria_scores: dict[str, float] = {}
+        for i, criterion in enumerate(criteria):
+            # Try matching by index key or exact criterion text
+            key_candidates = [
+                f"criterion_{i+1}",
+                criterion,
+                criterion[:50],
+                str(i + 1),
+            ]
+            score = 0.0
+            for key in key_candidates:
+                if key in scores_raw:
+                    score = float(scores_raw[key])
+                    break
+            # If no match found, try values in order
+            if score == 0.0 and i < len(scores_raw):
+                score = float(list(scores_raw.values())[i])
+            criteria_scores[criterion] = score
+
+        reasoning = str(parsed.get("reasoning", ""))
+        return criteria_scores, reasoning
+    except Exception as exc:
+        logger.warning("journey_eval: LLM judge failed: %s", exc)
+        return {c: 0.0 for c in criteria}, f"Judge error: {exc}"
+
+
+# ── Journey runner ───────────────────────────────────────────────────────
+
+
+async def run_single_journey_eval(
+    fixture: JourneyFixture,
+    model: str,
+    *,
+    judge_model: str = "gpt-4o-mini",
+    data_dir: Path | None = None,
+) -> JourneyEvalResult:
+    """Execute one journey eval: start \u2192 messages \u2192 score template."""
+    from shared.config import settings
+
+    # When data_dir is given, use its parent as MockExecutor root
+    # and its name as the journey directory so the LLM sees a
+    # meaningful path (not ".").
+    if data_dir:
+        mock_root = data_dir.parent
+        journey_directory = data_dir.name
+    else:
+        mock_root = fixture.fixture_path.parent
+        journey_directory = fixture.directory
+
+    mock = MockExecutor(
+        fixture_dir=mock_root,
+        seed_records={},
+    )
+
+    original_model = settings.LLM_MODEL
+    settings.LLM_MODEL = model
+
+    eval_user_id = f"eval-journey-{uuid.uuid4().hex[:8]}"
+
+    logger.info(
+        "journey_eval: starting %s | model=%s",
+        fixture.name, model,
+    )
+    start_time = time.time()
+
+    prompt_template: str | None = None
+    conversation: list[dict[str, str]] = []
+    done = False
+
+    try:
+        from shared.ws_context import set_current_user, clear_current_user
+        from app.journey import handle_journey_start, handle_journey_message, _sessions
+
+        set_current_user(eval_user_id)
+        with mock.patch():
+            # ── Start the journey ────────────────────────────────
+            start_frame: dict[str, Any] = {
+                "agent_type": "local",
+                "directory": journey_directory,
+                "data_types": fixture.data_types,
+                "session_id": f"eval-{uuid.uuid4().hex[:8]}",
+            }
+
+            reply = await handle_journey_start(eval_user_id, start_frame)
+            session_id = reply["session_id"]
+            conversation.append({"role": "assistant", "content": reply["message"]})
+
+            logger.info(
+                "journey_eval: start reply (%d chars), done=%s",
+                len(reply["message"]), reply["done"],
+            )
+
+            if reply["done"]:
+                prompt_template = reply.get("prompt_template")
+                done = True
+            else:
+                # ── Send user messages ───────────────────────────
+                for i, user_msg in enumerate(fixture.user_messages):
+                    if done:
+                        break
+
+                    conversation.append({"role": "user", "content": user_msg})
+
+                    msg_frame: dict[str, Any] = {
+                        "session_id": session_id,
+                        "message": user_msg,
+                    }
+                    reply = await handle_journey_message(eval_user_id, msg_frame)
+                    conversation.append({"role": "assistant", "content": reply["message"]})
+
+                    logger.info(
+                        "journey_eval: turn %d reply (%d chars), done=%s",
+                        i + 1, len(reply["message"]), reply["done"],
+                    )
+
+                    if reply["done"]:
+                        prompt_template = reply.get("prompt_template")
+                        done = True
+
+                # If not done after all user messages, send a final nudge
+                if not done:
+                    nudge = "Please generate the final prompt_template now. I'm satisfied with the configuration."
+                    conversation.append({"role": "user", "content": nudge})
+
+                    nudge_frame: dict[str, Any] = {
+                        "session_id": session_id,
+                        "message": nudge,
+                    }
+                    reply = await handle_journey_message(eval_user_id, nudge_frame)
+                    conversation.append({"role": "assistant", "content": reply["message"]})
+                    if reply["done"]:
+                        prompt_template = reply.get("prompt_template")
+                        done = True
+
+    except Exception as exc:
+        logger.error("journey_eval: pipeline failed for %s/%s: %s", fixture.name, model, exc)
+    finally:
+        settings.LLM_MODEL = original_model
+        from shared.ws_context import clear_current_user
+        clear_current_user()
+
+    elapsed = time.time() - start_time
+    turns = len([c for c in conversation if c["role"] == "user"])
+
+    logger.info(
+        "journey_eval: completed in %.1fs — %d turns, done=%s, template=%s",
+        elapsed, turns, done, "yes" if prompt_template else "no",
+    )
+
+    # ── Score the template ───────────────────────────────────────
+    criteria_scores: dict[str, float] = {}
+    judge_reasoning = ""
+
+    if prompt_template and fixture.expected_template_criteria:
+        criteria_scores, judge_reasoning = await _judge_template(
+            prompt_template,
+            fixture.expected_template_criteria,
+            judge_model=judge_model,
+        )
+    elif not prompt_template:
+        criteria_scores = {c: 0.0 for c in fixture.expected_template_criteria}
+        judge_reasoning = "No prompt_template was generated — journey did not complete."
+
+    overall = (
+        sum(criteria_scores.values()) / len(criteria_scores)
+        if criteria_scores
+        else 0.0
+    )
+
+    result = JourneyEvalResult(
+        fixture_name=fixture.name,
+        model=model,
+        prompt_template=prompt_template,
+        conversation_turns=turns,
+        done=done,
+        criteria_scores=criteria_scores,
+        overall_score=overall,
+        judge_reasoning=judge_reasoning,
+        elapsed_seconds=elapsed,
+    )
+
+    # ── Report to Langfuse ───────────────────────────────────────
+    trace_id = langfuse_eval.log_eval_trace(
+        fixture_name=fixture.name,
+        model=model,
+        prompt_variant="journey",
+        prompt_template=prompt_template or "(not generated)",
+        actual_mutations=[{"conversation": conversation[:20]}],
+        scores_summary=result.summary(),
+        langfuse_prompt_names=["journey_system"],
+    )
+
+    if trace_id:
+        from eval.scorer import EvalScores
+        scores_obj = EvalScores(
+            fixture_name=fixture.name,
+            model=model,
+            prompt_variant="journey",
+            precision=overall,
+            recall=float(done),
+            f1=overall,
+            llm_judge_score=overall,
+            llm_judge_reasoning=judge_reasoning,
+        )
+        langfuse_eval.post_eval_scores(scores_obj, trace_id=trace_id)
+
+    return result
+
+
+async def run_journey_fixture_eval(
+    fixture: JourneyFixture,
+    models: list[str],
+    *,
+    judge_model: str = "gpt-4o-mini",
+    data_dir: Path | None = None,
+) -> list[JourneyEvalResult]:
+    """Run all models for a journey fixture."""
