---
phase: 33-persistent-memory
plan: 03
subsystem: server/routes + ui/pages
tags: [ai-streaming, memory-injection, sse, handoff, react, express]

requires:
  - phase: 33-persistent-memory/01
    provides: assistantMemoryService (get/append/clear)
  - phase: 33-persistent-memory/02
    provides: PersonalAssistant page, chatApi, useNexusMode
  - phase: 31-puter-js-zero-config-cloud
    provides: puterProxyService.chatStream, nexusSettingsService

provides:
  - Real AI streaming via puterProxyService with echo fallback (no puter token)
  - Memory facts injected as system message prefix in AI calls
  - Facts appended to memory after each assistant turn
  - SSE format fixed to {type:"token"}, {type:"done"}, {type:"error"}
  - POST /conversations/:id/assistant-handoff route
  - chatApi.assistantHandoff UI method
  - Wired "Turn into project" button in PersonalAssistant

affects:
  - server/src/routes/chat.ts
  - server/src/routes/assistant-handoff.ts
  - server/src/app.ts
  - ui/src/api/chat.ts
  - ui/src/pages/PersonalAssistant.tsx

tech-stack:
  added: []
  patterns:
    - Pre-fetch conversation/settings/memory BEFORE flushHeaders (avoids SSE header race)
    - puterProxyService try/catch for graceful echo fallback
    - buildHandoffSummary pure function (exported + unit tested separately)
    - Non-blocking memory append via .catch(() => {})

key-files:
  created:
    - server/src/routes/assistant-handoff.ts
    - server/src/__tests__/33-assistant-handoff.test.ts
  modified:
    - server/src/routes/chat.ts
    - server/src/app.ts
    - ui/src/api/chat.ts
    - ui/src/pages/PersonalAssistant.tsx

key-decisions:
  - "Pre-fetch conversation + settings + memory BEFORE res.flushHeaders() to avoid race condition (Pitfall 3 from research)"
  - "puterProxyService.resolveToken wrapped in try/catch — graceful fallback to streamEcho when no puter token configured"
  - "buildHandoffSummary exported as named function for direct unit testing without route test harness"
  - "Personal Assistant handoff navigates to /dashboard after success — PM conversation will appear in conversation list"
  - "Memory injection skipped for project_builder mode via isAssistant flag from nexus settings"

metrics:
  duration: 20
  completed_date: "2026-04-01"
  tasks_completed: 2
  files_changed: 6
---

# Phase 33 Plan 03: AI Streaming + Memory Injection + Handoff Summary

**Real AI streaming via puterProxyService with memory-injected system prompt, SSE format fix, and assistant-to-PM handoff route with wired UI button.**

## Performance

- **Duration:** ~20 min
- **Completed:** 2026-04-01
- **Tasks:** 2
- **Files modified:** 6

## Accomplishments

### Task 1: Replace streamEcho with real AI streaming + memory injection

- `server/src/routes/chat.ts` updated with:
  - Imports for `assistantMemoryService`, `nexusSettingsService`, `puterProxyService`
  - Pre-flushHeaders block: resolves conversation, nexus settings, memory facts
  - `puterProxyService(db).resolveToken()` try/catch for echo fallback
  - Builds `messagesWithMemory` array: system message with memory facts (capped 2000 chars) + conversation history + new user message
  - Uses `puterProxyService.chatStream` when token available, falls back to `streamEcho`
  - **SSE format fixed**: `{type:"token", token}`, `{type:"done", messageId, content}`, `{type:"error", error}` — matches client parser
  - Non-blocking memory fact append after each assistant turn
  - Memory injection skipped when mode is `project_builder`

