consreview

Convene a code-review council on the cli-council engine (council review).

Input. If this ran as /consreview <ref|files>, that argument is the target. If it auto-triggered from the conversation, determine the target yourself from context (the change under discussion — current diff, a git ref, specific files, or pasted code). If the change is trivial (a one-line diff, a comment tweak), do NOT convene the council — just review it directly. A run is $0 via the user's subscriptions but burns a few minutes.

What this is

A local code-review council on the cli-council engine (council review, pipeline.run_review_pipeline). One council review call runs the whole shape: every enrolled voice reviews → anonymized peer-ranking → a chairman synthesizes one severity-classified review → verification panels (audit + redteam, each independent and parallel, worst-wins — a rule, not a chairman, so no model can soften the gate) + mechanical checks. $0 (subprocess to the subscription CLIs you authed), stdlib-only.

Not consilium — that runs the same engine on a DECISION (council decide; redteam off, since a recommendation has no ground truth to refute). Code HAS ground truth, so review keeps the redteam.

Voices: whatever you enrolled in council.toml ([council].voices, and the [review] audit/redteam panels). The chairman is also a voice — it reviews, ranks, AND synthesizes — and is kept OFF its own audit panel (no self-approval). See council.example.toml to enrol your own.

Procedure

1. Preflight. council on PATH (repo install.sh, or put bin/ on PATH) + council.toml enrolling your voices — the blessed path is python3 installer/doctor.py enroll claude <voice>...: it re-smokes each and writes a GATED config with the [review] audit + redteam panels already filled in from your voices. Hand-copying council.example.toml works too, but its [review] block ships commented out — leave it commented and the run is unverified (no audit gate, no AUDIT_VERDICT.md). The engine auto-discovers council.toml; an explicit --config <path> to a file that doesn't exist fails loud.

2. Determine target (one council review input form):

   • Empty / current changes → council review (uncommitted git diff HEAD)
   • Git ref → council review <ref>
   • File paths → council review --files <paths...> (reviews file contents, not a diff)
   • Plan / pasted code → write it to a file, council review --prompt-file <path>

   Run from the repo being reviewed so git diff resolves. The engine builds the review prompt itself, with anti-injection fencing (a stray ``` in the diff can't break out).

3. Run the review (background it — a few minutes with slow reasoning voices + 2 panels):

   DIR="$(mktemp -d)/review"; mkdir -p "$DIR"
   council review "${REF:-HEAD}" \
     --scope "Review all material bugs, correctness, security, concurrency, resource and error handling, edge cases, regressions, contradictions, and blind spots. Focal points are hints, not review boundaries." \
     --out "$DIR" 2>"$DIR/engine.log"
   echo "exit=$? · dir=$DIR"; tail -12 "$DIR/engine.log"

   • --files a.py b.py test_a.py instead of "${REF:-HEAD}" to review file contents (include the tests explicitly — there is no auto test-include).
   • --prompt-file <path> to review pasted / hand-built content.
   • --no-redteam requested → append --redteam ' ' (a whitespace-only list = empty panel; keeps audit, skips redteam). --no-verify → append --no-verify (skips BOTH panels → status unverified, no gate).
   • --audit v1,v2 / --redteam v1,v2 override the toml panels ad-hoc.
   • Non-zero exit = the council itself failed (all voices dead / git error) — read engine.log. A degraded status is exit 0: verification flagged something (steps 5–6).
4. Artifacts in $DIR: SYNTHESIS.md · v-<voice>.md (raw reviews) · a-/r-<voice>.md (audit / redteam panelists) · AUDIT_VERDICT.md / REDTEAM_VERDICT.md / MECHANICAL.md · RANKINGS.md · pipeline-status.json (machine-readable gate) · PIPELINE_DEGRADED.md (only when not clean). Read pipeline-status.json first, then SYNTHESIS.md.

5. Read pipeline-status.json + SYNTHESIS.md + AUDIT_VERDICT.md + REDTEAM_VERDICT.md (+ RANKINGS.md) and report:

   • Verdict (SHIP / SHIP-WITH-EDITS / FIX / REWORK) + status (clean / degraded [finding|infra|mixed] / unverified) + reviewers / opinion_errors.
   • Peer leaderboard (RANKINGS.md): a weighting SIGNAL only — never present rank as proof a finding is right/wrong, never drop a finding because of rank.
   • Convergent findings (2+ voices) with verbatim quotes · disagreements.
   • Audit verdict (CLEAN / ISSUES / INVALID, worst-wins) + which panelist(s) set it. Redteam verdict (HOLDS / WEAK / REFUTED; SKIPPED/UNAVAILABLE = no verdict) + per-finding refutations + the missed-claims sweep (redteam's highest- signal output).
   • Severity — SYNTHESIS.md headings `## BLOCKER / ## IMPORTANT / ## CHECK /

     ACCEPT / ## NOISE`. Present in that order.

   status: degraded is NOT a pipeline failure — a gate (audit ISSUES/INVALID, redteam WEAK/REFUTED, or a mechanical fail) flagged the synthesis. degraded_kind: finding = hand-verify before applying; infra = a verifier or the chairman synthesis couldn't complete (re-run); mixed = both.

   Severity protocol: BLOCKER fix, direct evidence · IMPORTANT fix or accept with documented risk · CHECK verify first (grep / query / test), then reclassify · ACCEPT real tradeoff, document · NOISE dismiss with a one-line reason. Spot-check BLOCKER/IMPORTANT claims with specific numbers/paths/lines against the code first.

   Redteam apply: HOLDS eligible · WEAK demote one tier + verify from source · REFUTED drop (re-synth if load-bearing) · MISSED by all voices = new candidate, verify directly · SKIPPED/UNAVAILABLE = proceed on audit + mechanical + voices, note the gap.

6. 🚨 Re-synth gate — MANDATORY when degraded_kind is finding or mixed (i.e. AUDIT_VERDICT is ISSUES/INVALID, or REDTEAM_VERDICT is WEAK/REFUTED): do NOT apply raw SYNTHESIS.md. The engine gates but does not auto-re-synth (an auto-loop would let the panel out-vote a correct minority). Re-synth + primary-source verify: a. Read each v-*.md directly + the flagging a-*.md / r-*.md. b. Rebuild by hand: recount convergences, drop inventions, normalize severity, add missed convergences, fix attributions. c. Verify each claim with a concrete identifier (path, line) — grep code, run a query. d. Apply only verified; drop disproven.

   degraded_kind: infra (a verifier couldn't run, or the chairman synthesis failed) is a DIFFERENT case — NOT a finding. Re-run verification (check the config / voices); do not hand-verify blind.

Dependencies

• cli-council on PATH (bin/council → python3 -m council; Python 3.11+, stdlib only)
  • a council.toml enrolling your voices (per-machine, git-ignored — copy council.example.toml).
• The subscription CLIs you enrolled, each authed (python3 installer/doctor.py smoke <voice> PASSES).

Related

• consilium — decision/opinion council on the same engine (council decide; redteam off — a recommendation has no ground truth to refute).
• commands/consreview.md — the explicit slash-command form of this exact procedure.
• Engine internals: council/{pipeline,review,audit,panel,stages}.py.