triage

name

triage

description

Morning read-only triage heartbeat for the build (project specifics from the profile). Reads the tasks file, git history, and GitHub issues/PRs, then writes a fresh state snapshot to the out-of-repo inbox. Surfaces the next unblocked task, stale checkboxes, open PRs needing review, blocked/owner items, and upcoming constitution risks. Use when the user says "triage", "morning triage", "what's the state of the build", or when fired on a schedule. NEVER edits code, opens issues/PRs, or commits.

/triage — Claude Code binding

The workflow logic is runtime-neutral and lives in .claude/workflow/triage.md — read that file now and execute it. Mapping of its abstract [roles]:

Neutral role	Claude Code mechanism
[workflow] (next-task)	the /next-task skill the human runs later
[headless run]	claude -p "/triage" (the scheduled heartbeat)
[environment block]	.claude/skills/next-task/SKILL.md → "This environment's concrete forms" (the single copy of the OS/shell/CLI gotchas — e.g. the gh PATH fallback)
inbox write	the Write tool (out-of-repo path; allowed on main because it's outside the repo root)

This procedure is read-only on the repo: no Edit/Write/MultiEdit under the repo root, no git add/commit/push, no gh issue/pr create. The PreToolUse guard (.claude/hooks/guard.sh) is the deterministic backstop.

Profile read — packet first (spec 007 US3.AC3). Read .claude/PROJECT.compact.md by default for the routing facts (paths, conventions, required check, autonomy status, blocked list — drift-checked by .claude/hooks/compact-packet-drift.sh in CI). The snapshot's look-ahead sections need profile facts the packet deliberately omits — the constitution-watch list and the maker-eval records path — so escalate to the full .claude/PROJECT.md explicitly for those sections (a named escalation case, not a default full read); the full profile stays the source of truth.

Gate-trends effective-fix rate — concrete instantiation (triage.md §2 / §4)

triage.md's "Gate trends" section defers the effective-fix rate's concrete recipe to the adapter (constitution P3 — a non-scorer re-runs a deterministic recipe, never a model estimate). This is that recipe. All reads are read-only; triage writes nothing to the stream.

• The recipe. bash .claude/hooks/effective-fix-rate.sh <telemetry-stream> [--since <ISO>] [--until <ISO>] reads the profile → "Paths" → Telemetry stream (the same <triage inbox dir>/<repo-basename>-telemetry.jsonl the other gate-trends reads resolve), filters gate-run records to the snapshot window by their timestamp (inclusive bounds; ISO-8601 UTC sorts lexicographically, so it is a plain string compare), and prints one compact JSON object — never appending to the stream. Malformed lines are skipped and surfaced as .skipped_malformed.
• Render the numerator and denominator, never a bare percentage (US9.AC2). The object carries .numerator (flips), .denominator (FAIL-triggered re-dispatches), .pct (a convenience percentage), and .by_auditor{<auditor>:{numerator,denominator}}. Render the window line as <numerator>/<denominator> (<pct>%) and one <auditor>: <n>/<d> per by_auditor entry.
• The two explicit empty states map to .state: "no-data" → the "no data yet — telemetry stream absent/empty" line; "no-fix-rounds" → the "no fix rounds in window" line (gate-run records but denominator zero). "rate" renders the numbers — including a genuine 0-of-N (.numerator 0 with .denominator > 0), which is not the "no-fix-rounds" state.
• Observe-only (telemetry.md law / US9.AC4): the recipe reads the stream and prints; it feeds no gate, tier, guard, or selection path, and writes nothing.

Verification-machinery freshness — concrete instantiation (triage.md §1.7 / §2)

triage.md's PROBES-STALE check needs two adapter facts the neutral doc defers:

• Recompute the current [guard]-machinery fingerprint using the recipe defined once in .claude/adapters/claude-code-probes.md → "Probe-run fingerprint" (the guard-script content hash + the matcher-only hook-wiring hash), rendered guard=<sha7> wiring=<sha7>. Reuse that recipe — never re-derive it here — so a future change to what the fingerprint covers stays a one-place edit and the comparison hashes exactly what the probe runs hashed. Both halves are content hashes: recomputing is read-only, no repo write.
• The recorded baseline is the most recent non-placeholder row of that same file's "## Probe results" table: its Fingerprint cell is the value to compare against, and its date is the age source. No non-placeholder row ⇒ the neutral "no probe run recorded yet" state.

