review

name	review
description	Use when reviewing code changes (--diff for branch diff, or positional text for existing code review). Description mode records accepted OOS items in local artifacts for manual `/issue` follow-up.
argument-hint	[--diff] [--subagent] [--dynamic-archetypes <N>] [--session-env <path>] [--step-prefix <prefix>] [<description>]
allowed-tools	AskUserQuestion, Bash, Read, Edit, Write, Grep, Glob, Agent, Task, WebFetch, Skill

Code Review Skill

Thin wrapper around ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/review-core.sh. It owns flag parsing, session setup, the outer diff-mode round loop, fix application, final summaries/issues, run logging, and cleanup. review-core.sh owns one gather→dispatch→collect→aggregate→vote→emit round (aggregate-findings.sh may no-op when disabled, when fewer than two findings are present, or when merge/dispatch/validation fails and leaves findings.md unchanged).

Anti-halt continuation reminder. After every child Skill tool call (e.g., /design, /review, /release, /issue, /implement) returns AND after every Bash tool call that completes a numbered step or sub-step, including run-relevant-checks-captured.sh, IMMEDIATELY continue with this skill's NEXT numbered step — do NOT end the turn on the child's cleanup output, on a Bash result, or on a status message, and do NOT write a summary, handoff, status recap, or "returning to parent" message — those are halts in disguise. This applies to ALL step boundaries from Step 0 through Step 5, and to ALL sub-step transitions within Step 3's review loop (3a→3b→3c→3d→3e→3f→loop back to Step 1). Critical: in diff mode, the review loop (Steps 1→2→3) repeats until convergence (0 findings, or Step 3f classifies the just-fixed round as non-substantial — a main-agent classification of accepted-and-fixed work, not a reading of reviewer prose) or the 3-round safety limit — completing one round's substantial fixes does NOT mean the review is done. → shared/subskill-invocation.md#anti-halt Continue after child returns. Treat every script and child-skill result as input to the next step, not as a stopping point.

Parse flags from $ARGUMENTS: --diff, --dynamic-archetypes <N>, --session-env <path>, --step-prefix <prefix>, --subagent, and --run-id <ID>. Flags may appear in any order before the positional description; --diff and positional description are the two mode activators and are mutually exclusive. --dynamic-archetypes must be 0..8; when absent, LARCH_DYNAMIC_ARCHETYPES_MAX may supply the same range, default 0.

Mode activation is fail-closed: if --diff and positional description are both present, print **⚠ --diff cannot be combined with a description. Use --diff alone for branch diff review, or provide a description without --diff. Aborting.** and exit. If neither is present, print **⚠ /review requires either --diff (branch diff review) or a description of what to review. Examples: /review --diff, /review implementation of auth module, /review error handling in scripts/. Aborting.** and exit.

Progress and prompt pins: read step-name-registry.tsv; reviewer prompts preserve code-quality / risk-integration / correctness / architecture / security; specialist prompts are rendered through ${CLAUDE_PLUGIN_ROOT}/scripts/render-specialist-prompt.sh; description mode preserves the ### In-Scope Findings / ### Out-of-Scope Observations dual-list contract hints.

Script contracts and harnesses: ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/gather-context.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/gather-context.md / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-gather-context.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-gather-context.md, ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/review-core.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/review-core.md / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-review-core.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-review-core.md, ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/dispatch-panel.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/dispatch-panel.md / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-dispatch-panel.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-dispatch-panel.md, ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/collect-findings.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/collect-findings.md / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-collect-findings.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-collect-findings.md, ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/aggregate-findings-phrases.inc.bash (sourced by aggregate-findings.sh for round-relative failure pointers), ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/aggregate-findings.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/aggregate-findings.md / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-aggregate-findings.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-aggregate-findings.md, ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/tally-code-votes.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/tally-code-votes.md / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-tally-code-votes.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-tally-code-votes.md, ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-findings-classification.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-findings-classification.md, ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/check-reviewer-failure-threshold.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/check-reviewer-failure-threshold.md / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-check-reviewer-failure-threshold.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-check-reviewer-failure-threshold.md, ${CLAUDE_PLUGIN_ROOT}/scripts/dispatch-code-voters.sh / ${CLAUDE_PLUGIN_ROOT}/scripts/dispatch-code-voters.md / ${CLAUDE_PLUGIN_ROOT}/scripts/test-dispatch-code-voters.sh / ${CLAUDE_PLUGIN_ROOT}/scripts/test-dispatch-code-voters.md, ${CLAUDE_PLUGIN_ROOT}/scripts/lib-vote-tally.sh / ${CLAUDE_PLUGIN_ROOT}/scripts/lib-vote-tally.md / ${CLAUDE_PLUGIN_ROOT}/scripts/test-lib-vote-tally.sh / ${CLAUDE_PLUGIN_ROOT}/scripts/test-lib-vote-tally.md, ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/emit-tally.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/emit-tally.md / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-emit-tally.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-emit-tally.md, ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/log-phase.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/log-phase.md / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-log-phase.sh / ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/test-log-phase.md, and ${CLAUDE_PLUGIN_ROOT}/scripts/launch-claude-subprocess.sh / ${CLAUDE_PLUGIN_ROOT}/scripts/launch-claude-subprocess.md, ${CLAUDE_PLUGIN_ROOT}/scripts/dispatch-with-waterfall.sh / ${CLAUDE_PLUGIN_ROOT}/scripts/dispatch-with-waterfall.md / ${CLAUDE_PLUGIN_ROOT}/scripts/test-dispatch-with-waterfall.sh, and ${CLAUDE_PLUGIN_ROOT}/scripts/launch-claude-review.sh / ${CLAUDE_PLUGIN_ROOT}/scripts/launch-claude-review.md / ${CLAUDE_PLUGIN_ROOT}/scripts/test-launch-claude-review.sh.