+    langfuse_eval.sync_journey_fixture_to_dataset(fixture)
+
+    results: list[JourneyEvalResult] = []
+    for model in models:
+        result = await run_single_journey_eval(
+            fixture, model, judge_model=judge_model,
+            data_dir=data_dir,
+        )
+        results.append(result)
+
+    return results
+
+
+def print_journey_results(results: list[JourneyEvalResult]) -> None:
+    """Print a formatted summary of journey eval results."""
+    if not results:
+        print("\nNo journey eval results.")
+        return
+
+    print("\n" + "=" * 95)
+    print(f"{'Fixture':<25} {'Model':<25} {'Done':>5} {'Turns':>6} {'Score':>7} {'Time':>7}")
+    print("-" * 95)
+
+    for r in results:
+        done_str = "yes" if r.done else "NO"
+        print(
+            f"{r.fixture_name:<25} {r.model:<25} {done_str:>5} "
+            f"{r.conversation_turns:>6} {r.overall_score:>7.2f} {r.elapsed_seconds:>6.1f}s"
+        )
+
+    print("=" * 95)
+
+    # Criteria breakdown
+    for r in results:
+        if r.criteria_scores:
+            print(f"\n[{r.model}] Criteria scores:")
+            for criterion, score in r.criteria_scores.items():
+                indicator = "PASS" if score >= 0.7 else "PARTIAL" if score >= 0.4 else "FAIL"
+                print(f"  {indicator:>7} ({score:.1f}) {criterion}")
+
+        if r.judge_reasoning:
+            print(f"  Judge: {r.judge_reasoning}")
+
+        if r.prompt_template:
+            preview = r.prompt_template[:200].replace("\n", " ")
+            print(f"  Template preview: {preview}...")
+
+    print()

services/batch-agent/eval/langfuse_eval.py

@@ -0,0 +1,327 @@
+"""Langfuse evaluation integration — datasets, runs, and scoring.
+
+Uses the Langfuse Python SDK v4 (OpenTelemetry-based) to:
+
+1. **Sync fixtures → Langfuse datasets**: Each YAML fixture becomes a dataset,
+   each prompt variant + expected pair becomes a dataset item.
+
+2. **Track eval runs**: Each (fixture × model × prompt_variant) execution
+   is recorded as a trace with linked scores.
+
+3. **Post scores**: precision, recall, F1, field_accuracy, llm_judge are
+   posted as numeric scores on the trace.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any
+
+from shared.config import settings
+from eval.config import EvalFixture
+from eval.scorer import EvalScores
+
+logger = logging.getLogger(__name__)
+
+
+def _get_langfuse():
+    """Get or create a Langfuse client instance (SDK v4)."""
+    if not settings.LANGFUSE_SECRET_KEY or not settings.LANGFUSE_PUBLIC_KEY:
+        return None
+    try:
+        os.environ.setdefault("LANGFUSE_SECRET_KEY", settings.LANGFUSE_SECRET_KEY)
+        os.environ.setdefault("LANGFUSE_PUBLIC_KEY", settings.LANGFUSE_PUBLIC_KEY)
+        if settings.LANGFUSE_HOST:
+            os.environ.setdefault("LANGFUSE_HOST", settings.LANGFUSE_HOST)
+        from langfuse import get_client
+        return get_client()
+    except Exception as exc:
+        logger.warning("langfuse_eval: failed to create client: %s", exc)
+        return None
+
+
+def sync_fixture_to_dataset(fixture: EvalFixture) -> str | None:
+    """Create or update a Langfuse dataset from a fixture.
+
+    Each prompt variant becomes a separate dataset item with:
+    - input: {directory, data_types, prompt_template, seed_records}
+    - expected_output: {expected records}
+
+    Returns the dataset name, or None if Langfuse is unavailable.
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        logger.info("langfuse_eval: Langfuse not configured — skipping dataset sync")
+        return None
+
+    dataset_name = f"batch-eval-{fixture.name}"
+
+    try:
+        lf.create_dataset(
+            name=dataset_name,
+            description=fixture.description,
+            metadata={
+                "data_types": ",".join(fixture.data_types),
+                "file_extensions": ",".join(fixture.file_extensions) if fixture.file_extensions else "",
+            },
+        )
+    except Exception:
+        # Dataset may already exist — that's fine
+        pass
+
+    # Build expected_output appropriate to the fixture's mode
+    expected_output: dict[str, Any] = {}
+    if fixture.mode in ("step1", "full") and fixture.expected_classification:
+        expected_output["classifications"] = [
+            {"file": ec.file, "project_id": ec.project_id, "domains": ec.domains}
+            for ec in fixture.expected_classification
+        ]
+    if fixture.mode in ("step2", "full") and fixture.expected:
+        for rec in fixture.expected:
+            expected_output.setdefault(rec.table, []).append(rec.fields)
+
+    item_id = f"{fixture.name}--{fixture.mode}"
+    try:
+        lf.create_dataset_item(
+            dataset_name=dataset_name,
+            id=item_id,
+            input={
+                "directory": fixture.directory,
+                "data_types": fixture.data_types,
+                "mode": fixture.mode,
+                "seed_records": fixture.seed_records,
+            },
+            expected_output=expected_output,
+            metadata={"mode": fixture.mode},
+        )
+    except Exception as exc:
+        logger.warning(
+            "langfuse_eval: failed to upsert dataset item %s: %s", item_id, exc
+        )
+
+    lf.flush()
+    logger.info("langfuse_eval: synced fixture '%s' → dataset '%s'", fixture.name, dataset_name)
+    return dataset_name
+
+
+def sync_journey_fixture_to_dataset(fixture) -> str | None:
+    """Create or update a Langfuse dataset from a journey fixture.
+
+    Each journey fixture becomes a single dataset item with:
+    - input: {directory, data_types, user_messages}
+    - expected_output: {criteria}
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        logger.info("langfuse_eval: Langfuse not configured — skipping journey dataset sync")
+        return None
+
+    dataset_name = f"journey-eval-{fixture.name}"
+
+    try:
+        lf.create_dataset(
+            name=dataset_name,
+            description=fixture.description,
+            metadata={"type": "journey", "data_types": ",".join(fixture.data_types)},
+        )
+    except Exception:
+        pass  # Dataset may already exist
+
+    item_id = f"{fixture.name}--journey"
+    try:
+        lf.create_dataset_item(
+            dataset_name=dataset_name,
+            id=item_id,
+            input={
+                "directory": fixture.directory,
+                "data_types": fixture.data_types,
+                "user_messages": fixture.user_messages,
+            },
+            expected_output={
+                "criteria": fixture.expected_template_criteria,
+            },
+            metadata={"type": "journey"},
+        )
+    except Exception as exc:
+        logger.warning("langfuse_eval: failed to upsert journey dataset item %s: %s", item_id, exc)
+
+    lf.flush()
+    logger.info("langfuse_eval: synced journey fixture '%s' → dataset '%s'", fixture.name, dataset_name)
+    return dataset_name
+
+
+def create_eval_run(
+    dataset_name: str,
+    run_name: str,
+    *,
+    metadata: dict[str, Any] | None = None,
+) -> str:
+    """Create a dataset run in Langfuse. Returns the run name.
+
+    Note: In SDK v4, dataset runs are created implicitly via
+    dataset.run_experiment(). This function is kept for backwards
+    compatibility but may not create a run.
