Cynical Code Review

Cynical Code Review

Use this workflow only after explicit selection as $cynical-code-review or a binding task instruction requiring it.

The job is to treat the implementation story as unproven, read current code until the real authority paths are visible, find the places the work lied by name, status, proof, or structure, save a findings-first review artifact, and report the verdict plus path.

This skill does not implement, repair, commit, push, open PRs, route repair waves, run external review subprocesses, or decide the user's broader workflow.

Use When

• The user asks for a cynical, skeptical, adversarial, hostile, suspicious, or implementation-integrity code review.
• The user says a plan, phase, branch, diff, or completion claim was implemented but may have missed the point, stopped short, or only looked done.
• The user wants to find where code was implemented in name but not fact: duplicate truth, split-brain owners, side doors, old paths still live, partial unification, false simplification, or current code rationalized as design.
• The user says docs, status, tests, worklogs, labels, wrappers, or reviewer receipts may be lying about what the code really does.
• The user wants the audit run repeatedly after repairs to surface the next layer of code-reality issues.

Do Not Use When

• The user wants a normal high-signal code review. Use the host agent's normal review response.
• The user wants exhaustive line-by-line, file-by-file, or coverage-ledger code review where coverage itself is the deliverable. Use exhaustive-code-review.
• The user wants a plan audited before work starts. Use plan-audit.
• The user wants code reviewed mainly against a plan artifact's architecture and quality bar, without investigating whether the completion story is truthful. Use plan-audit implementation-audit.
• The user wants only a harsh maintainability, spaghetti, file-size, or code-judo pass. Use thermo-nuclear-code-quality-review.
• The user wants an external Codex, Claude, Cursor, Grok, or Kimi second opinion. Use the explicitly requested consult or delegation skill.
• The user wants fixes, implementation, PR shipping, or workflow orchestration.

Non-Negotiables

• Review only. Do not edit reviewed files.
• Save the review artifact under /tmp/cynical-code-review/<slug>-<timestamp>/ unless the user explicitly asks for a repo doc path.
• This is prompt-only doctrine. Do not build a rule engine, runner, controller, scorer, harness, script, or formal parameter interface.
• Apply ../_shared/agent-orchestration-policy.md whenever the review uses child agents.
• For broad targets, fan out only across genuinely independent review lenses or path families. Start each independent slice as a new clean same-host native child when the active host supports it, keep the slices non-overlapping, and bound fanout by host slots, shared-file or shared-state collision risk, and the parent's ability to integrate every result.
• Use the strongest read-only capability the host exposes, also tell every review child not to edit or write, and have the parent compare repository status and diffs with the pre-dispatch state before accepting child evidence.
• Children may use native sub-agents within authorized scope and host limits; they may not start external agents.
• The parent owns child accounting, deduplication, integration, scope disposition, the saved artifact, and the final verdict.
• Do not manually spawn codex, claude, agent, grok, kimi, or any other coding-harness executable.
• Do not invoke external agent, delegation, consult, or review skills as the review mechanism.
• Start from distrust: the implementation story, completion claim, names, wrappers, comments, status blocks, docs, tests, logs, and reviewer launches are claims, not proof.
• Current code behavior and structure are the authority for what exists. Human requests and explicit human approvals are the authority for scope. Plans, docs, worklogs, tests, and status surfaces help aim review; an agent-authored revision cannot retroactively authorize code.
• Apply ../_shared/scope-and-convergence.md. For plan-, branch-, conductor-, PR-, or history-backed work with a recoverable baseline, reconstruct the initial human scope, pre-approval convergence closure, later human approvals, plan/review waves, and final code. Unauthorized post-approval growth or scope cycling is a REQUIRED REPAIR and forces not-approved, even when it works and tests pass. The normal repair target is subtraction.
• If provenance should be available but cannot be recovered, return coverage-incomplete. For a standalone code target with no plan, completion claim, or human-scope history, mark the provenance lane not applicable rather than inventing a baseline.
• Findings must be concrete current-code or requested-scope risks. Drop style preferences, generic maintainability advice, missing-test nits, and doc hygiene unless they expose a false implementation story.
• A clean review is allowed, but only after the likely ways the work could be lying have been checked and recorded honestly.

