Update AI instructions

This commit is contained in:
Roberto
2026-02-26 14:27:01 +01:00
parent 3ec501388c
commit f4eb278692
5 changed files with 35 additions and 1584 deletions

View File

@@ -67,10 +67,44 @@ To add a new table or column: edit `schema.ts`, run `drizzle-kit generate`, then
3. **Types**`AppRouter` is exported from `src/main/router/index.ts` and imported in `src/renderer/lib/trpc.ts` — types flow automatically
4. **UI** — Create components under `src/renderer/components/<feature>/`, use `trpc.*.*useQuery()` for data
### AI Subsystem (`src/main/ai/`)
LangGraph-based agentic system with pluggable LLM providers (OpenAI, Anthropic, GitHub Copilot).
**Orchestrator** (`orchestrator.ts`): Classifies user intent → routes to one of three specialist agents:
- **Project agent** — project-scoped Q&A with tools: `read_project_notes`, `add_task`, `get_summary`, `suggest_checkpoints`, `suggest_tasks`
- **Knowledge agent** — cross-project semantic search via `vector_search_all`
- **General agent** — workspace-wide `add_task`
Tool-calling strategy differs by provider: OpenAI/Anthropic use LangChain `bindTools()` + ToolMessage loop (max 5 iterations); Copilot uses SDK-native tools (loop handled internally).
**Streaming**: Orchestrator calls `sendStreamChunk(sender, token, done)` over IPC channel `'ai:stream'`. Renderer subscribes via `window.electronAI.onStreamChunk()` in `AIChatPanel.tsx`. `<tool_call>` blocks are filtered before sending to renderer.
**Provider factory** (`llm.ts`): `gpt-4o-mini` (OpenAI), `claude-sonnet-4-20250514` (Anthropic), or ChatCopilot wrapper — all with `temperature: 0.3` and streaming enabled.
**Token storage** (`token.ts`) — three-tier fallback:
1. keytar (OS keychain) — preferred, encrypted per-user
2. electron-store + `safeStorage` — encrypted at rest
3. Plain electron-store — WSL fallback
Keytar service name is `'adiuva'`. Once keytar fails, `keytarFailed` flag skips it for the session.
**AI approval pattern**: Tasks and checkpoints have `isAiSuggested` (bool) and `isApproved` (bool) columns. AI-suggested items appear in the UI pending user approval before being treated as real records.
### Vector Embeddings (`src/main/db/vectordb.ts`)
LanceDB stored in `{userData}/vectors/`. Table schema: `{ id, projectId, content, vector }`. Vectors are 1536-dimensional (`text-embedding-3-small`). Embeddings use a priority chain: Copilot CLI token → OpenAI token.
- Note create/update fires `upsertNoteEmbedding()` (fire-and-forget, errors swallowed)
- `migrateNotesIfNeeded()` backfills existing notes on first startup
- `searchNotes(query, limit=5)` is called by the Knowledge agent tool
### Key Config Notes
- Vite configs use `.mts` extension (not `.ts`) to avoid ESM/CJS conflicts with electron-forge's externalize-deps plugin
- `@/*` path alias resolves to `src/renderer/*` (TypeScript + Vite + shadcn/ui all share this alias)
- shadcn/ui style: **new-york**, base color: **neutral**
- Icons: **lucide-react** throughout — do not introduce other icon libraries
- Tailwind 4 (not 3) — use CSS variable theming via `globals.css`, not `tailwind.config.js`
- Tailwind 4 (not 3) — use CSS variable theming via `globals.css`, not `tailwind.config.js`
- Notes use Milkdown (`@milkdown/crepe`) as the markdown editor (`src/renderer/components/notes/MilkdownEditor.tsx`)
- Routes: `index`, `tasks`, `timeline`, `projects`, `notes.$noteId` (note ID is a URL param)