| name | critical-review |
| description | Parallel adversarial review — dispatches reviewer agents on distinct challenge angles, then synthesizes their findings. 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>] |
Critical Review
One fresh reviewer agent per angle, dispatched in parallel — no anchoring to the reasoning that produced the artifact — then a synthesis pass.
Step 1: Find the artifact
If a lifecycle is active, take the most relevant of cortex/lifecycle/{feature}/plan.md → spec.md → research.md, in that order; otherwise use conversation context. Nothing clear enough to challenge → ask "What should I critically review?" first. Resolve it to an absolute path — reviewers read that literal path.
Step 2: Derive angles
Derive the challenge angles from the artifact yourself, in this conversation. Default 2; escalate to 3–4 only when criticality is high/critical, or the artifact introduces claims its inputs lacked (mechanisms, measured figures, or verification approaches absent from the spec or research it derives from).
Each angle must cite a specific section, claim, assumption, or design choice in this artifact, and no two may re-phrase the same concern. "Fragile assumptions" alone is not an angle; "the retry logic in §3 assumes idempotent endpoints, which breaks for the payment webhook in §5" is. Architectural risk, integration risk, and scope creep are a diversity nudge, not a checklist — weight toward the failure modes of whatever domain the artifact lives in.
Step 3: Assemble project context
Build a ## Project Context block for the reviewer prompts from cortex/requirements/project.md's Overview (~250 words), a **Project type:** {type} prefix from cortex/lifecycle.config.md when it carries a valid type:, and cortex/requirements/glossary.md's ## Language section verbatim. None available → omit the section entirely, no placeholder.
Deliberately narrow. Critical-review skips 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. Read only ## Language, not ## Relationships, ## Example dialogue, or ## Flagged ambiguities. Do not "fix" this by wiring tag-based loading into the dispatch path.
Step 4: Dispatch reviewers
One general-purpose agent per angle, all in parallel, using ${CLAUDE_SKILL_DIR}/references/reviewer-prompt.md verbatim with {artifact_path}, {angle name}, {angle description}, and the Step 3 context block substituted.
Extract each reviewer's envelope: split on the last <!--findings-json--> line, json.loads the tail, and assert top-level angle: str and findings: list, each finding carrying class ∈ {A,B,C}, finding, and evidence_quote. The envelope is the reviewer's whole deliverable, so a malformed one leaves nothing to salvage — warn ⚠ Reviewer {angle} emitted malformed JSON envelope ({reason}) and drop it.
Some reviewers failing → synthesize from the rest and prefix "N of M reviewer angles completed." Never wait on a silent agent. All failing → dispatch a single general-purpose agent to derive 3–4 angles itself and produce the same output shape, prefix Note: parallel dispatch failed, falling back to single reviewer, and skip synthesis.
Step 5: Synthesize
Dispatch one synthesizer with ${CLAUDE_SKILL_DIR}/references/synthesizer-prompt.md verbatim, with the reviewer findings substituted. Synthesis is the judgment step of this skill — weigh the model accordingly.
Step 6: Write B-class residue
With ≥1 B-class finding, write the sidecar the morning report reads — the verb resolves the feature from the session id itself:
cortex-critical-review-write-residue --session-id "$LIFECYCLE_SESSION_ID" <<< "$PAYLOAD_JSON"
Payload: ts, feature, artifact, synthesis_status (ok|failed), reviewers: {completed, dispatched}, and findings — each {class: "B", finding, reviewer_angle, evidence_quote}. Zero B-class findings → skip the call, no file, no note. state: no-context (zero matches) or state: ambiguous (multiple active lifecycle sessions matched) → nothing written; relay the returned note verbatim. Synthesis failure still writes, with synthesis_status: "failed" and the B-class findings from the reviewers' own envelopes.
Step 7: Present and apply
Output the synthesis directly. Do not soften or editorialize.
Then work through each objection independently, without waiting for the user: Apply when the fix is unambiguous and confidence is high, Dismiss when the artifact already addresses it or it misreads a stated constraint, Ask when it turns on user preference, a scope decision, or genuine uncertainty. Default ambiguous to Ask.
Dismissals must point at 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 before classifying; re-reading the artifact is not new evidence.
Then re-read the artifact in full, write the updated version with every Apply incorporated and everything else preserved, and present a compact summary: Apply bullets describing the direction of change (each opening with strengthened / narrowed / clarified / added / removed / inverted), one Dismiss: N objections line (omitted at zero), and all Ask items consolidated into a single message.