Equal ⇒ machinery current; different ⇒ PROBES-STALE (report the last run's age either way). The GUARD-SILENT check needs no extra adapter fact — it reads the telemetry stream (profile → "Paths" → Telemetry, the source triage.md §1.6 already reads) and tells gate-run from [guard] evaluation records by the neutral record field (workflow/telemetry.md).

triage.md's CORPUS-STALE check (auditor-liveness corpus currency) needs the auditor analog of the two PROBES-STALE facts:

• Recompute the current reviewer-spec fingerprint using the recipe defined once in .claude/skills/auditor-liveness/SKILL.md → "The reviewer-spec fingerprint" (the git hash-object over the four *-auditor.md specs — including spec-quality-auditor.md — + the evasion register, rendered specs=<sha7>). Reuse that recipe — never re-derive it here — so what the fingerprint covers stays a one-place edit. It is a content hash: recomputing is read-only.
• The recorded baseline is the most recent non-placeholder full-scope row of that same file's "## Corpus-run results" table — skip partial:<fixture-id> rows, since a scoped single-fixture diagnostic has not re-confirmed every auditor and so never serves as the freshness baseline (workflow/auditor-liveness.md → "Re-run policy"). Its Reviewer-spec fingerprint cell is the value to compare against, and its date is the age source. No non-placeholder full row ⇒ the neutral "no corpus run recorded yet" state.

Equal ⇒ auditors confirmed-live as of that run; different ⇒ CORPUS-STALE (report the last run's age either way). All three checks mutate nothing, honoring the read-only-on-the-repo contract above.

Maker-eval surfacing — concrete instantiation (triage.md §1.8 / §2)

triage.md's "Maker eval" section needs the adapter facts the neutral doc defers — where the records live, how to recompute the current maker-behavior fingerprint, and the record fields the differential reads. All reads here are read-only (honoring the contract above); triage never runs the eval or writes a record.

• The records + the packet link. The channel is the profile → "Paths" → Maker-eval records (out-of-repo, beside the telemetry stream — the same path .claude/hooks/maker-eval-emit.sh resolves). Read <channel>/records.jsonl — one append-only record per (corpus task × maker tier) per run, each carrying run_id, task_id, maker_tier, the fingerprint {maker_behavior, judge_identity, eval_instrument} object, the per-dimension dimensions verdicts (each with its lifecycle), the overall verdict, and a relative packet path. A flagged regression's packet link is that packet field, which resolves only inside the channel (the US2.AC3 fence, T804) — render it as the path, not a repo link.
• The two runs to difference. Group records.jsonl by run_id; a run is complete iff every corpus task has a record under it at every maker tier — the read-only completeness oracle is bash .claude/hooks/maker-eval-emit.sh complete --run-id <id> (exit 0 = complete, exit 3 = incomplete). The last complete run and the prior complete run are the differential's two inputs; an incomplete latest run (including a single-tier run) is never a silent baseline.
• Recompute the current maker-behavior fingerprint with the single-source recipe bash .claude/hooks/maker-eval-emit.sh fingerprint → the .maker_behavior field of the printed triple-fingerprint object. Reuse that recipe — never re-derive it here — so what the maker-behavior surface covers (the model rows + the instruction/runtime surfaces) stays a one-place edit; it is a content hash, so recomputing is read-only.
• MAKER-EVAL-STALE compares that recomputed .maker_behavior against the last run's recorded fingerprint.maker_behavior. Equal ⇒ current; different ⇒ MAKER-EVAL-STALE (an eval is overdue). This is the maker analog of PROBES-STALE/CORPUS-STALE above.
• JUDGE-CHANGED / INSTRUMENT-CHANGED compare the two runs being differenced (not the live tree): the last complete run's fingerprint.judge_identity / fingerprint.eval_instrument against the prior complete run's. A difference makes the pair not-comparable and suppresses the regression call.
• JUDGE-MISCALIBRATED reads the recorded judge↔owner agreement and its floor from the run's records (added by US1.AC5 / T806 — the calibration set and agreement computation). Until those fields are emitted, no agreement is present ⇒ render the neutral "no agreement recorded yet" state. All of this stays observe-only (maker-eval.md → "Observe-only"): the section feeds no gate, tier, guard, or selection path.