Exhaustive Code Review

Exhaustive Code Review

Use this workflow only after explicit selection as $exhaustive-code-review or a binding task instruction requiring it. The job is to review the requested code scope, save the review artifact to disk, and report the verdict plus path.

This skill does not dictate the user's workflow. It does not implement, repair, commit, push, open PRs, route repair waves, run an external review subprocess, or decide what workflow the user should use next.

Use When

• The user asks for exhaustive, meticulous, line-by-line, file-by-file, abstraction-by-abstraction, feature-by-feature, or coverage-ledger code review.
• The user wants a full branch, current diff, commit range, explicit path set, plan scope, or completion claim reviewed with every relevant touched surface accounted for.
• The user specifically cares about split-brain abstractions, bypassed centralized owners, partial migrations, side doors, drift surfaces, stale docs/prompts/generated artifacts, or proof that tests the wrong thing.

Do Not Use When

• The user wants a normal high-signal general review. Use the host agent's normal review response.
• The user wants implemented code reviewed mainly against a plan artifact. Use plan-audit implementation-audit.
• The user wants only a harsh maintainability pass. Use thermo-nuclear-code-quality-review.
• The user wants an external Codex/Claude/Cursor second opinion. Use the appropriate 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/exhaustive-code-review/<slug>-<timestamp>/.
• Apply ../_shared/agent-orchestration-policy.md whenever the review uses child agents.
• Build a coverage-led slice plan only when distinct lenses or path families improve the review. Start every independent slice as a new clean same-host native child when supported, keep scopes non-overlapping, and bound fanout by host slots, shared-file or shared-state collision risk, and parent integration capacity.
• 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, evidence spot-checking, deduplication, integration, finding scope disposition, the artifact, and the final verdict.
• Do not manually spawn codex, claude, agent, or other coding-harness executables.
• Do not invoke external agent/delegation/review skills as the review mechanism.
• Do not build a rule engine, runner, controller, scorer, harness, or script.
• Read repo truth directly: changed files, local instructions, relevant callers, owners, tests, docs, schemas, generated artifacts, prompts, examples, and configs.
• Findings must be concrete. Drop style preferences, generic advice, and "maybe centralize this" comments unless they name real changed-code risk.
• Competing-path, side-door, stale-truth, and competing-owner detection is default review behavior, not a special mode. If a live duplicate path affects the requested scope, treat it as a required repair unless the review can name the genuinely different contract or controlling out-of-scope anchor. For a fixed-scope plan or history-backed change, apply ../_shared/scope-and-convergence.md: a reviewer-discovered adjacent path cannot enter repair scope. Require subtraction or redesign inside the approved boundary, or stop for an explicit human scope decision.
• A clean review is allowed. Do not invent findings to justify the run.

First Move

1. Resolve the requested review target from normal 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/exhaustive-code-review/.
4. Read references/review-catalog.md.
5. Read references/output-contract.md.
6. Read ../_shared/agent-orchestration-policy.md before creating or resuming any child.
7. For a fixed-scope plan or history-backed change, read ../_shared/scope-and-convergence.md.

Workflow

1. Build the review target summary and save it as target.md.
2. Map the changed files, changed hunks, touched symbols, touched abstractions, visible features or behavior obligations, and likely adjacent surfaces.
3. Build a proportional coverage plan. When parallel slices materially improve coverage, launch new clean native read-only children over non-overlapping lenses or path families, account for every final state, and synthesize in the parent. In Codex use fork_turns: "none"; in Claude use a clean named or custom subagent rather than an ambiguous conversation fork or skill context: fork shorthand. Use inherited context only for a named chat-only dependency, not the parent's completion narrative.
4. Review every touched file and changed hunk. Read surrounding code when the hunk depends on a function, class, module, caller, lifecycle, or contract.
5. Review touched abstractions: canonical owner, old and new paths, callers, readers, writers, side doors, invariants, and drift surfaces.
6. Review competing ways to accomplish the same goal: old APIs, sibling callers, direct writers, alternate readers, duplicate helpers, command aliases, generated artifacts, docs, prompts, examples, tests, and side doors. Classify each in-scope competing path as a required repair, observation, genuinely different contract, or named out-of-scope follow-up. When scope is signed off, classify every material finding against its human-scope or pre-approval convergence anchor before naming a repair; review discovery is not scope authority.
7. Review touched behavior: entrypoints, success paths, failure paths, state, persistence, user-visible or externally observable effects, and proof.
8. Review live truth surfaces: tests, fixtures, docs, examples, comments, schemas, generated artifacts, prompts, agent instructions, config, package metadata, stable IDs, telemetry names, and install commands when they describe or consume changed behavior.
9. Use the review catalog to identify concrete risks.
10. Save coverage.md, findings.md, and verdict.md.
11. 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 - concrete review patterns, evidence to read, required-repair conditions, safe-difference guards, and example findings
• references/output-contract.md - required saved files, verdicts, finding shape, and final chat summary shape
• ../_shared/agent-orchestration-policy.md - transport, starting context, continuation, isolation, topology, and parent-integration policy