| name | critical-review |
| description | Parallel adversarial review — dispatches reviewer agents on distinct challenge angles, then synthesizes findings with an Opus agent. Use when user says "critical review", "pressure test", "adversarial review", or "challenge from multiple angles". Auto-triggers in the lifecycle for Complex + medium/high/critical features before spec approval. |
| when_to_use | Use when you want to stress-test a plan, spec, or research artifact before committing ("poke holes in the plan"). Different from /devils-advocate — devils-advocate runs inline in the current agent context for a lightweight solo deliberation; critical-review dispatches parallel sub-agents and synthesizes findings. |
| argument-hint | [<artifact-path>] |
| inputs | ["artifact-path: string (optional) — path to plan.md, spec.md, or research.md to review; if omitted, auto-detect from current lifecycle"] |
| outputs | ["Synthesis prose presented in conversation","Optional residue write at cortex/lifecycle/{feature}/critical-review-residue.json"] |
| preconditions | ["Run from project root","Artifact path resolves to an existing markdown file"] |
Critical Review
Dispatches one fresh reviewer agent per angle in parallel — full independence, no anchoring to the reasoning that produced the artifact — then synthesizes with an Opus agent.
Step 1: Find the Artifact
If a lifecycle is active, read the most relevant artifact (cortex/lifecycle/{feature}/plan.md → spec.md → research.md, in that order); otherwise use conversation context. If nothing is clear enough to challenge, ask "What should I critically review?" first.
Step 2: Review Setup and Dispatch
Step 2a: Load Domain Context
Assemble a ## Project Context block for the reviewer prompts from:
cortex/requirements/project.md — the Overview section (or the first top-level section if none is labeled Overview), up to ~250 words.cortex/lifecycle.config.md — a valid type: field (present, non-empty, uncommented) as a one-line prefix **Project type:** {type} before the overview.cortex/requirements/glossary.md — the ## Language section verbatim, inline. Read only that section — NOT ## Relationships, ## Example dialogue, or ## Flagged ambiguities (those approach existing-reasoning territory). Skip silently when absent.
None of these available → omit the ## Project Context section entirely — no empty placeholder.
Requirements loading: deliberately exempt. Critical-review narrows its context to cortex/requirements/project.md's Overview and the glossary ## Language section, skipping the tag-based requirements-loading protocol other skills use — broader project context (priorities, area tags, decisions) would dilute the fresh-eyes stance. Vocabulary is admitted because it's definitional, not reasoning-shaped. Do not "fix" this by wiring tag-based loading into the dispatch path.
Step 2a.5: Pre-Dispatch (atomic path + SHA pin)
Fuse path validation and SHA-256 computation before any dispatch via cortex-critical-review prepare-dispatch <artifact-path> [--feature <name>], binding {artifact_path}/{artifact_sha256} for every downstream dispatch site. Full invocation contract and exit-code routing: ${CLAUDE_SKILL_DIR}/references/verification-gates.md.
Step 2b: Derive Angles
The orchestrator (main conversation context) derives the challenge angles from the artifact — distinct, artifact-specific, and picked to reveal real problems for this artifact, not generic category labels. Count and acceptance criteria: ${CLAUDE_SKILL_DIR}/references/angle-menu.md.
Step 2c: Dispatch Parallel Reviewers
Dispatch one general-purpose agent per angle as a parallel Task sub-task, all simultaneously, using the canonical reviewer prompt from ${CLAUDE_SKILL_DIR}/references/reviewer-prompt.md verbatim, substituting {artifact_path}, {artifact_sha256}, {angle name}, {angle description}, and the Step 2a Project Context block at runtime. The template carries a ~40-turn cap — on hit the reviewer stops investigating and returns what it has, and its partial findings route through the Partial branch below.
Failure Handling
- Partial (some agents succeed, some fail) → proceed to Step 2d with the successful findings only, prefixing the synthesis per the partial-coverage rule in
${CLAUDE_SKILL_DIR}/references/verification-gates.md.