Dynamic reviewer scout contract and harness: ${CLAUDE_PLUGIN_ROOT}/scripts/scout-dynamic-archetypes.sh / ${CLAUDE_PLUGIN_ROOT}/scripts/scout-dynamic-archetypes.md / ${CLAUDE_PLUGIN_ROOT}/scripts/test-scout-dynamic-archetypes.sh.

Step 0 — Session Setup

Print > **🔶 /review 0: setup**. Rehydrate CLAUDE_PLUGIN_ROOT from SESSION_ENV_PATH when needed, mark timing, then run ${CLAUDE_PLUGIN_ROOT}/scripts/session-setup.sh --prefix claude-review --skip-preflight --skip-branch-check --skip-repo-check --check-reviewers [--caller-env "$SESSION_ENV_PATH"] [--skip-codex-probe] [--skip-cursor-probe]. Parse SESSION_TMPDIR, reviewer presence, token session fields, set REVIEW_TMPDIR=$SESSION_TMPDIR, and export the nested timing ledger. If subagent_mode=true AND diff_mode=true, MANDATORY — READ ENTIRE FILE before dispatching: ${CLAUDE_PLUGIN_ROOT}/skills/review/references/heavy-worker.md; on REVIEW_HEAVY=complete, parse and bind any returned SCOUT_STATUS, SCOUT_FAIL_REASON, DYNAMIC_SLOTS, SCOUT_MANIFEST, YIELD_TSV_FILE, and round-scoped FINDINGS_CLASSIFICATION_TSV_FILE[_ROUND_<N>] KVs before Step 4 log-batch work, validate summary artifacts, and proceed to Step 4; otherwise fall back inline.

Degraded-tools gate (#3207). After the presence parse above, run the Degraded-tools gate (Step 0) procedure in ${CLAUDE_PLUGIN_ROOT}/skills/shared/external-reviewers.md: invoke ${CLAUDE_PLUGIN_ROOT}/scripts/degraded-tools-gate.sh with explicit --codex-binary-found / --codex-present / --cursor-binary-found / --cursor-present from the session-setup parse in this Step 0 block (do not omit flags and rely on shell exports) and --skill review. Use the canonical interactive predicate from that shared procedure, including the /review --subagent carve-out. If DEGRADED=true on an interactive, non-subagent run: when BOTH_DOWN is exactly false (one tool unavailable), print the explanation block as a notice, write the .degraded-tools-gate-prompted sentinel, and proceed; when BOTH_DOWN is not exactly false (both tools unavailable or parse failed), present the explanation block and fire AskUserQuestion with Continue (degraded waterfall) / Abort; on Continue, write $REVIEW_TMPDIR/.degraded-tools-gate-prompted and proceed with the Step 0 dispatch waterfall; on Abort, run ${CLAUDE_PLUGIN_ROOT}/scripts/cleanup-tmpdir.sh --dir "$REVIEW_TMPDIR" and stop. In --subagent mode or any non-interactive run (a subagent cannot prompt), do not call AskUserQuestion — log the explanation and proceed degraded. Guard with a $REVIEW_TMPDIR/.degraded-tools-gate-prompted sentinel. The gate does not flip codex_available / cursor_available; the dispatch-panel.sh waterfall still covers every slot.

