Memory capture stores every tool output empty: PostToolUse reads tool_output, Claude Code sends tool_response #164

Summary

handle_post_tool_use (src/context_engine/memory/hooks.py) reads the tool result from tool_output or tool_output_json only:

tool_output = data.get("tool_output") or data.get("tool_output_json") or ""

Claude Code's PostToolUse hook payload names that field tool_response, and the installed hook script forwards stdin unchanged. Every live-captured tool event is therefore stored with an empty raw_output. The request succeeds, so nothing signals the loss.

Evidence

On one project's memory.db, every tool_event_payloads.raw_output written between 2026-09-23 and 2026-10-04 by the live hooks, and by a transcript-replay script that also used Claude Code's tool_response name, is empty (~11,000 rows). The only non-empty rows are ones replayed after that script switched to sending tool_output.

Checked against v0.4.26 and current main (7507115). A code search finds no mapping of tool_response anywhere.

Suggested fix

Accept Claude Code's field name as well, e.g. data.get("tool_output") or data.get("tool_output_json") or data.get("tool_response") or "", plus a handler test that posts a real Claude Code PostToolUse payload.

Related: #163 (another silent memory-capture loss, the serve.port rendezvous).

🤖 Generated with Claude Code