claude-mem/context/temp/PHASE2-PROMPT.md

Phase 2 Implementation Prompt

Use this prompt to start a new chat for Phase 2 implementation:

---

Context

I'm implementing a refactor of the claude-mem memory system based on REFACTOR-PLAN.md.

Phase 1 is complete (see PHASE1-COMPLETE.md):

• ✅ Database schema with migration 004
• ✅ HooksDatabase shared layer
• ✅ All four hook functions (context, new, save, summary)
• ✅ CLI integration and tests passing

Task

Implement Phase 2: SDK Worker Process

According to REFACTOR-PLAN.md (lines 296-423), I need to:

1. Create SDK Worker Process (src/sdk/worker.ts)

   • Uses Agent SDK streaming input mode
   • AsyncIterable message generator that:
     • Yields initial prompt
     • Polls observation_queue table
     • Yields observation prompts
     • Handles FINALIZE message
   • Parses SDK responses for <observation> and <summary> XML blocks
   • Stores results using HooksDatabase methods

2. Create SDK Prompts (src/sdk/prompts.ts)

   • buildInitPrompt() - Initialize agent (see REFACTOR-PLAN.md:537-595)
   • buildObservationPrompt() - Send tool observation (see REFACTOR-PLAN.md:601-634)
   • buildFinalizePrompt() - Request summary (see REFACTOR-PLAN.md:640-692)

3. Create XML Parser (src/sdk/parser.ts)

   • Parse <observation> blocks with <type> and <text>
   • Parse <summary> blocks with 8 required fields
   • Extract file arrays from <file> child elements

4. Update newHook (src/hooks/new.ts)

   • Uncomment SDK worker spawn code
   • Pass session ID to worker
   • Detached process with stdio: 'ignore'

5. Test End-to-End

   • Create test that simulates full lifecycle
   • Verify observations are queued, processed, and stored
   • Verify summary generation works

Key Requirements

From REFACTOR-PLAN.md:

• Use @anthropic-ai/claude-agent-sdk query function with streaming input mode
• Model: claude-sonnet-4-5
• Use disallowedTools: ['Glob', 'Grep', 'ListMcpResourcesTool', 'WebSearch']
• Message generator yields { role: "user", content: string } objects
• Capture SDK session ID from system init message
• Poll observation queue every 1 second
• Use AbortController for graceful cancellation
• Parse XML with a library (not regex) - suggest fast-xml-parser
• Store observations and summaries using HooksDatabase methods

Architecture Reference

The SDK worker is a synthesis engine that:

• Receives tool observations (not raw data)
• Extracts meaningful insights
• Stores atomic observations in SQLite
• Generates structured summaries at session end

See REFACTOR-PLAN.md (lines 69-119) for the full architecture diagram.

Files to Create

1. src/sdk/worker.ts - Main SDK worker process
2. src/sdk/prompts.ts - Prompt builders
3. src/sdk/parser.ts - XML response parser
4. src/sdk/index.ts - Exports
5. test-phase2.ts - End-to-end tests

Files to Modify

1. src/hooks/new.ts - Spawn worker process
2. package.json - May need to add fast-xml-parser dependency

Testing Strategy

1. Unit tests for prompts (verify prompt structure)
2. Unit tests for parser (verify XML parsing)
3. Integration test for worker (mock SDK responses)
4. End-to-end test (simulate full observation → summary flow)

Success Criteria

• SDK worker runs as detached process
• Worker polls observation queue continuously
• Worker sends observations to Claude SDK
• Worker parses <observation> and <summary> XML correctly
• Worker stores results in database using HooksDatabase
• Worker handles FINALIZE message and exits gracefully
• All tests pass
• No blocking of main Claude Code session

Notes

• Keep hooks fast and non-blocking (they already are)
• SDK worker is fire-and-forget background process
• Use HooksDatabase methods (already implemented in Phase 1)
• Follow the exact prompt formats from REFACTOR-PLAN.md
• Use proper TypeScript types from Agent SDK

---

Start with: Create the SDK prompts module first, then the parser, then the worker. Test each piece before integrating.