14 KiB
14 KiB
| phase | plan | type | wave | depends_on | files_modified | autonomous | requirements | must_haves | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 33-persistent-memory | 03 | execute | 2 |
|
|
true |
|
|
Purpose: ASST-01 requires memory to shape future responses (needs real AI, not echo). ASST-03 requires one-click handoff to PM agent with context transfer. Output: Working AI streaming with memory, handoff route, and wired UI button.
<execution_context> @$HOME/.claude/get-shit-done/workflows/execute-plan.md @$HOME/.claude/get-shit-done/templates/summary.md </execution_context>
@.planning/PROJECT.md @.planning/ROADMAP.md @.planning/STATE.md @.planning/phases/33-persistent-memory/33-RESEARCH.md @.planning/phases/33-persistent-memory/33-01-SUMMARY.md @.planning/phases/33-persistent-memory/33-02-SUMMARY.md @server/src/routes/chat.ts @server/src/services/puter-proxy.ts @server/src/services/chat.ts @ui/src/api/chat.ts @ui/src/hooks/useStreamingChat.ts From server/src/services/assistant-memory.ts (created in 33-01): ```typescript export function assistantMemoryService() { return { get(companyId: string): Promise<{ facts: string[]; updatedAt: string | null }>, append(companyId: string, rawFact: string): Promise<{ facts: string[]; updatedAt: string | null }>, clear(companyId: string): Promise, }; } ```From server/src/services/nexus-settings.ts:
export function nexusSettingsService() {
return {
get(): Promise<{ mode: "personal_ai" | "project_builder" | "both" }>,
set(patch: Partial<NexusSettings>): Promise<NexusSettings>,
};
}
From server/src/services/puter-proxy.ts:
export function puterProxyService(db: Db) {
return {
chatStream(
companyId: string,
agentId: string | null | undefined,
messages: unknown[],
model: string | undefined,
signal: AbortSignal | undefined,
): AsyncGenerator<string>,
resolveToken(companyId: string): Promise<string>,
};
}
From server/src/services/chat.ts:
// chatService(db) returns object with:
// addMessage(conversationId, { role, content, agentId? })
// addSystemMessage(conversationId, { content, messageType, agentId? })
// getConversation(id) -> { id, companyId, agentId, ... }
// listMessages(conversationId, { cursor?, limit? })
// createConversation(companyId, { title?, agentId? })
From ui/src/api/chat.ts (SSE client parser - current):
// Client expects: { type: "token", token: string } and { type: "done", messageId: string, content: string }
// Server currently sends: { token: string } and { done: true, messageId, content } (MISMATCH)
// This plan MUST fix the server to send the format the client expects
From ui/src/pages/PersonalAssistant.tsx (created in 33-02):
// Has a disabled "Turn this into a project" button — this plan wires it
1. BEFORE `res.flushHeaders()` (this is critical — Pitfall 3 from research), add:
- Resolve the conversation: `const conversation = await svc.getConversation(req.params.id!);`
- Get nexus mode: `const settings = await nexusSettingsService().get();`
- Check if assistant mode: `const isAssistant = settings.mode !== "project_builder";`
- If isAssistant, load memory: `const memory = await assistantMemoryService().get(conversation.companyId);`
- Try resolving puter token: wrap `puterProxyService(db).resolveToken(conversation.companyId)` in try/catch — if no token available, fall back to echo stub
2. Build messages array for AI call:
- Fetch recent messages: `const recentMsgs = await svc.listMessages(req.params.id!, { limit: 20 });`
- Build OpenAI-format messages array from recentMsgs.items (reverse to chronological order)
- If isAssistant AND memory.facts.length > 0, prepend system message:
```
{ role: "system", content: `[Memory from previous sessions]\n${memory.facts.map(f => "- " + f).join("\n")}\n\nUse these facts to personalize your responses. Do not mention that you have a memory system unless asked.` }
```
- Add the new user message: `{ role: "user", content }`
- Cap the injected system prefix at 2000 characters
3. After `res.flushHeaders()` and `:ok`, replace `svc.streamEcho(content, abort.signal)` with:
- If puter token available: `puterProxyService(db).chatStream(conversation.companyId, agentId, messagesWithMemory, undefined, abort.signal)`
- Else: fall back to `svc.streamEcho(content, abort.signal)` (keeps existing behavior for users without puter token)
4. FIX SSE data format to match client expectations:
- Token events: `res.write(\`data: ${JSON.stringify({ type: "token", token })}\n\n\`)` (add `type: "token"`)
- Done events: `res.write(\`data: ${JSON.stringify({ type: "done", messageId: message.id, content: fullContent.trim() })}\n\n\`)` (change `done: true` to `type: "done"`)
- Error events: `res.write(\`data: ${JSON.stringify({ type: "error", error: "Stream error" })}\n\n\`)` (add `type: "error"`)
5. After the done event (after saving the assistant message), if isAssistant:
- Extract a brief fact from the exchange: use the last user message + first 200 chars of assistant response as a simple fact: `User asked about: ${content.slice(0, 100)}. Assistant topic: ${fullContent.slice(0, 100)}`
- Call `assistantMemoryService().append(conversation.companyId, fact)` — non-blocking (`.catch(() => {})`)
6. Move the conversation fetch for push notification to use the already-fetched `conversation` variable (avoid redundant DB call).
Import at top of file:
- `import { assistantMemoryService } from "../services/assistant-memory.js";`
- `import { nexusSettingsService } from "../services/nexus-settings.js";`
- `import { puterProxyService } from "../services/puter-proxy.js";`
- The route function already receives `db: Db` as parameter.
Wire in `server/src/app.ts`:
- Import and mount: `app.use("/api", assistantHandoffRoutes(db))`
Update `ui/src/api/chat.ts`:
- Add method: `assistantHandoff(conversationId: string)` that POSTs to `/conversations/${conversationId}/assistant-handoff` and returns `{ targetConversationId: string }`
Update `ui/src/pages/PersonalAssistant.tsx`:
- Wire the "Turn this into a project" button:
- Remove disabled state
- On click: call `chatApi.assistantHandoff(currentConversationId)`
- On success: navigate to `/dashboard` (or `/projects` — the PM conversation will appear in the conversation list)
- Show toast on success: "Conversation handed off to PM"
- Show toast on error
Create `server/src/__tests__/33-assistant-handoff.test.ts`:
- Test the route handler creates a new conversation with handoff_context system message
- Test summary is built from user messages only
- Test summary is capped at 1500 chars
- Use mock chatService
<success_criteria>
- Stream endpoint uses puterProxyService for real AI (with echo fallback)
- Memory facts prepended as system message to AI calls
- Facts appended to memory after each assistant turn
- SSE format fixed to include type field
- Handoff creates PM conversation with context summary
- UI button triggers handoff and navigates
- All tests pass, TypeScript clean </success_criteria>