feat: session_end() method and end_session MCP tool

[helinakdogan]

Summary

• sdk/magnet/client.py — BehavioralMemory.session_end():

  • Takes last max_messages=20 turns from the session
  • Calls LLM with prompt: "Extract key decisions, learned preferences, and important context… Format as bullet points."
  • Falls back to _auto_summarize() if LLM call fails
  • Stores summary as episode with importance=0.9 in EpisodicStore (Layer 2)
  • Also writes to Redis key magnet:session_summary:{tenant_id} with 30-day TTL
  • Returns {"status": "ok", "summary": str}
  • Also stored self._redis = redis_client in __init__ (needed for the TTL write)

• sdk/magnet/mcp_server.py — end_session MCP tool:

  • Description: "Summarize and save current session to memory"
  • Input: user_id, project_id, messages (array)
  • Output: {"status": "ok", "summary": string}
  • Delegates to session_end() via asyncio.to_thread

Replaces Claude Code's compact feature — automatically saves what matters, forgets the rest.

Test plan

• Call session_end(user_id, project_id, messages) — returns {"status": "ok", "summary": "..."} with bullet points
• Redis key magnet:session_summary:{tenant_id} exists with 30-day TTL after call
• EpisodicStore.recall() returns the session summary as a top result (importance=0.9)
• LLM failure → graceful fallback to _auto_summarize, no exception raised
• MCP end_session tool callable from Claude Code / Cursor via stdio

🤖 Generated with Claude Code