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        return run_name
+
+    try:
+        if hasattr(lf, "create_dataset_run"):
+            lf.create_dataset_run(
+                dataset_name=dataset_name,
+                run_name=run_name,
+                metadata=metadata or {},
+            )
+            lf.flush()
+        else:
+            logger.debug("langfuse_eval: create_dataset_run not available in SDK v4")
+    except Exception as exc:
+        logger.warning("langfuse_eval: failed to create run %s: %s", run_name, exc)
+
+    return run_name
+
+
+def post_eval_scores(
+    scores: EvalScores,
+    *,
+    trace_id: str | None = None,
+    dataset_name: str | None = None,
+    run_name: str | None = None,
+) -> None:
+    """Post evaluation scores to Langfuse.
+
+    If trace_id is provided, scores are attached to that trace.
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        return
+
+    score_data = [
+        ("precision", scores.precision),
+        ("recall", scores.recall),
+        ("f1", scores.f1),
+    ]
+    # Only post field_accuracy when there are field-level scores (step2/full)
+    if scores.field_scores:
+        score_data.append(("field_accuracy", scores.field_accuracy))
+    if scores.llm_judge_score is not None:
+        score_data.append(("llm_judge", scores.llm_judge_score))
+
+    for name, value in score_data:
+        try:
+            lf.create_score(
+                name=name,
+                value=value,
+                trace_id=trace_id,
+                data_type="NUMERIC",
+                comment=f"{scores.fixture_name} | {scores.model} | {scores.prompt_variant}",
+            )
+        except Exception as exc:
+            logger.warning("langfuse_eval: failed to post score %s: %s", name, exc)
+
+    lf.flush()
+    logger.info(
+        "langfuse_eval: posted %d scores for %s/%s/%s",
+        len(score_data), scores.fixture_name, scores.model, scores.prompt_variant,
+    )
+
+
+def log_eval_trace(
+    *,
+    fixture_name: str,
+    model: str,
+    prompt_variant: str,
+    prompt_template: str,
+    actual_mutations: list[dict],
+    scores_summary: dict[str, Any],
+    step1_results: list[dict] | None = None,
+    dataset_name: str | None = None,
+    run_name: str | None = None,
+    dataset_item_id: str | None = None,
+    langfuse_prompt_names: list[str] | None = None,
+) -> str | None:
+    """Create a Langfuse trace for one eval execution and link it to a dataset run.
+
+    Uses SDK v4 observation API (traces are created implicitly by root spans).
+    ``langfuse_prompt_names`` can contain one or two prompt names to link
+    (e.g. ``["batch_file_classifier", "batch_processing"]`` for full mode).
+    Each prompt gets its own generation-type observation for per-version
+    metrics tracking.
+
+    Returns the trace_id, or None if Langfuse is unavailable.
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        return None
+
+    try:
+        from langfuse import propagate_attributes
+
+        # Fetch prompt objects for linking
+        prompt_objs: list[tuple[str, Any]] = []
+        for pname in (langfuse_prompt_names or []):
+            try:
+                obj = lf.get_prompt(name=pname, cache_ttl_seconds=300)
+                prompt_objs.append((pname, obj))
+                logger.info("langfuse_eval: linked prompt '%s' (type=%s)", pname, type(obj).__name__)
+            except Exception as exc:
+                logger.warning("langfuse_eval: prompt '%s' not found — %s", pname, exc)
+
+        # Build trace output dict
+        trace_output: dict[str, Any] = {"scores": scores_summary}
+        if step1_results:
+            trace_output["classifications"] = step1_results
+        if actual_mutations:
+            trace_output["mutations"] = actual_mutations[:50]
+
+        with propagate_attributes(
+            trace_name=f"eval-{fixture_name}",
+            metadata={
+                "eval": "true",
+                "fixture": fixture_name,
+                "model": model,
+                "prompt_variant": prompt_variant,
+            },
+            tags=["eval", f"model:{model}", f"variant:{prompt_variant}"],
+        ):
+            # Root span for the eval run
+            span = lf.start_observation(name=f"eval-{fixture_name}")
+            span.update(
+                input={
+                    "prompt_template": prompt_template,
+                    "model": model,
+                    "prompt_variant": prompt_variant,
+                },
+                output=trace_output,
+            )
+            trace_id = span.trace_id
+
+            # Create a generation-type observation per linked prompt
+            for pname, pobj in prompt_objs:
+                gen = lf.start_observation(
+                    name=f"prompt-{pname}",
+                    prompt=pobj,
+                    as_type="generation",
+                )
+                gen.end()
+
+            # Link to dataset run if available
+            if dataset_name and run_name and dataset_item_id:
+                try:
+                    dataset = lf.get_dataset(dataset_name)
+                    for item in dataset.items:
+                        if item.id == dataset_item_id:
+                            item.link(span, run_name)
+                            break
+                except Exception as exc:
+                    logger.warning("langfuse_eval: failed to link trace to dataset run: %s", exc)
+
+            span.end()
+
+        lf.flush()
+        return trace_id
+    except Exception as exc:
+        logger.warning("langfuse_eval: failed to create eval trace: %s", exc)
+        return None

services/batch-agent/eval/mock_executor.py

@@ -0,0 +1,258 @@
+"""Mock executor — intercepts execute_on_client for offline E2E testing.
+
+Patches ``execute_on_client`` at all usage sites so agent pipeline runs don't
+require a live Electron client or Redis.  Instead:
+
+- **Filesystem actions** (list_directory, read_file_content, get_file_metadata)
+  are served from local fixture files on disk.
+- **Read actions** (select, get) return preseeded records from an in-memory
+  store provided by the test fixture.
+- **Write actions** (insert, update, delete) are captured as *mutations* and
+  stored for later comparison against expected results.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import time
+import uuid
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+from contextlib import contextmanager, asynccontextmanager
+from unittest.mock import AsyncMock, patch
+
+
+@dataclass
+class Mutation:
+    """A single recorded write operation."""
+
+    action: str  # insert | update | delete
+    table: str
+    data: dict[str, Any]
+    timestamp: float = field(default_factory=time.time)
+
+
+# ── Fake DB helpers (used to bypass async_session in full mode) ───────
+
+class _FakeRow:
+    """Mimics an AgentRunLog row returned by SQLAlchemy."""
+    id = 0
+    status = "running"
+    items_processed = 0
+    items_created = 0
+    errors: list[str] = []
+    completed_at = None
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        object.__setattr__(self, name, value)
+
+
+class _FakeResult:
+    """Mimics a SQLAlchemy ``Result`` with ``scalar_one_or_none``."""
+    def __init__(self, row: _FakeRow) -> None:
+        self._row = row
+
+    def scalar_one_or_none(self) -> _FakeRow:
+        return self._row
+
+
+@dataclass
+class MockExecutor:
+    """In-memory executor that replaces Redis-based tool round-trip.
+
+    Parameters
+    ----------
+    fixture_dir : Path
+        Directory containing sample files for filesystem tool calls.
+    seed_records : dict[str, list[dict]]
+        Pre-existing records per table, e.g. ``{"tasks": [...], "projects": [...]}``.
+        The executor returns these for ``select`` / ``get`` actions and auto-updates
+        them on ``insert`` / ``update`` / ``delete`` so subsequent selects reflect changes.
