Claude Code sends failed tool calls to the PostToolUseFailure event.
ADAM registered only PostToolUse, so it never saw a real failure, and
every tool_error_loop came from ERROR_RE matching successful output.
Handle PostToolUseFailure in adam-observe.mjs through the PostToolUse
detectors. Shape the error string as an is_error response. An interrupt
is an abort, so it counts as a tool event but gets no fingerprint.
Register the event in settings.json.example.
Tests 9b and 9c added: 150 passed, 0 failed.
Claude-Session: https://claude.ai/code/session_014prTNuJnKuwcAd28vUZAnN
- 8 friction signals via lightweight hook (correction, retry_loop, weak_agent,
tool_error_loop, dead_end, edit_churn, build_loop, subagent_dispatch_pattern)
- Deterministic confidence rubric with cross-session evidence gate
- /reflect skill to dispatch the analyst subagent and walk the queue
- Skill overlap detection (prefer skill_edit over skill_new on collision)
- Solution synthesis from transcript context for new skill drafts
- Soft-delete trash, never hard rm
- 18 tests covering all signals
lukaszraczylo