First Move

1. Resolve the requested review target from natural language: current worktree, branch diff, commit range, explicit paths, plan scope, or completion claim.
2. Read local instructions and nearby review-relevant conventions.
3. Create the run directory under /tmp/cynical-code-review/.
4. Write the implementation story in plain English: what the work claims is now true, which source or plan says so if present, and what current code would have to do for that claim to be real.
5. When scope history exists, write its human baseline, initial closure and approval anchor, explicit later approvals, and revision/review waves before accepting the latest plan as context.
6. Read references/review-catalog.md.
7. Read references/output-contract.md.
8. Read ../_shared/agent-orchestration-policy.md before creating or resuming any child.
9. Read references/agent-slices.md before launching native review slices.
10. Read references/examples.md when the review involves completion claims, plan-backed implementation, unification, simplification, proof/status claims, or confusing peer review lanes.

Workflow

1. Save the review target summary as target.md.
2. Build suspicion-map.md: implementation story, controlling source truth, proxy evidence, expected code behavior, old authority paths, duplicate truth risks, side doors, adjacent same-contract surfaces, user job, and likely scope contamination. For scope-backed work, start from the initial human scope, not the latest plan revision.
3. Map changed files, changed hunks, touched symbols, touched abstractions, visible behavior obligations, relevant old paths, and likely adjacent surfaces.
4. When broader coverage warrants children, dispatch a proportional set of new clean native read-only slices with distinct lenses or path families: claim mapping, old-path hunting, runtime flow tracing, split-brain hunting, user-job review, overbuild/scope review, and proof-surface review. Integrate and account for every slice in the parent.
5. Trace actual control flow, data flow, callers, readers, writers, lifecycle, persistence, generated artifacts, prompts, configs, and public entrypoints until the current authority path is visible.
6. Treat every attractive label as suspect: unified, canonical, owner, stable, complete, migrated, deleted, simplified, verified, ready, single source, and strict.
7. Review competing ways to accomplish the same goal: old APIs, sibling callers, direct writers, fallback readers, command aliases, duplicate helpers, compatibility flags, docs, examples, tests, fixtures, generated artifacts, prompt surfaces, and side doors.
8. Check whether the user's actual job works from the starting state that matters. A real surface that only supports a narrower internal job is not complete.
9. Check whether the implementation created machinery that hides the old problem: harnesses, policy layers, flag matrices, wrappers, fake abstractions, process receipts, or proof systems that add concepts without making the intended code behavior true.
10. Review proof surfaces only as code-reality claims. Changed tests, fixtures, docs, worklogs, examples, comments, prompts, logs, and status blocks matter when they prove the wrong thing, mask code gaps, keep old paths alive, or would mislead the next agent.
11. Run the Unauthorized Scope Ratchet Or Cycling pattern whenever scope provenance is recoverable. A reviewer discovery can be real and still lack authority; do not recommend another generalized system as the repair.
12. Use references/review-catalog.md to classify concrete risks. Use references/examples.md as illustrations of reasoning, not a lookup table.
13. Save coverage.md, findings.md, and verdict.md.
14. Return a short findings-first answer with the verdict and run directory.

Output Expectations

The final chat reply includes:

• verdict: approve, not-approved, or coverage-incomplete
• required repairs, if any
• observations, if material
• run directory path
• one next action from the review, not a workflow prescription

The full saved artifact follows references/output-contract.md.

Reference Map

• references/review-catalog.md - cynical implementation-integrity review patterns, what evidence to read, required-repair conditions, safe-difference guards, and example findings
• references/agent-slices.md - native parallel-agent slice guidance and accounting rules
• ../_shared/agent-orchestration-policy.md - transport, starting context, continuation, isolation, topology, and parent-integration policy
• references/output-contract.md - required saved files, verdicts, finding shape, suspicion map, and final chat summary shape
• references/examples.md - concrete failure-pattern examples and stronger review moves derived from recurring agent-history failures