+    """
+
+    fixture_dir: Path
+    seed_records: dict[str, list[dict]] = field(default_factory=dict)
+    mutations: list[Mutation] = field(default_factory=list)
+    _id_counter: int = field(default=1000, repr=False)
+
+    # ── Public API ───────────────────────────────────────────────────
+
+    def reset(self) -> None:
+        """Clear recorded mutations (keep seed_records intact)."""
+        self.mutations.clear()
+
+    def get_mutations(self, *, table: str | None = None, action: str | None = None) -> list[Mutation]:
+        """Filter mutations by table and/or action."""
+        result = self.mutations
+        if table:
+            result = [m for m in result if m.table == table]
+        if action:
+            result = [m for m in result if m.action == action]
+        return result
+
+    def created_records(self, table: str) -> list[dict]:
+        """Return data dicts of all inserts into *table*."""
+        return [m.data for m in self.mutations if m.table == table and m.action == "insert"]
+
+    def updated_records(self, table: str) -> list[dict]:
+        """Return data dicts of all updates to *table*."""
+        return [m.data for m in self.mutations if m.table == table and m.action == "update"]
+
+    # ── Context manager for patching ──────────────────────────────
+
+    @contextmanager
+    def patch(self):
+        """Patch execute_on_client and DB session at all usage sites."""
+        mock_fn = AsyncMock(side_effect=self._handle)
+        targets = [
+            "shared.ws_context.execute_on_client",
+            "app.agent_runner.execute_on_client",
+            "app.agents.filesystem_agent.execute_on_client",
+        ]
+
+        # Mock async_session so run_local_agent / _finalize_run skip real DB
+        fake_row = _FakeRow()
+        fake_db = AsyncMock()
+        fake_db.commit = AsyncMock()
+        fake_db.refresh = AsyncMock()
+        fake_db.execute = AsyncMock(return_value=_FakeResult(fake_row))
+        fake_db.add = lambda obj: None  # noqa: ARG005
+
+        @asynccontextmanager
+        async def _fake_session():
+            yield fake_db
+
+        patches = [patch(t, new=mock_fn) for t in targets]
+        patches.append(patch("app.agent_runner.async_session", _fake_session))
+        for p in patches:
+            p.start()
+        try:
+            yield mock_fn
+        finally:
+            for p in patches:
+                p.stop()
+
+    # ── Internal dispatch ─────────────────────────────────────────
+
+    async def _handle(
+        self,
+        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]:
+        # Filesystem
+        if action == "list_directory":
+            return self._list_directory(data or {})
+        if action == "read_file_content":
+            return self._read_file(data or {})
+        if action == "get_file_metadata":
+            return self._get_file_metadata(data or {})
+
+        # CRUD
+        if action == "select":
+            return self._select(table or "", filters)
+        if action == "get":
+            return self._get(table or "", data or {})
+        if action == "insert":
+            return self._insert(table or "", data or {})
+        if action == "update":
+            return self._update(table or "", data or {})
+        if action == "delete":
+            return self._delete(table or "", data or {})
+
+        # Vector (no-op for eval)
+        if action in ("vector_upsert", "vector_search"):
+            return {"rows": []}
+
+        return {"error": f"Unknown action: {action}"}
+
+    # ── Filesystem handlers ───────────────────────────────────────
+
+    def _list_directory(self, data: dict) -> dict:
+        rel_path = data.get("path", "")
+        abs_path = self.fixture_dir / rel_path.lstrip("/\\")
+        if not abs_path.is_dir():
+            return {"entries": []}
+        entries: list[dict] = []
+        for child in sorted(abs_path.iterdir()):
+            entry_type = "directory" if child.is_dir() else "file"
+            # Return paths relative to fixture_dir but with the original prefix
+            entry_path = rel_path.rstrip("/\\") + "/" + child.name
+            entries.append({
+                "name": child.name,
+                "path": entry_path,
+                "type": entry_type,
+            })
+        return {"entries": entries}
+
+    def _read_file(self, data: dict) -> dict:
+        rel_path = data.get("path", "")
+        abs_path = self.fixture_dir / rel_path.lstrip("/\\")
+        if not abs_path.is_file():
+            return {"content": "", "error": f"File not found: {rel_path}"}
+        return {"content": abs_path.read_text(encoding="utf-8", errors="replace")}
+
+    def _get_file_metadata(self, data: dict) -> dict:
+        rel_path = data.get("path", "")
+        abs_path = self.fixture_dir / rel_path.lstrip("/\\")
+        if not abs_path.exists():
+            return {"error": f"Not found: {rel_path}"}
+        stat = abs_path.stat()
+        return {
+            "path": rel_path,
+            "size": stat.st_size,
+            "modifiedAt": int(stat.st_mtime * 1000),
+            "createdAt": int(stat.st_ctime * 1000),
+            "isDirectory": abs_path.is_dir(),
+        }
+
+    # ── CRUD handlers ─────────────────────────────────────────────
+
+    def _select(self, table: str, filters: dict | None) -> dict:
+        rows = list(self.seed_records.get(table, []))
+        if filters:
+            rows = [
+                r for r in rows
+                if all(r.get(k) == v for k, v in filters.items() if v is not None)
+            ]
+        return {"rows": rows}
+
+    def _get(self, table: str, data: dict) -> dict:
+        record_id = data.get("id", "")
+        rows = self.seed_records.get(table, [])
+        for r in rows:
+            if r.get("id") == record_id:
+                return {"row": r}
+        return {"row": None}
+
+    def _insert(self, table: str, data: dict) -> dict:
+        self._id_counter += 1
+        record = {**data, "id": str(self._id_counter)}
+        # Add to seed so subsequent selects can find it
+        self.seed_records.setdefault(table, []).append(record)
+        self.mutations.append(Mutation(action="insert", table=table, data=record))
+        return {"row": record}
+
+    def _update(self, table: str, data: dict) -> dict:
+        record_id = data.get("id", "")
+        rows = self.seed_records.get(table, [])
+        for r in rows:
+            if r.get("id") == record_id:
+                r.update({k: v for k, v in data.items() if v is not None and v != ""})
+                self.mutations.append(Mutation(action="update", table=table, data=dict(r)))
+                return {"row": r}
+        # Record not found — still log the mutation
+        self.mutations.append(Mutation(action="update", table=table, data=data))
+        return {"row": data}
+
+    def _delete(self, table: str, data: dict) -> dict:
+        record_id = data.get("id", "")
+        rows = self.seed_records.get(table, [])
+        self.seed_records[table] = [r for r in rows if r.get("id") != record_id]
+        self.mutations.append(Mutation(action="delete", table=table, data={"id": record_id}))
+        return {"deleted": True}

services/batch-agent/eval/requirements.txt

@@ -0,0 +1,2 @@
+# Extra dependencies for the eval harness (on top of the service requirements.txt)
+pyyaml>=6.0.0

services/batch-agent/eval/runner.py

@@ -0,0 +1,545 @@
+"""Eval runner — orchestrates fixture → mock → agent pipeline → scoring.
+
+Supports three eval modes:
+
+- **step1**: Test classification prompt only (``_STEP1_SYSTEM_PROMPT``).
+  Calls the LLM with fixture-provided ``domain_definitions`` and
+  ``projects_list`` and compares output against ``expected_classification``.