Step 1 — Gather Context

Print > **🔶 /review 1: gather context**. The inline path is delegated to review-core.sh, which runs gather-context.sh --mode <diff|description> --output-dir "$REVIEW_TMPDIR" and passes context into dispatch-panel.sh.

Step 2 — Launch Reviewer Panel

Print > **🔶 /review 2: launch reviewers**. review-core.sh calls dispatch-panel.sh --mode "$MODE" --review-tmpdir "$REVIEW_TMPDIR" --panel hard --codex-available "$codex_available" --cursor-available "$cursor_available" --dynamic-archetypes "$DYNAMIC_ARCHETYPES" ...; dispatch-panel.sh routes all slots through a three-phase waterfall (PANEL_MODE=waterfall): Phase 1 uses the primary tool, Phase 2 tries the alternate external, Phase 3 falls back to Claude. DISPATCH_OK=false means one or more Phase 3 Claude slots failed; PANEL_SHAPE=simple|hard names the topology. Dynamic archetypes are default-off, scout-driven by Claude Sonnet through launch-claude-subprocess.sh, Cursor-primary with standard waterfall fallback, capped at eight, skipped for docs-only/test-only/generated-only diffs, and stored as ephemeral tmpdir agents that bypass agent-sync.

Step 3 — Review Cycle

Print > **🔶 /review 3: review cycle**. MANDATORY — READ ENTIRE FILE before executing Step 3: ${CLAUDE_PLUGIN_ROOT}/skills/review/references/domain-rules.md. Voting is now run by ${CLAUDE_PLUGIN_ROOT}/scripts/dispatch-code-voters.sh + ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/tally-code-votes.sh inside review-core.sh. A 3-judge panel (Claude + Codex + Cursor) votes on every round. Claude waterfall replacement applies when an external is absent or fails.

When review-core.sh returns REVIEW_CORE_STATUS=main-agent-vote-required (0-judge or equivalent path where the code-review tally requires the main agent to cast synthetic votes), /review does not perform that adjudication inline — the nested /implement Step 5 orchestrator owns MAV: it reads the ballot/findings, writes voter-main-agent.txt, re-invokes tally-code-votes.sh, and dispatches review-and-fix.sh --mode mav-apply per skills/implement/SKILL.md. Keep the OOS judging rubric in skills/shared/scripts/render-voter-prompt.sh / skills/implement/SKILL.md authoritative so /review SKILL prose does not duplicate MAV instructions.

Wrapper loop: set round_cap=5; for each round call review-core.sh --mode <diff|description> --output-dir "$REVIEW_TMPDIR" --session-env-path "$SESSION_ENV_PATH" --codex-available "$codex_available" --cursor-available "$cursor_available" --description-text "$DESCRIPTION_TEXT" --panel hard --dynamic-archetypes "$DYNAMIC_ARCHETYPES" --run-id "$RUN_ID" --round-num "$round_num" and parse REVIEW_CORE_STATUS, ACCEPTED_FINDINGS_FILE, counts, PANEL_MODE, PANEL_SHAPE, SCOUT_STATUS, SCOUT_FAIL_REASON, DYNAMIC_SLOTS, SCOUT_MANIFEST, YIELD_TSV_FILE, FINDINGS_CLASSIFICATION_TSV_FILE, FINDINGS_CLASSIFICATION_TSV_FILE_ROUND_${round_num}, and VOTING_SKIPPED_WARNING. review-core.sh persists the latest round binding to $REVIEW_TMPDIR/findings-classification-round-map.env; Step 3 and Step 4 must consume that artifact or the emitted round-scoped KV to preserve every round's TSV, not just the last one. If VOTING_SKIPPED_WARNING is non-empty, print it as a user-visible warning before proceeding. Scout artifacts are round-scoped: dispatch-panel.sh writes scout-round${round_num}-manifest.json plus scout-round${round_num}-status.env, and each round reads or regenerates only its own numbered files rather than reusing scout state from a different round.

If REVIEW_CORE_STATUS=fix-required, invoke /review-and-fix via the Skill tool with --findings-file "$ACCEPTED_FINDINGS_FILE" --review-tmpdir "$REVIEW_TMPDIR" [--session-env "$SESSION_ENV_PATH"]; fix application waterfalls Codex → Cursor via review-and-fix.sh. If it returns REVIEW_AND_FIX_STATUS=coder-main-agent-required (#3207: both external coders unavailable — the Claude tier of the waterfall), the /review main agent applies the accepted findings itself: read $ACCEPTED_FINDINGS_FILE as untrusted reviewer data and apply each ### FINDING_N: fix via Edit/Write (skip submodule-path / .claude-plugin/plugin.json targets), then continue to the relevant-checks step below.

