audit - review code quality against the project standards

Context reuse: Reuse any required file already loaded in project instructions or the current session. Read it again only if absent, changed, or exact current bytes or line references are needed.

First action: Before project inspection, preflight, or any other tool call, publish running to blueprint/.state/run.json using the dashboard activity contract in AGENTS.md.

Where this sits in the workflow:

/implement or /autopilot  ->  [audit]  ->  fixes or /complete
(code exists)                 (review +    (repair quality issues
                               ledger)      or close the feature)

/check proves behavior against the spec. /doctor checks Blueprint setup and workflow health. This skill checks the code itself through either a broad review or one focused lens: quality, security, performance, or tests.

It reviews code without changing it: it never edits source files, installs dependencies, commits, merges, pushes, or starts product work. A normal audit's one write is the findings ledger at blueprint/context/findings.md (Step 4), the durable record of findings and their status. Independent mode may also write blueprint/context/review.md using the exact record contract in reference/independent-review.md.

The quality-gate config controls when another workflow invokes this skill automatically. An explicit /audit or $audit request always selects the audit regardless of whether the applicable gate is manual, conditional, or always. A selected independentReview gate invokes independent mode instead of letting the builder satisfy its own review. When both audit and independent review are selected, one passing independent review satisfies the audit gate. A missing config means built-in defaults. If it exists but is invalid, stop and point to /doctor before writing the findings ledger.

Input

Treat scope and lens as separate controls. Arguments may appear in either order, such as /audit security current or /audit src/auth tests.

Optional scope:

• no scope argument: use current when an active feature exists, otherwise use changed when local changes exist, otherwise use full
• current: audit the active current-feature.md, every committed feature-branch change from its merge base through HEAD, staged and unstaged changes, untracked source files, and nearby code affected by the feature
• changed: audit staged, unstaged, and untracked source files plus nearby code
• full: audit all project-owned source, tests, and configuration while excluding dependencies, generated files, build output, coverage output, caches, vendored code, and minified assets unless the user explicitly includes them
• path or directory: audit that area and the tests or callers needed to understand it

Optional lens:

• no lens: review all four lenses
• quality: maintainability, duplication, dead code, consistency, complexity, and standards drift
• security: authorization, input trust, injection, data exposure, secret handling, and unsafe configuration
• performance: query, network, rendering, memory, payload, concurrency, and unbounded-work risks
• tests: missing coverage for important logic, weak assertions, skipped or focused tests, poor isolation, brittle mocks, and likely flakiness

full is always the full-project scope, not a lens. /audit full therefore runs all lenses across the full project. When only a lens is supplied, select scope with the normal no-scope rules. A focused pass may name one or more lenses. If the request names multiple lenses, review their union and report them separately.

If the requested scope is unclear, pick the smallest useful scope and state it. If the lens is unclear, use all lenses and state that choice.

Optional review mode:

• independent: prepare or complete an independent review of current across all four lenses. It cannot be combined with changed, full, a path scope, or a focused lens because a completion receipt must cover the whole active work item.

Independent mode

/audit independent current is a two-context workflow. The builder prepares a request. With review.independentExecution: "manual", the selected fresh reviewer session runs the same command to complete it. With automatic, the current adapter may start a fresh isolated reviewer subagent after preparing the request. Blueprint verifies the exact target and later staleness. The adapter, model, and fresh-context identity remain declared metadata. An explicit invocation uses this execution setting even when the active workflow's independentReview gate is manual; that gate value disables only automatic selection by the workflow.

Read reference/independent-review.md before either phase.

Phase A - prepare the handoff

Use this phase when blueprint/context/review.md has no current pending request for HEAD and the current spec hash.

A current pending request without Requested execution is legacy and manual only. Never add execution fields to it or run a subagent against it. Stop with the fresh-session handoff; Phase B must omit Actual execution so the legacy request and receipt keep both execution fields absent.