+
+- **step2**: Test processing prompt only (``_PROCESSING_SYSTEM_PROMPT``).
+  Compiles the prompt with fixture-provided ``existing_context``,
+  ``project_context``, ``data_types``, and ``custom_prompt_section``,
+  then runs the tool-calling loop.  Mutations are scored against
+  ``expected`` records.
+
+- **full**: Run ``run_local_agent()`` end-to-end (both steps).
+  Scored on both classification and extraction.
+"""
+
+from __future__ import annotations
+
+import copy
+import json
+import logging
+import time
+import uuid
+from typing import Any
+
+from eval.config import EvalFixture, ExpectedClassification
+from eval.mock_executor import MockExecutor
+from eval.scorer import (
+    EvalScores,
+    FieldScore,
+    compute_precision_recall,
+    llm_judge_score,
+    score_field_match,
+)
+from eval import langfuse_eval
+
+logger = logging.getLogger(__name__)
+
+
+# ── Step 1 runner ─────────────────────────────────────────────────────────
+
+
+async def _run_step1(
+    fixture: EvalFixture,
+    model: str,
+    mock: MockExecutor,
+) -> list[dict[str, Any]]:
+    """Run step-1 classification for every file in the fixture directory.
+
+    Scans the directory recursively, classifies each file, and returns
+    a list of result dicts:
+    ``[{file, project_id, domains, new_project_name}, ...]``
+    """
+    from app.agent_runner import _classify_file
+
+    # Build project name lookup for display
+    proj_names: dict[str, str] = {
+        p.get("id", ""): p.get("name", "") for p in fixture.projects_list
+    }
+
+    # Discover all files in the fixture directory
+    all_files = await _scan_fixture_files(mock, fixture.directory)
+    print(f"\n  Scanning {len(all_files)} files in {fixture.directory}\n")
+
+    results: list[dict[str, Any]] = []
+    for i, file_path in enumerate(all_files, 1):
+        file_result = await mock._handle(
+            action="read_file_content",
+            data={"path": file_path},
+        )
+        file_content: str = file_result.get("content", "")
+        if not file_content.strip():
+            continue
+
+        project_id, domains, new_name = await _classify_file(
+            file_path=file_path,
+            file_content=file_content,
+            projects=fixture.projects_list,
+            config_data_types=fixture.data_types,
+            custom_system_prompt=fixture.custom_step1_prompt or None,
+        )
+
+        short_name = file_path.rsplit("/", 1)[-1] if "/" in file_path else file_path
+        proj_label = proj_names.get(project_id, new_name or "?")
+        print(f"  [{i}/{len(all_files)}] {short_name}  →  {project_id} ({proj_label})  {domains}")
+
+        results.append({
+            "file": file_path,
+            "project_id": project_id,
+            "domains": domains,
+            "new_project_name": new_name,
+        })
+    return results
+
+
+async def _scan_fixture_files(mock: MockExecutor, directory: str) -> list[str]:
+    """Recursively list all files under *directory* via the mock executor."""
+    files: list[str] = []
+
+    async def _walk(path: str) -> None:
+        result = await mock._handle(action="list_directory", data={"path": path})
+        for entry in result.get("entries", []):
+            if entry.get("type") == "directory":
+                await _walk(entry["path"])
+            elif entry.get("type") == "file":
+                files.append(entry["path"])
+
+    await _walk(directory)
+    return sorted(files)
+
+
+def _score_step1(
+    fixture: EvalFixture,
+    results: list[dict[str, Any]],
+) -> tuple[float, float, float, str]:
+    """Score step-1 results. Returns (precision, recall, f1, reasoning).
+
+    Files with expected classifications are scored (OK/FAIL).
+    Files without expectations are shown as informational (INFO).
+    """
+    if not fixture.expected_classification:
+        return 0.0, 0.0, 0.0, "No expected classifications"
+
+    # Build project name lookup
+    proj_names: dict[str, str] = {
+        p.get("id", ""): p.get("name", "") for p in fixture.projects_list
+    }
+    proj_names["new"] = "(new project)"
+
+    def _proj_label(pid: str, new_name: str | None = None) -> str:
+        name = proj_names.get(pid, "?")
+        if pid == "new" and new_name:
+            return f"new → \"{new_name}\""
+        return f"{pid} ({name})" if name and name != "?" else pid
+
+    def _short_file(path: str) -> str:
+        """Use just the filename for cleaner display."""
+        return path.rsplit("/", 1)[-1] if "/" in path else path
+
+    expected_files = {ec.file for ec in fixture.expected_classification}
+    total = len(fixture.expected_classification)
+    matched = 0
+
+    scored_lines: list[str] = []
+    info_lines: list[str] = []
+
+    # Score expected files
+    for ec in fixture.expected_classification:
+        actual = next((r for r in results if r["file"] == ec.file), None)
+        fname = _short_file(ec.file)
+        if actual is None:
+            scored_lines.append(f"  MISS  {fname}")
+            scored_lines.append(f"          expected: {_proj_label(ec.project_id)}")
+            continue
+
+        pid_ok = actual["project_id"] == ec.project_id
+        domains_ok = set(actual["domains"]) == set(ec.domains) if ec.domains else True
+
+        if pid_ok and domains_ok:
+            matched += 1
+            scored_lines.append(f"  OK    {fname}")
+            scored_lines.append(f"          project: {_proj_label(actual['project_id'])}")
+            scored_lines.append(f"          domains: {actual['domains']}")
+        else:
+            scored_lines.append(f"  FAIL  {fname}")
+            if not pid_ok:
+                scored_lines.append(f"          project: {_proj_label(actual['project_id'])}  (expected: {_proj_label(ec.project_id)})")
+            else:
+                scored_lines.append(f"          project: {_proj_label(actual['project_id'])}")
+            if not domains_ok:
+                scored_lines.append(f"          domains: {actual['domains']}  (expected: {ec.domains})")
+            else:
+                scored_lines.append(f"          domains: {actual['domains']}")
+
+    # Show unscored files
+    for r in results:
+        if r["file"] not in expected_files:
+            fname = _short_file(r["file"])
+            proj = _proj_label(r["project_id"], r.get("new_project_name"))
+            info_lines.append(f"  ·     {fname}")
+            info_lines.append(f"          project: {proj}  |  domains: {r['domains']}")
+
+    precision = matched / total if total > 0 else 0.0
+    recall = precision
+    f1 = precision
+
+    parts: list[str] = []
+    if scored_lines:
+        parts.append(f"Scored ({matched}/{total}):")
+        parts.extend(scored_lines)
+    if info_lines:
+        parts.append(f"\nOther files ({len(info_lines) // 2}):")
+        parts.extend(info_lines)
+
+    return precision, recall, f1, "\n".join(parts)
+
+
+# ── Step 2 runner ─────────────────────────────────────────────────────────
+
+
+async def _run_step2(
+    fixture: EvalFixture,
+    model: str,
+    mock: MockExecutor,
+) -> None:
+    """Run step-2 processing for each file in the fixture directory.
+
+    Compiles ``_PROCESSING_SYSTEM_PROMPT`` with fixture-provided variables
+    and runs the tool-calling loop.  Mutations are captured by the mock.