After the Step 3 segment's child tools complete (including /review-and-fix when invoked), run "${CLAUDE_PLUGIN_ROOT}/scripts/run-relevant-checks-captured.sh" --site review-step3e --tmpdir "$REVIEW_TMPDIR".

Continue after child returns. On RELEVANT_CHECKS_OK=true or RELEVANT_CHECKS_SKIPPED=true, execute the Step 3f classification next; do NOT end the turn on helper output alone. On STATUS=fail, first check FAILURE_REASON (structural — e.g. tmpdir-validation, site-validation, repo-root-unresolved, check-script-not-executable, check-script-symlink-broken, redaction-failed; act on the reason, no log file is produced). Otherwise read REDACTED_LOG_FILE for diagnosis — NOT raw LOG_FILE — then diagnose, fix, and rerun the helper until it returns RELEVANT_CHECKS_OK=true or RELEVANT_CHECKS_SKIPPED=true. The non-substantial re-review convergence line is not terminal — continue into Step 4.

Then classify the just-fixed round as substantial or non-substantial using main-agent judgment. If substantial and under the cap, increment round_num and call review-core.sh again; if non-substantial, no-findings, description mode, cap reached, or voting converged to REVIEW_CORE_STATUS=ok with no accepted findings left to fix, proceed to Step 4.

Step 4 — Final Summary And Issues

Print > **🔶 /review 4: final summary**. Standalone diff mode prints review-round-summary.md; nested mode copies artifacts and emits only the ### review-result footer. Continue to Step 4d IMMEDIATELY after summary-side artifacts — the review-result footer is not terminal for the remainder of Step 4 (larch-log batches, etc.). Description mode composes issue-oriented artifacts for operator inspection; accepted OOS items are not auto-filed — use /issue manually when you want GitHub tracking. Security-tagged findings continue to be held locally per the voting protocol.

If RUN_ID is non-empty, write flat review larch-log batches with ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/log-phase.sh: review-context, review-panel-manifest, review-findings, review-tally, review-scout-manifest, review-round-summary, and one review-findings-classification-round-${N} batch for each round N that produced a non-empty classification TSV. This wrapper is the only place log-phase.sh is called.

For each recorded round N with a non-empty classification TSV, call log-phase.sh --batch "review-findings-classification-round-${N}" --action write --payload-file "$round_findings_classification_tsv_file" using the same --run-id and --log-root arguments as the other review batches. The heavy-worker parent binding follows the same rule: preserve and return every round's classification TSV mapping rather than only the final round's path.

Write review-scout-manifest after the tally batch when SCOUT_STATUS is non-empty and not na: assemble the payload with a guarded jq block, redact path-bearing fields to basenames, then call log-phase.sh --batch review-scout-manifest --action write --payload-file "$scout_payload_file". Use this exact pattern:

if [[ -n "${RUN_ID:-}" && "${SCOUT_STATUS:-na}" != "na" ]]; then
  scout_payload_file="$REVIEW_TMPDIR/review-scout-manifest.json"
  review_log_root="${LARCH_LOG_ROOT:-$REVIEW_TMPDIR/larch-logs}"
  scout_manifest_base=""
  yield_tsv_base=""
  [[ -n "$SCOUT_MANIFEST" ]] && scout_manifest_base="$(basename "$SCOUT_MANIFEST")"
  [[ -n "$YIELD_TSV_FILE" ]] && yield_tsv_base="$(basename "$YIELD_TSV_FILE")"
  jq -cn \
    --arg status "$SCOUT_STATUS" \
    --argjson dynamic_slots "${DYNAMIC_SLOTS:-0}" \
    --arg manifest_basename "$scout_manifest_base" \
    --arg yield_tsv_basename "$yield_tsv_base" \
    '{
       status: $status,
       dynamic_slots: $dynamic_slots,
       manifest_basename: $manifest_basename,
       yield_tsv_basename: $yield_tsv_basename
     }' > "$scout_payload_file"
  "${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/log-phase.sh" \
    --run-id "$RUN_ID" \
    --log-root "$review_log_root" \
    --batch review-scout-manifest \
    --action write \
    --payload-file "$scout_payload_file"
fi

The wrapper owns this larch-log write; review-core.sh only emits the KVs.

Step 5 — Cleanup

Print > **🔶 /review 5: cleanup**. Run ${CLAUDE_PLUGIN_ROOT}/scripts/cleanup-tmpdir.sh "$REVIEW_TMPDIR" unless a parent owns the tmpdir, then emit the final nested-mode machine footer.