1. Require an active spec with every build step checked and status verified, a non-default work branch, a reliable merge base, and a working tree matching the target except existing blueprint/context/review.md and blueprint/context/findings.md evidence. Independent mode accepts only a locally recorded remote default branch, local main, or local master as its enforceable base ref. Stop when none reliably covers the active work. The current full HEAD must be the approved application-code checkpoint. Include the verified spec when tracked; for a new ignored-spec request, prepare the exact local Spec snapshot under the reference contract. Never force-add it or change ignore visibility. Never create a commit inside Audit. If any tracked, staged, unstaged, or untracked path other than those two evidence paths differs from the target, stop and ask the user to approve a review checkpoint through /implement, even when normal checkpoint commits are disabled. Do not create a checkpoint solely for review/findings changes. This normal Phase A exception never allows snapshot Git differences or overwriting conflicting completion-recovery evidence.
2. Read installed adapters from blueprint/.state/manifest.json when valid. For older installs, detect .agents/skills as codex and .claude/skills as claude. These files prove project support, not that the external runtime is installed or authenticated.
3. Resolve the review executor from review.independentExecution:
   • For manual, ask which detected adapter and available model should review. Recommend an equal-or-stronger coding model, a different model family when practical, and high reasoning for sensitive work. Offer a fresh session in the current adapter as the fallback. Do not invent available models or offer an adapter that is not installed in the project.
   • For automatic, use only a live child-agent capability in the current adapter that can start with no builder transcript, disclose the exact reviewer adapter and model, and wait for completion. Spawn a generic fresh isolated child through the current runtime. Do not discover, select, or depend on a globally installed role, skill, prompt, or another workflow such as TraversyFlow. If the runtime cannot start that generic child from project-local instructions, or capability, isolation, identity, model, completion, or access to the same ignored spec/snapshot inputs cannot be confirmed, use the manual path in the original checkout.
4. Record the full target SHA, full merge-base SHA, the exact local base ref used to calculate it, exact spec SHA-256, current adapter and model, requested reviewer adapter and model, requested execution from review.independentExecution, workflow, and whether the configured Check gate is required. For a new ignored-spec request, create or reuse the exact snapshot first and record Spec snapshot as defined in the reference. Never add that field to an existing pending or completed record. Write the pending template exactly. Copy the full model identifier exposed by the active runtime or session metadata (for example, gpt-5.6-sol), never a generic family label such as GPT-5. If the runtime does not expose an exact identifier, record unknown (runtime did not expose exact model) instead of guessing. When the reviewer runtime cannot select a specific model before opening the session, record the exact runtime-default sentinel from the reference contract.
5. Execute the configured path:
   • For manual, set dashboard activity to ready and give the exact handoff command for the selected adapter. Claude Code and Google Antigravity use /audit independent current; Codex uses $audit independent current; Copilot and OpenCode receive the equivalent plain-language instruction. Tell the user to open a fresh session in the original checkout with only the handoff, not the builder chat. Include target/base SHAs and, when present, the exact snapshot path and spec hash.
   • For automatic, freeze all parent product, test, spec, and config changes. Start one generic isolated child without the builder transcript. Instruct it to read the project-local Audit skill and audit/reference/independent-review.md from the current adapter tree, then execute Phase B using the same local spec/snapshot inputs against the prepared request. All review instructions come from that installed Blueprint project. The reviewer may write only blueprint/context/findings.md and blueprint/context/review.md; it must not repair code, change the spec, commit, or perform external actions. Wait for completion, then reread and validate the normal receipt before continuing. Record fresh subagent as its reviewer context.

If automatic execution fails or any required property becomes uncertain, keep the pending request intact, set activity to ready, and stop with the existing manual fresh-session handoff. Never let the builder review its own work or skip a selected independent-review gate.

The builder never performs Phase B itself. It may continue only after a manual reviewer session or automatic isolated reviewer produced a valid current receipt.

Phase B - perform the review

Use this phase when a current pending request exists.