+    """
+    from app.agent_runner import (
+        _PROCESSING_SYSTEM_PROMPT,
+        _build_processing_tools,
+        _run_agent_with_tools,
+        _MAX_PROCESSING_STEPS,
+    )
+    from app import tracing
+
+    # Compile the processing prompt with fixture variables
+    system_prompt = tracing.compile_prompt(
+        "batch_processing",
+        fallback=_PROCESSING_SYSTEM_PROMPT,
+        variables={
+            "existing_context": fixture.existing_context,
+            "project_context": fixture.project_context,
+            "data_types": ", ".join(fixture.data_types),
+            "custom_prompt_section": fixture.custom_prompt_section,
+        },
+    )
+
+    tools = _build_processing_tools(fixture.data_types)
+
+    # Scan files in the fixture directory
+    file_entries = await mock._handle(
+        action="list_directory",
+        data={"path": fixture.directory},
+    )
+    for entry in file_entries.get("entries", []):
+        if entry.get("type") != "file":
+            continue
+        # Filter by extension if specified
+        if fixture.file_extensions:
+            ext = entry["name"].rsplit(".", 1)[-1] if "." in entry["name"] else ""
+            if ext not in fixture.file_extensions:
+                continue
+
+        file_result = await mock._handle(
+            action="read_file_content",
+            data={"path": entry["path"]},
+        )
+        file_content: str = file_result.get("content", "")
+        if not file_content.strip():
+            continue
+
+        await _run_agent_with_tools(
+            system_prompt=system_prompt,
+            user_message=(
+                f"Process this file and extract relevant information.\n\n"
+                f"File: {entry['path']}\n\nContent:\n{file_content}"
+            ),
+            tools=tools,
+            max_steps=_MAX_PROCESSING_STEPS,
+        )
+
+
+# ── Full runner ───────────────────────────────────────────────────────────
+
+
+async def _run_full(
+    fixture: EvalFixture,
+    model: str,
+    mock: MockExecutor,
+    user_id: str,
+) -> None:
+    """Run the full two-step pipeline via ``run_local_agent``."""
+    from app.agent_runner import run_local_agent
+
+    trigger_data: dict[str, Any] = {
+        "type": "agent_trigger",
+        "directory": fixture.directory,
+        "directory_paths": [fixture.directory],
+        "data_types": fixture.data_types,
+        "file_extensions": fixture.file_extensions,
+        "prompt_template": fixture.custom_prompt_section,
+        "device_id": "eval-harness",
+        "run_context": {
+            "agent_id": f"eval-{fixture.name}",
+            "run_id": None,
+        },
+    }
+
+    with mock.patch():
+        await run_local_agent(user_id, trigger_data)
+
+
+# ── Scoring helpers ───────────────────────────────────────────────────────
+
+
+def _score_mutations(
+    fixture: EvalFixture,
+    mock: MockExecutor,
+) -> tuple[list[FieldScore], float, float, float, int, int]:
+    """Score mutations against expected records.
+
+    Returns (field_scores, precision, recall, f1, extra, missing).
+    """
+    all_field_scores: list[FieldScore] = []
+    total_expected = 0
+    total_actual = 0
+    total_matched = 0
+    total_extra = 0
+    total_missing = 0
+
+    expected_by_table: dict[str, list[dict]] = {}
+    for rec in fixture.expected:
+        expected_by_table.setdefault(rec.table, []).append(rec.fields)
+
+    tables = set(expected_by_table.keys()) | {m.table for m in mock.mutations}
+    for table in tables:
+        expected_records = expected_by_table.get(table, [])
+        actual_records = mock.created_records(table) + mock.updated_records(table)
+
+        field_scores, extra, missing = score_field_match(expected_records, actual_records, table)
+        all_field_scores.extend(field_scores)
+
+        matched = sum(1 for s in field_scores if s.best_match is not None)
+        total_expected += len(expected_records)
+        total_actual += len(actual_records)
+        total_matched += matched
+        total_extra += extra
+        total_missing += missing
+
+    precision, recall, f1 = compute_precision_recall(total_expected, total_actual, total_matched)
+    return all_field_scores, precision, recall, f1, total_extra, total_missing
+
+
+# ── Main entry point ──────────────────────────────────────────────────────
+
+
+async def run_single_eval(
+    fixture: EvalFixture,
+    model: str,
+    *,
+    use_llm_judge: bool = True,
+    judge_model: str = "gpt-4o-mini",
+) -> EvalScores:
+    """Execute one eval run for a fixture + model.  Mode is read from the fixture."""
+    from shared.config import settings
+    from shared.ws_context import set_current_user, clear_current_user
+
+    seed = copy.deepcopy(fixture.seed_records)
+    mock = MockExecutor(
+        fixture_dir=fixture.fixture_path.parent,
+        seed_records=seed,
+    )
+
+    original_model = settings.LLM_MODEL
+    settings.LLM_MODEL = model
+    eval_user_id = str(uuid.uuid4())
+
+    logger.info(
+        "eval: starting %s | mode=%s | model=%s",
+        fixture.name, fixture.mode, model,
+    )
+    start_time = time.time()
+
+    step1_results: list[dict[str, Any]] = []
+    step1_reasoning = ""
+
+    try:
+        set_current_user(eval_user_id)
+
+        if fixture.mode == "step1":
+            with mock.patch():
+                step1_results = await _run_step1(fixture, model, mock)
+
+        elif fixture.mode == "step2":
+            with mock.patch():
+                await _run_step2(fixture, model, mock)
+
+        elif fixture.mode == "full":
+            with mock.patch():
+                # Step 1 — classification (independent from run_local_agent)
+                if fixture.expected_classification:
+                    step1_results = await _run_step1(fixture, model, mock)
+
+            # Step 2 — full pipeline (run_local_agent handles both steps)
+            await _run_full(fixture, model, mock, eval_user_id)
+
+    except Exception as exc:
+        logger.error("eval: pipeline failed for %s/%s: %s", fixture.name, model, exc)
+    finally:
+        settings.LLM_MODEL = original_model
+        clear_current_user()
+
+    elapsed = time.time() - start_time
+    logger.info("eval: completed in %.1fs — %d mutations", elapsed, len(mock.mutations))
+
+    # ── Score ─────────────────────────────────────────────────────
+
+    if fixture.mode == "step1":
+        s1_precision, s1_recall, s1_f1, step1_reasoning = _score_step1(fixture, step1_results)
+        scores = EvalScores(
+            fixture_name=fixture.name,
+            model=model,
+            prompt_variant=fixture.mode,
+            precision=s1_precision,
+            recall=s1_recall,
+            f1=s1_f1,
+            llm_judge_reasoning=step1_reasoning,
+        )
+    else:
+        # step2 or full — score mutations
+        field_scores, precision, recall, f1, extra, missing = _score_mutations(fixture, mock)
+        scores = EvalScores(
+            fixture_name=fixture.name,
+            model=model,
+            prompt_variant=fixture.mode,
+            field_scores=field_scores,
+            precision=precision,
+            recall=recall,
+            f1=f1,
+            extra_records=extra,
+            missing_records=missing,
+        )
+
+        # Add step1 classification scores for full mode
+        if fixture.mode == "full" and fixture.expected_classification:
+            s1_p, s1_r, s1_f1, step1_reasoning = _score_step1(fixture, step1_results)
