run-feedback

Run Feedback

Purpose: Errors from runs evaporate today — a gate fails, gets retried, and the knowledge dies with the session. This skill closes the loop: deterministic capture (retro step, hooks, transcript miner) → LLM triage → deterministic filing into the existing ledgers, both of them thin indexes over record files per known-issues-format: defects to docs/issues/ + docs/KNOWN_ISSUES.md, work-items to docs/backlog/ + docs/BACKLOG.md. Filed issues marked auto_fixable: true feed the /heal-issues harness. Machine state lives under the gitignored .agent/feedback/; the repo-visible trace is the ledgers themselves.

1. Red Flags (Anti-Rationalization)

STOP and READ THIS if you are thinking:

• "I'll append a line to KNOWN_ISSUES.md without creating the issue file" -> WRONG. Lockstep or nothing — the file subcommand writes both or neither, for both ledgers. Never hand-edit one side. Scope of this ban: record files and the index's - **ID** pointer lines. The prefix→category TABLE row in the index preamble is the ONE hand-edit that is yours (§7 step 6) — the script cannot write it.
• "A work-item is small, its body can just live in the backlog index line" -> WRONG. Both indexes are pointer indexes: one line per record, body in docs/backlog/<slug>.md. An inlined body is how a single entry once reached 7 849 characters — unreadable, undiffable, impossible to close in parts. If a repo really runs a one-file backlog (backlog_layout: "flat"), the engine REFUSES a body it would have to flatten; do not "fix" that by squashing your body into one line — switch the layout or file the body where it belongs.
• "This finding is obviously a dup, I'll file it anyway for completeness" -> WRONG. A duplicate goes to --as noise --reason "duplicate of <ID>"; the recurrence is already counted by the fingerprint merge.
• "I'll classify straight from the inbox JSON without running triage" -> WRONG. Dup candidates come from the script (fingerprint + title overlap). Eyeballing misses them.
• "I'll invent severity HIGH / status handled for a new issue" -> WRONG. The write vocabulary is owned by known-issues-format (SEV-2/3/4/LOW; open/fixed/documented/by-design/mitigated/wontfix). Reading tolerates local extensions; writing never emits them.
• "The filing is simple, I'll skip --dry-run" -> WRONG. Dry-run first, always — it previews the allocated ID and the exact index line; a mis-filed ID pollutes a hand-maintained ledger.
• "I'll edit a finding JSON in the inbox by hand" -> WRONG. Only the CLI mutates inbox state; hand edits break fingerprint dedup and the audit trail.
• "The body I just filed came out malformed, I'll fix the issue file" -> WRONG. Ledger writes are create-only, so a filed body cannot be repaired. The gate now catches the two failures that used to land silently (RF-2, fixed): an unterminated fence and a defect body with no ## Reproduction section are refused at exit 4, and --dry-run echoes the rendered record so you can read the body before it lands. Still compose in a REAL file rather than a heredoc you cannot re-read — the gate checks structure, not whether the prose is right. Landed broken → leave it, tell the human.
• "The retro step failed, I should retry it / fail the workflow" -> WRONG. The retro is non-blocking by contract: report one line and move on; it NEVER changes the calling workflow's verdict.