nexus/.planning/phases/22-agent-streaming/22-01-SUMMARY.md

118 lines
6.2 KiB
Markdown

---
phase: 22-agent-streaming
plan: "01"
subsystem: chat-streaming
tags: [sse, streaming, chat, websockets-alternative, abort]
dependency_graph:
requires: [22-00]
provides: [chat-streaming-endpoint, useStreamingChat-hook, chat-edit-truncate-api]
affects: [server/src/routes/chat.ts, server/src/services/chat.ts, ui/src/hooks/useStreamingChat.ts, ui/src/api/chat.ts]
tech_stack:
added: ["@testing-library/react ^16.0.0 (devDep)"]
patterns:
- "SSE via fetch ReadableStream (not EventSource — POST-based stream)"
- "AbortController for stop generation (CHAT-12)"
- "startTransition for non-blocking token accumulation (PERF-02)"
- "res.flushHeaders() before for-await loop for sub-100ms first-token (PERF-02)"
- "res.writable guard on all res.write() calls (prevents write-after-end)"
key_files:
created:
- ui/src/hooks/useStreamingChat.ts
- ui/src/hooks/useStreamingChat.test.ts
modified:
- server/src/services/chat.ts
- server/src/routes/chat.ts
- ui/src/api/chat.ts
- ui/package.json
decisions:
- "Used fetch with ReadableStream instead of EventSource because endpoint is POST-based (EventSource only supports GET)"
- "Partial content on stop saved with [stopped] suffix via savePartialMessage (Open Question 3 resolved)"
- "streamEcho is a stub async generator yielding word-by-word with 50ms delay — real LLM in Phase 23"
- "chatMessages schema has no updatedAt column so editMessage only sets content (adapted from plan which assumed updatedAt)"
- "Added @testing-library/react as devDep (plan said already installed but it was not present)"
metrics:
duration: "~6 minutes"
completed: "2026-04-01"
tasks: 2
files: 6
---
# Phase 22 Plan 01: SSE Streaming Endpoint and useStreamingChat Hook Summary
One-liner: SSE streaming endpoint with stub echo generator, AbortController stop, partial-content persistence, and fetch ReadableStream client hook.
## Tasks Completed
| Task | Name | Commit | Key Files |
|------|------|--------|-----------|
| 1 | Server SSE streaming endpoint + edit/truncate service methods | 2d711c7e | server/src/services/chat.ts, server/src/routes/chat.ts |
| 2 | useStreamingChat hook, chat API stream method, real unit tests | 78742239 | ui/src/hooks/useStreamingChat.ts, ui/src/hooks/useStreamingChat.test.ts, ui/src/api/chat.ts |
## What Was Built
### Server (Task 1)
**`server/src/services/chat.ts`** — Three new methods added to `chatService`:
- `editMessage(messageId, content)` — Updates message content in DB (adapted: no updatedAt column in chatMessages schema)
- `truncateMessagesAfter(conversationId, messageId)` — Deletes all messages after a given createdAt timestamp; imports `gt` from drizzle-orm
- `streamEcho(content, signal)` — Async generator stub yielding words with 50ms delay, respects AbortSignal
**`server/src/routes/chat.ts`** — Three new routes:
- `POST /conversations/:id/stream` — SSE endpoint; headers flushed (PERF-02) before for-await loop; saves full content as assistant message on completion; handles req.on("close") abort
- `PATCH /conversations/:id/messages/:msgId` — Edit message content
- `DELETE /conversations/:id/messages/after/:msgId` — Truncate subsequent messages
### Client (Task 2)
**`ui/src/api/chat.ts`** — Two new methods:
- `postMessageAndStream` — Opens POST fetch with ReadableStream, parses SSE data lines, dispatches onToken/onDone/onError callbacks
- `savePartialMessage` — Delegates to `postMessage` to persist partial content on stop
**`ui/src/hooks/useStreamingChat.ts`** — New hook:
- `startStream(userMessage, agentId?)` — Creates AbortController, calls postMessageAndStream, wraps setStreamingContent in startTransition
- `stop()` — Aborts controller, saves partial content with " [stopped]" suffix, invalidates queries
- Returns: `{ streamingContent, isStreaming, startStream, stop }`
**`ui/src/hooks/useStreamingChat.test.ts`** — 5 passing unit tests:
1. Token accumulation into streamingContent
2. isStreaming lifecycle (true on start, false on done)
3. stop() aborts controller and sets isStreaming=false
4. SSE error sets isStreaming=false
5. null conversationId guard
## Verification Results
- `pnpm --filter @paperclipai/ui exec -- tsc --noEmit` — PASS (0 errors)
- `pnpm --filter @paperclipai/server exec -- tsc --noEmit` (chat files) — PASS (0 chat errors; pre-existing plugin-sdk errors in unrelated files)
- vitest: 5/5 tests passing, 0 todos
- PERF-02: flushHeaders() position (pre-loop) verified via Python script
## Deviations from Plan
### Auto-fixed Issues
**1. [Rule 2 - Missing Dependency] @testing-library/react not installed**
- **Found during:** Task 2 test execution
- **Issue:** Plan stated `renderHook` from `@testing-library/react` was "already installed" but the package was not present in ui/package.json or the pnpm lockfile
- **Fix:** Added `"@testing-library/react": "^16.0.0"` to ui/package.json devDependencies and ran `pnpm install --filter @paperclipai/ui`
- **Files modified:** ui/package.json, pnpm-lock.yaml
- **Commit:** 78742239
**2. [Rule 1 - Bug] chatMessages schema has no updatedAt column**
- **Found during:** Task 1 — implementing editMessage
- **Issue:** Plan's `editMessage` code included `.set({ content, updatedAt: new Date() })` but the `chatMessages` schema only has `id`, `conversationId`, `role`, `content`, `agentId`, `createdAt` — no `updatedAt`
- **Fix:** Removed `updatedAt: new Date()` from the update set; only `content` is updated
- **Files modified:** server/src/services/chat.ts
- **Commit:** 2d711c7e
**3. [Rule 3 - Blocking] Worktree branch was behind phase 22 foundation**
- **Found during:** Initial setup
- **Issue:** The worktree branch `worktree-agent-a8157dfc` was on nexus/main commits that diverged from phase 22; chat files did not exist
- **Fix:** Reset worktree branch to `gsd/phase-22-agent-streaming` with `git reset --hard`
- **Impact:** Nexus-specific commits (phase 01-04 branding/onboarding work) are not present on this worktree; those exist on nexus/main branch and are not related to phase 22
## Known Stubs
- `streamEcho` in `server/src/services/chat.ts` is a stub that echoes back the user message word-by-word with 50ms delay. This is intentional for Phase 22 (no LLM). Phase 23 will replace this with real LLM adapter calls.
## Self-Check: PASSED