+            scores.llm_judge_reasoning = f"Step1 classification:\n{step1_reasoning}"
+
+        # Optional LLM judge for extraction quality
+        if use_llm_judge and fixture.expected:
+            all_expected = [r.fields for r in fixture.expected]
+            all_actual = [m.data for m in mock.mutations if m.action in ("insert", "update")]
+            judge_score, reasoning = await llm_judge_score(
+                all_expected, all_actual, judge_model=judge_model,
+            )
+            scores.llm_judge_score = judge_score
+            if step1_reasoning:
+                scores.llm_judge_reasoning += f"\n\nLLM judge:\n{reasoning}"
+            else:
+                scores.llm_judge_reasoning = reasoning
+
+    # ── Report to Langfuse ────────────────────────────────────────
+    prompt_names = {
+        "step1": ["batch_file_classifier"],
+        "step2": ["batch_processing"],
+        "full": ["batch_file_classifier", "batch_processing"],
+    }.get(fixture.mode, ["batch_processing"])
+
+    trace_id = langfuse_eval.log_eval_trace(
+        fixture_name=fixture.name,
+        model=model,
+        prompt_variant=fixture.mode,
+        prompt_template=fixture.custom_prompt_section or "(default)",
+        actual_mutations=[{"action": m.action, "table": m.table, "data": m.data} for m in mock.mutations],
+        scores_summary=scores.summary(),
+        step1_results=step1_results or None,
+        langfuse_prompt_names=prompt_names,
+    )
+
+    if trace_id:
+        langfuse_eval.post_eval_scores(scores, trace_id=trace_id)
+
+        # For full mode, post classification scores separately
+        if fixture.mode == "full" and fixture.expected_classification:
+            s1_p, s1_r, s1_f1, _ = _score_step1(fixture, step1_results)
+            for name, value in [
+                ("classification_precision", s1_p),
+                ("classification_recall", s1_r),
+                ("classification_f1", s1_f1),
+            ]:
+                try:
+                    from langfuse import get_client
+                    lf = get_client()
+                    if lf:
+                        lf.create_score(
+                            name=name,
+                            value=value,
+                            trace_id=trace_id,
+                            data_type="NUMERIC",
+                            comment=f"{fixture.name} | {model} | full",
+                        )
+                except Exception:
+                    pass
+
+    return scores
+
+
+async def run_fixture_eval(
+    fixture: EvalFixture,
+    models: list[str],
+    *,
+    use_llm_judge: bool = True,
+    judge_model: str = "gpt-4o-mini",
+) -> list[EvalScores]:
+    """Run all models for a fixture."""
+    langfuse_eval.sync_fixture_to_dataset(fixture)
+
+    results: list[EvalScores] = []
+    for model in models:
+        scores = await run_single_eval(
+            fixture, model,
+            use_llm_judge=use_llm_judge,
+            judge_model=judge_model,
+        )
+        results.append(scores)
+
+    return results
+
+
+def print_results(results: list[EvalScores]) -> None:
+    """Print a formatted summary table of eval results."""
+    if not results:
+        print("\nNo eval results.")
+        return
+
+    W = 90
+
+    print("\n" + "=" * W)
+    print(f"{'Fixture':<25} {'Mode':<6} {'Model':<25} {'P':>6} {'R':>6} {'F1':>6} {'FA':>6} {'LLM':>6}")
+    print("-" * W)
+
+    for s in results:
+        llm_str = f"{s.llm_judge_score:.2f}" if s.llm_judge_score is not None else "  --"
+        fa_str = f"{s.field_accuracy:.2f}" if s.field_scores else "  --"
+        print(
+            f"{s.fixture_name:<25} {s.prompt_variant:<6} {s.model:<25} "
+            f"{s.precision:>6.2f} {s.recall:>6.2f} {s.f1:>6.2f} "
+            f"{fa_str:>6} {llm_str:>6}"
+        )
+
+    print("=" * W)
+
+    for s in results:
+        if s.llm_judge_reasoning:
+            print(f"\n{'─' * W}")
+            print(f"  {s.fixture_name}  |  {s.model}  |  {s.prompt_variant}")
+            print(f"{'─' * W}")
+            print(s.llm_judge_reasoning)
+
+    print()

services/batch-agent/eval/scorer.py

@@ -0,0 +1,268 @@
+"""Scoring functions for batch agent evaluation.
+
+Two scoring strategies:
+
+1. **FieldMatchScorer** — deterministic check: for each expected record,
+   find the best-matching actual record and compare specified fields.
+   Returns precision, recall, and per-field accuracy.
+
+2. **LLMJudgeScorer** — uses a secondary LLM to semantically evaluate
+   whether the actual extractions satisfy the expected intent, even if
+   wording differs.  Returns a 0-1 score + reasoning.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import dataclass, field
+from difflib import SequenceMatcher
+from typing import Any
+
+from langchain_core.messages import HumanMessage, SystemMessage
+
+logger = logging.getLogger(__name__)
+
+
+# ── Result types ─────────────────────────────────────────────────────────
+
+
+@dataclass
+class FieldScore:
+    """Score for a single expected record against its best match."""
+
+    expected: dict[str, Any]
+    best_match: dict[str, Any] | None
+    matched_fields: dict[str, bool]
+    similarity: float  # 0-1 overall similarity
+
+    @property
+    def field_accuracy(self) -> float:
+        if not self.matched_fields:
+            return 0.0
+        return sum(self.matched_fields.values()) / len(self.matched_fields)
+
+
+@dataclass
+class EvalScores:
+    """Aggregated scores for one eval run."""
+
+    fixture_name: str
+    model: str
+    prompt_variant: str
+    field_scores: list[FieldScore] = field(default_factory=list)
+    precision: float = 0.0
+    recall: float = 0.0
+    f1: float = 0.0
+    llm_judge_score: float | None = None
+    llm_judge_reasoning: str = ""
+    extra_records: int = 0  # records created but not expected
+    missing_records: int = 0  # expected but not found
+
+    @property
+    def field_accuracy(self) -> float:
+        if not self.field_scores:
+            return 0.0
+        return sum(s.field_accuracy for s in self.field_scores) / len(self.field_scores)
+
+    def summary(self) -> dict[str, Any]:
+        return {
+            "fixture": self.fixture_name,
+            "model": self.model,
+            "prompt_variant": self.prompt_variant,
+            "precision": round(self.precision, 3),
+            "recall": round(self.recall, 3),
+            "f1": round(self.f1, 3),
+            "field_accuracy": round(self.field_accuracy, 3),
+            "llm_judge_score": round(self.llm_judge_score, 3) if self.llm_judge_score is not None else None,
+            "extra_records": self.extra_records,
+            "missing_records": self.missing_records,
+        }
+
+
+# ── Field Match Scorer ───────────────────────────────────────────────────
+
+
+def _normalize(value: Any) -> str:
+    """Normalize a value for comparison."""
+    if value is None:
+        return ""
+    return str(value).strip().lower()
+
+
+def _text_similarity(a: str, b: str) -> float:
+    """Fuzzy text similarity using SequenceMatcher."""
+    if not a and not b:
+        return 1.0
+    if not a or not b:
+        return 0.0
+    return SequenceMatcher(None, a.lower(), b.lower()).ratio()
+
+
+def _find_best_match(
+    expected: dict[str, Any],
+    actuals: list[dict[str, Any]],
+) -> tuple[dict[str, Any] | None, float]:
+    """Find the actual record most similar to expected, return (match, similarity)."""
