Incident Memory

incident-memory

Every chain writes here. Including the ones that failed.

A store containing only successes teaches classify that everything is known and everything works. The failures — the unproven causes, the exhausted attempt bounds, the loops that could not be built — are what make the known/novel judgement honest.

Store: .harness/incidents.jsonl, append-only, one JSON object per line.

Phase 1 — Write the record

{
  "id": "INC-<UTC timestamp>",
  "ts": "2026-08-01T03:40:00Z",
  "symptom": "what was observed, in the reporter's words",
  "loop": "the exact command from prove-the-failure",
  "loop_built": true,
  "contained": "what was stopped, and its undo",
  "left_alone": "what was deliberately not stopped, and why",
  "cause": "one sentence, plain language",
  "cause_proven": true,
  "evidence": "file:line",
  "rejected": ["hypotheses ruled out, and what ruled them out"],
  "classification": "known|novel",
  "blast_radius": "low|high",
  "blast_reasoning": "what else changes, what that could trigger",
  "attempts": 1,
  "approaches": ["how each attempt differed in kind"],
  "diff": "path to the proposed diff, or null",
  "reviewer_model": "which model reviewed",
  "reviewer_verdict": "PASS|FAIL|ABSENT",
  "reviewer_findings": ["what it flagged"],
  "verified": true,
  "immunisation": "the guard added, or the fence change proposed-unapplied, or none-possible",
  "outcome": "fixed|gave-up|stopped-at-gate|breaker-tripped",
  "gated_on": "what a human must decide, if anything"
}

outcome is the load-bearing field. Its four values are all legitimate:

• fixed — verified green, immunised
• gave-up — attempt bound exhausted; the diagnosis is the deliverable
• stopped-at-gate — classify said novel or high-radius, or a gated action was required
• breaker-tripped — the chain halted; a human was called

Phase 2 — Make it findable

classify matches on cause, not symptom. Write the cause so a later match is possible:

• State the mechanism, not the manifestation. "Workflow pushes directly to a protected branch" not "the drain job failed".
• Name the surface concretely — repo, file, job, database.
• Avoid words that only make sense with this incident's context in mind.

Search is grep over the JSONL. Do not build an index; the file is small and will stay small, and an index is a second thing that can be wrong.

Completion

• Record appended, valid JSON on one line
• cause written as a mechanism, matchable by a future chain
• outcome set to one of the four values
• Failures recorded with the same care as successes
• gated_on populated whenever a human must act

Stop conditions

Never rewrite or delete a prior record. Append-only. A correction is a new record referencing the old id.

Never mark outcome: fixed without verified: true. Applied is not verified; verified means verify ran and pasted both outputs.

This file is not on the self-modification list, but treat it as evidence. An agent editing its own incident history to make a failure look known is the same class of act as clearing its own denials — if you find yourself wanting to adjust a past record so the current fix qualifies, that is the signal to stop.

Next

End of chain. If outcome is anything other than fixed, the report goes to the founder with the diagnosis attached.