### Task 2: Assistant handoff route and wired UI button (TDD)

- RED: `server/src/__tests__/33-assistant-handoff.test.ts` — 7 tests written before implementation (all failing)
- GREEN: `server/src/routes/assistant-handoff.ts` created:
  - `buildHandoffSummary()` pure function: filters user-role messages, concatenates, caps at 1500 chars
  - `assistantHandoffRoutes(db)`: `POST /conversations/:id/assistant-handoff` with assertBoard auth
  - Creates new conversation, inserts `handoff_context` system message with user summary
  - Returns `{ targetConversationId }`
- `server/src/app.ts`: `assistantHandoffRoutes(db)` mounted via `api.use()`
- `ui/src/api/chat.ts`: `assistantHandoff(conversationId)` method added
- `ui/src/pages/PersonalAssistant.tsx`: "Turn into project" button fully wired
  - `handleHandoff` callback using `chatApi.assistantHandoff`
  - Success toast + navigate to `/dashboard`
  - Error toast on failure
  - Loading spinner during handoff
  - Button disabled when no conversation selected or handing off

## Task Commits

1. **Task 1: Real AI streaming + memory injection** — `d9a00d25`
2. **Task 2 (RED): Failing tests for handoff route** — `b5028a5d`
3. **Task 2 (GREEN): Handoff route + app wiring + UI button** — `17564d9c`

## Files Created/Modified

- `server/src/routes/chat.ts` — Updated stream endpoint with real AI, memory injection, SSE format fix
- `server/src/routes/assistant-handoff.ts` — New handoff route with buildHandoffSummary
- `server/src/app.ts` — Added assistantHandoffRoutes mount
- `server/src/__tests__/33-assistant-handoff.test.ts` — 7 unit tests (4 for buildHandoffSummary, 3 for route handler)
- `ui/src/api/chat.ts` — Added assistantHandoff method
- `ui/src/pages/PersonalAssistant.tsx` — Wired "Turn into project" button with handoff flow

## Deviations from Plan

### Auto-fixed Issues

**1. [Rule 3 - Blocking] Prerequisite files missing from worktree**

- **Found during:** Task 1 setup
- **Issue:** Worktree `worktree-agent-acd379e3` is based on an older commit (4c8cfcd8) that predates phase-21 chat system, phase-30 nexus settings, and phase-33 memory services. Files referenced by this plan did not exist.
- **Fix:** Used `git checkout gsd/phase-33-persistent-memory -- ...` to bring 33-01/02 files, and `git checkout gsd/phase-31-puter-js-zero-config-cloud -- ...` for puter-proxy and nexus-settings. Also checked out `pushService.ts` from PAP-878 branch.
- **Files:** 18 prerequisite files checked out in Task 1 commit
- **Commit:** d9a00d25

**2. [Rule 3 - Blocking] Tooltip component removed from PersonalAssistant imports**

- **Found during:** Task 2 (wiring UI button)
- **Issue:** The plan called for removing the disabled state from the button. The previous `Tooltip` wrapper was removed since the button is now fully functional (no need for "Coming soon" tooltip).
- **Fix:** Removed unused `Tooltip`, `TooltipTrigger`, `TooltipContent` imports; simplified button to direct `onClick` handler.
- **Files:** `ui/src/pages/PersonalAssistant.tsx`
- **Commit:** 17564d9c

## Known Stubs

None — all data paths are fully wired. The `puterProxyService` fallback to `streamEcho` is intentional behavior (not a stub) for users without a configured puter token.

## Test Results

- **24 tests pass** across 3 test files:
  - `33-memory-sanitization.test.ts`: 10 tests
  - `33-assistant-memory.test.ts`: 7 tests
  - `33-assistant-handoff.test.ts`: 7 tests (4 buildHandoffSummary + 3 route integration)

## Self-Check: PASSED

Files exist:
- FOUND: /opt/nexus/.claude/worktrees/agent-acd379e3/server/src/routes/chat.ts
- FOUND: /opt/nexus/.claude/worktrees/agent-acd379e3/server/src/routes/assistant-handoff.ts
- FOUND: /opt/nexus/.claude/worktrees/agent-acd379e3/server/src/app.ts
- FOUND: /opt/nexus/.claude/worktrees/agent-acd379e3/server/src/__tests__/33-assistant-handoff.test.ts
- FOUND: /opt/nexus/.claude/worktrees/agent-acd379e3/ui/src/api/chat.ts
- FOUND: /opt/nexus/.claude/worktrees/agent-acd379e3/ui/src/pages/PersonalAssistant.tsx

Commits:
- FOUND: d9a00d25
- FOUND: b5028a5d
- FOUND: 17564d9c