prd-triage

Set dispositions on the active PRD's findings. /prd-approve is blocked by the runner's findings gate until every finding has a non-pending disposition.

Dispositions:

• accepted — will be fixed in this PRD. No rationale needed.
• rejected — will NOT be addressed. --rationale REQUIRED (why not).
• deferred — tracked but out of scope for this PRD. --rationale REQUIRED (where).
• pending — initial state; can be used to revert a premature disposition.

Do not edit the JSONL file by hand. The writer enforces the rationale rule and stamps resolved_at atomically with the disposition change.

Steps:

1. Resolve the active PRD id:

PRD_ID=$(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/prd_runner.py" status | \
  python3 -c "import sys,json; s=json.load(sys.stdin); print(s.get('prd_id') or '')")

If PRD_ID is empty, tell the author no PRD is active and stop.

1. Print the cross-PRD advisory. This surfaces prior findings from sibling PRDs that were already rejected or deferred and closely match a pending finding here, so the author triages with the prior rationale in view:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/findings_writer.py" advisory "$PRD_ID"

The advisory is deterministic (token-shingle similarity, no LLM) and printed by the runner, not assembled here. It is informational only: it never blocks, never auto-dispositions, and any xref failure is swallowed so triage proceeds. Empty output means no sibling PRD settled anything similar.

1. List pending findings:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/findings_writer.py" list "$PRD_ID" --only-pending

Show them to the author. For each, ask what disposition they want. Do NOT guess the author's intent; pending findings are theirs to resolve.

1. Run the blind judge BEFORE the author decides, once per finding. This is what makes a triage a calibration case: evaluate scores a receipt only when it carries BOTH a judge prediction and a human decision.

JUDGE_RUN="$(mktemp -t judge-run-XXXXXX).json"
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/judgment_compiler.py" \
  judge --prd "$PRD_ID" --finding <finding-id> --output "$JUDGE_RUN"

The judge runs with tools OFF and sees only the frozen context packet, which carries no disposition or rationale. Do NOT show its prediction to the author before their disposition is set. An author who sees the prediction and agrees inflates measured agreement, and the calibration set stops measuring anything. Report it afterwards, or not at all.

If the judge call fails, say so and continue WITHOUT --judge-run. A model outage must not wedge triage; it costs one calibration case, and a receipt with an honest missing judge is worth more than a blocked author.

1. Apply each disposition. Pass the structured reason code and its evidence so the judgment receipt captures a machine-checkable decision, not prose:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/findings_writer.py" \
  set-disposition "$PRD_ID" <finding-id> <accepted|rejected|deferred> \
  [--rationale "<text>"] \
  [--reason-code <valid-fix-now|already-remediated|duplicate|owned-by-other-prd|scope-removed|out-of-scope|superseded|defer-dependency|defer-ordering|invalid-finding|insufficient-context|needs-human>] \
  [--evidence <finding:prd/id | issue:id | prd:id | receipt:id | commit:sha | test:path | scope:path#section | judgment:receipt-id | spillover:id>]...
  [--judge-run "$JUDGE_RUN"]

Every adjudication appends an immutable receipt to .prd-os/judgments.jsonl (the Judgment Compiler ledger — judgment_compiler.py, enforced in code, not here). Evidence-requiring codes (duplicate, already-remediated, owned-by-other-prd, scope-removed, out-of-scope, superseded) are REFUSED by the writer without a stable --evidence reference; the findings file stays untouched on refusal. Omitting --reason-code still works and records an honest null — prefer passing it: receipts with codes are what the calibration evaluator and policy-candidate detector can learn from.