What problem does it solve? AI agent work sessions produce recurring friction—silent no-ops, false completion reports, unnecessary questions—that goes unrecorded and therefore never gets fixed systematically. This Skill captures each trouble as a structured, append-only evidence entry, then supports human-triggered triage that clusters entries by failure shape and proposes hooks, skills, or permission rules as countermeasures. ## Core Features & Use Cases - Structured trouble recording: Writes one Markdown entry per incident using a fixed 7-section template with machine-readable frontmatter (date, summary, skills, repo, canon, paths), validated deterministically by a shell script. - Status ledger management: Tracks triage_status and response_status per entry in TSV ledgers via a Python CLI (sync, summary, validate, update), keeping entry bodies append-only while statuses evolve. - Triage and effectiveness measurement: Clusters untriaged entries by failure shape, proposes warn-only hooks or skill/doc fixes, and measures hook effectiveness from Claude Code and Codex transcripts (correction rate, recurrence, false alarms). - Use Case: After a user points out "why did you ask that when it's already automated?", the agent immediately records the incident with verbatim commands and observed breakage; at the weekly triage, accumulated entries are clustered and matching shapes become hook proposals. ## Quick Start Tell the agent "record that trouble" right after pointing out a workflow mistake, and it will write a structured entry to the configured storage root.