- Returned nothing (an agent died loudly, or idled without reporting) → treat it as a failed agent and take the Partial branch with the reviewers that did return; never wait on it indefinitely. All agents returning nothing is the Total branch.
- Total (all fail) → fall back to one general-purpose agent with the canonical fallback prompt from
${CLAUDE_SKILL_DIR}/references/fallback-reviewer-prompt.md (substituting {artifact_path}, {artifact_sha256}). Output its result directly with no Step 2d synthesis, prefixed Note: parallel dispatch failed, falling back to single reviewer, then proceed to Step 3.
Step 2c.5: Sentinel-First Verification Gate
After parallel reviewers return, run the two-phase verification gate (sentinel check, then envelope extraction) before Step 2d synthesis. If every reviewer is excluded (all exit-3), surface verbatim and do NOT synthesize: All reviewers excluded — drift or Read failure detected; critical-review pass invalidated. Re-run after resolving concurrent write source. Full route table, record-exclusion contract, and Phase 2 schema assertions: ${CLAUDE_SKILL_DIR}/references/verification-gates.md.
Step 2d: Opus Synthesis
After parallel reviewers (or the successful subset) clear Step 2c.5, resolve the synthesizer model by running cortex-resolve-model --role synthesizer (no --criticality flag and no lifecycle-state read — the standalone path may have no lifecycle session, so a missing state must never block synthesis); on nonzero exit, halt and escalate rather than substitute a model. Dispatch one synthesizer agent with the resolved model and the canonical prompt from ${CLAUDE_SKILL_DIR}/references/synthesizer-prompt.md verbatim, with {artifact_path}, {artifact_sha256}, {a_to_b_rubric_path} ← the resolved absolute path of ${CLAUDE_SKILL_DIR}/references/a-to-b-downgrade-rubric.md (the synthesizer Reads it — do not Read or inline the rubric here), and the reviewer-findings payload substituted at runtime. The prompt directs the synthesizer to Read the artifact once at start and emit SYNTH_READ_OK: <path> <sha> in output before per-finding analysis.
The synthesizer applies the A→B downgrade rubric (Read by path via {a_to_b_rubric_path}) to each A-class finding's "fix_invalidation_argument" field.
Step 2d.5: Post-Synthesis (atomic SHA verification)
Pipe the synthesizer's full output through cortex-critical-review check-synth-stable --feature <name> --expected-sha <hex> --artifact-path <resolved_path> before surfacing anything or proceeding to Step 2e. Full contract and exit-code routing: ${CLAUDE_SKILL_DIR}/references/verification-gates.md.
Step 2e: Residue Write
After synthesis (or a Step 2c.5 pass-through), atomically write any B-class findings to cortex/lifecycle/{feature}/critical-review-residue.json via the cortex-critical-review-write-residue console-script; skip silently when zero B-class remain. Resolver, payload schema (R4), and zero/multiple-match gating: ${CLAUDE_SKILL_DIR}/references/residue-write.md.
Step 3: Present
Output the review result directly. Do not soften or editorialize.
Step 4: Apply Feedback
Immediately after presenting the synthesis, work through each objection independently — do not wait for the user.
- Apply when the fix is unambiguous and confidence is high.
- Dismiss when the artifact already addresses the objection or it misreads stated constraints.
- Ask when the fix involves user preference, a scope decision, or genuine uncertainty.
Default ambiguous to Ask. Anchor-checks: dismissals must point to artifact text, not memory; resolutions must rest on new evidence, not prior reasoning. For any empirical claim (latency, file size, blast radius, baseline behavior), run the actual measurement (time, wc -c, grep) before classifying Apply/Dismiss — re-reading the artifact text is not new evidence.
Then: (1) re-read the artifact in full; (2) write the updated artifact with all Apply fixes incorporated, preserving everything untouched; (3) present a compact summary — Apply bullets describe the direction of change (not the objection text), each opening with one of strengthened / narrowed / clarified / added / removed / inverted; a single Dismiss: N objections line (omit when N = 0); and any Ask items consolidated into one message.