+    if not actuals:
+        return None, 0.0
+
+    best_match = None
+    best_score = 0.0
+
+    # Primary matching key: title or name
+    expected_title = _normalize(expected.get("title", expected.get("name", "")))
+
+    for actual in actuals:
+        actual_title = _normalize(actual.get("title", actual.get("name", "")))
+        sim = _text_similarity(expected_title, actual_title)
+        if sim > best_score:
+            best_score = sim
+            best_match = actual
+
+    return best_match, best_score
+
+
+def _compare_fields(
+    expected: dict[str, Any],
+    actual: dict[str, Any],
+) -> dict[str, bool]:
+    """Compare each expected field against the actual record."""
+    results: dict[str, bool] = {}
+    for key, expected_val in expected.items():
+        actual_val = actual.get(key)
+        # Exact match for non-string types
+        if not isinstance(expected_val, str):
+            results[key] = actual_val == expected_val
+        else:
+            # Fuzzy match for strings (threshold: 0.7)
+            results[key] = _text_similarity(
+                _normalize(expected_val), _normalize(actual_val)
+            ) >= 0.7
+    return results
+
+
+def score_field_match(
+    expected_records: list[dict[str, Any]],
+    actual_records: list[dict[str, Any]],
+    table: str,
+) -> tuple[list[FieldScore], int, int]:
+    """Score actual extractions against expected records for one table.
+
+    Returns (field_scores, extra_count, missing_count).
+    """
+    field_scores: list[FieldScore] = []
+    matched_actuals: set[int] = set()
+
+    for exp in expected_records:
+        # Find best match among unmatched actuals
+        candidates = [
+            (i, a) for i, a in enumerate(actual_records) if i not in matched_actuals
+        ]
+        if not candidates:
+            field_scores.append(FieldScore(
+                expected=exp, best_match=None, matched_fields={}, similarity=0.0,
+            ))
+            continue
+
+        best_idx, best_match = None, None
+        best_sim = 0.0
+        for idx, actual in candidates:
+            _, sim = _find_best_match(exp, [actual])
+            if sim > best_sim:
+                best_sim = sim
+                best_idx = idx
+                best_match = actual
+
+        if best_sim >= 0.5 and best_match is not None:
+            matched_actuals.add(best_idx)
+            matched_fields = _compare_fields(exp, best_match)
+            field_scores.append(FieldScore(
+                expected=exp, best_match=best_match,
+                matched_fields=matched_fields, similarity=best_sim,
+            ))
+        else:
+            field_scores.append(FieldScore(
+                expected=exp, best_match=None, matched_fields={}, similarity=0.0,
+            ))
+
+    extra_count = len(actual_records) - len(matched_actuals)
+    missing_count = sum(1 for s in field_scores if s.best_match is None)
+
+    return field_scores, extra_count, missing_count
+
+
+def compute_precision_recall(
+    expected_count: int,
+    actual_count: int,
+    matched_count: int,
+) -> tuple[float, float, float]:
+    """Compute precision, recall, F1."""
+    precision = matched_count / actual_count if actual_count > 0 else 0.0
+    recall = matched_count / expected_count if expected_count > 0 else 0.0
+    f1 = (
+        2 * precision * recall / (precision + recall)
+        if (precision + recall) > 0
+        else 0.0
+    )
+    return precision, recall, f1
+
+
+# ── LLM Judge Scorer ─────────────────────────────────────────────────────
+
+_JUDGE_SYSTEM_PROMPT = """\
+You are an evaluation judge for a data extraction system.
+
+Your task is to compare the EXPECTED extractions against the ACTUAL extractions
+produced by an AI agent, and assess quality on a 0-1 scale.
+
+Scoring criteria:
+- 1.0: All expected records found with correct fields, no significant extras
+- 0.8: Most expected records found, minor field differences or extras
+- 0.6: Core extractions present but some missing or incorrect
+- 0.4: Partial match — several expected records missing or wrong
+- 0.2: Poor quality — most expected records missing or incorrect
+- 0.0: Complete failure — no meaningful overlap
+
+Consider semantic equivalence: "Send invoice" and "Email the invoice" are matches.
+Ignore field ordering and formatting differences.
+
+Respond with ONLY a JSON object:
+{"score": 0.85, "reasoning": "Brief explanation of the score"}
+"""
+
+
+async def llm_judge_score(
+    expected: list[dict[str, Any]],
+    actual: list[dict[str, Any]],
+    *,
+    judge_model: str = "gpt-4o-mini",
+) -> tuple[float, str]:
+    """Use an LLM to semantically evaluate extraction quality.
+
+    Returns (score, reasoning).
+    """
+    from shared.llm import get_llm
+
+    llm = get_llm(model=judge_model, temperature=0)
+
+    user_content = (
+        f"## Expected extractions\n```json\n{json.dumps(expected, indent=2, default=str)}\n```\n\n"
+        f"## Actual extractions\n```json\n{json.dumps(actual, indent=2, default=str)}\n```"
+    )
+
+    try:
+        response = await llm.ainvoke([
+            SystemMessage(content=_JUDGE_SYSTEM_PROMPT),
+            HumanMessage(content=user_content),
+        ])
+        raw = response.content.strip()
+        if raw.startswith("```"):
+            raw = raw.split("```")[1]
+            if raw.startswith("json"):
+                raw = raw[4:]
+        parsed = json.loads(raw.strip())
+        return float(parsed.get("score", 0.0)), str(parsed.get("reasoning", ""))
+    except Exception as exc:
+        logger.warning("eval: LLM judge failed: %s", exc)
+        return 0.0, f"Judge error: {exc}"

services/batch-agent/requirements.txt

@@ -0,0 +1,21 @@
+fastapi>=0.115.0
+uvicorn[standard]>=0.34.0
+gunicorn>=22.0.0
+pydantic>=2.10.0
+pydantic-settings>=2.7.0
+sqlalchemy>=2.0.0
+asyncpg>=0.30.0
+redis>=5.0.0
+cryptography>=42.0.0
+python-dotenv>=1.0.0
+langchain-core>=0.3.0
+langchain-openai>=0.3.0
+langchain-litellm>=0.3.0
+litellm>=1.50.0
+openai>=1.50.0
+httpx>=0.27.0
+langfuse>=3.0.0
+croniter>=2.0.0
+google-api-python-client>=2.130.0
+google-auth>=2.30.0
+msal>=1.28.0

services/billing/Dockerfile

@@ -0,0 +1,36 @@
+# ── builder ──────────────────────────────────────────────────────────────────
+FROM python:3.12-slim AS builder
+
+WORKDIR /build
+
+COPY services/billing/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
+
+# Shared module
+COPY shared/ shared/
+
+# Service source
+COPY services/billing/app/ app/
+
+RUN chown -R appuser:appgroup /app
+
+USER appuser
+
+EXPOSE 8000
+
+# Billing is lightweight — single worker is fine
+CMD ["gunicorn", "app.main:app", \
+     "-k", "uvicorn.workers.UvicornWorker", \
+     "--bind", "0.0.0.0:8000", \
+     "--workers", "1", \
+     "--timeout", "30"]