1. Confirm the current adapter matches Requested reviewer, the current model matches Requested model unless the runtime-default sentinel was selected, and a sentinel request now records the exact model exposed by the session, HEAD matches Target commit, the recorded base ref still produces the recorded merge base, the exact spec hash matches, and no path differs from the target except blueprint/context/review.md and blueprint/context/findings.md. When Spec snapshot is present, verify both raw spec/snapshot hashes and every path, visibility, and Git condition in the reference. Stop on any mismatch or stale state.
2. Proceed only from the fresh reviewer handoff. Record fresh session for a manual reviewer or fresh subagent for an automatic isolated reviewer. This is a declaration, never cryptographic proof. If the reviewer has the builder conversation or is the builder continuing in place, stop and request a fresh context. For a legacy request with no Requested execution, require a fresh reviewer session, record fresh session, and omit Actual execution.
3. Run Steps 1 through 3 across current with quality, security, performance, and tests together. Review the code fresh against the recorded Base commit and Target commit; exclude the request and findings files from the code scope. Existing findings are context, never the review checklist.
4. Run /check from the reviewer session when the request says Check is required. Follow Check's server and evidence boundaries. A required check that cannot run prevents a passing receipt.
5. Update the findings ledger through Step 4, then replace the pending request with a completed receipt. Use passed only when the whole target was reviewed, required checks passed, and no P0 or P1 finding is open or fixed. Copy the reviewer's full runtime model identifier using the same rule as Phase A. Record Check result and keep all four receipt sections non-empty, using an explicit None entry when appropriate. List every unavailable verification command under Remaining risk, even when Check was not required and the receipt may still pass. Otherwise use changes-requested and name the exact blockers. Record actual automatic with fresh subagent, or actual manual with fresh session, including an explicit manual fallback from an automatic request.
6. Report the receipt target, reviewer adapter and model, commands, evidence, findings, remaining risk, and whether the receipt passed. Never repair code from the reviewer session.

After changes are requested, the builder repairs through /implement, obtains approval for a new checkpoint, and prepares a new request. The next reviewer pass reviews the complete new delta, not only the old findings. A local-spec-only revision may reuse the same approved product HEAD after normal spec and verification gates, with a new snapshot/request and full fresh review; it never requires an empty commit.

Step 1 - gather context

Resolve this required context:

• AGENTS.md
• blueprint/config.json
• blueprint/context/project-overview.md
• blueprint/context/coding-standards.md
• blueprint/context/current-feature.md
• blueprint/context/findings.md, for existing IDs and statuses
• blueprint/context/review.md, for independent request and receipt state
• blueprint/context/ai-interaction.md
• blueprint/build-plan.md, when feature order matters
• git branch and working tree status
• relevant source files, tests, and configs for the chosen scope

For current, changed, and path scopes, begin with the diff or named area and follow only the callers, dependencies, tests, and contracts needed to verify a reachable finding. Do not survey unrelated directories. For full, preserve the declared exclusions and inspect by bounded area rather than dumping the project into one response.

For current, resolve the comparison base without network access:

1. Use a base branch declared by the active spec or project instructions.
2. Otherwise use the locally recorded remote default branch when available.
3. Otherwise use an existing local main, then master.
4. Find the merge base and inspect the committed delta through HEAD, then add staged, unstaged, and untracked work.
5. If no reliable base exists, say so and use the active spec plus local changes. Never claim that committed feature work was fully covered in that case.

Do not fetch or pull to discover the base. For full, state the excluded paths before reviewing so generated or third-party code does not consume the audit.

Prefer rg and targeted file reads. Do not dump large files into the response.

Step 2 - run available signals

Use existing commands only. Do not install tools.

Run or inspect only the signals relevant to the selected lens and scope:

• lint and typecheck commands when declared and relevant
• test command for the tests lens or when it directly validates a suspected risk
• build command when the selected lens needs compilation or bundle evidence
• existing security command for the security lens, when declared and locally runnable
• existing performance command for the performance lens, when declared and locally runnable
• targeted lightweight searches for the chosen lens, such as unused exports and copied logic for quality, unsafe trust boundaries for security, repeated or unbounded work for performance, and skipped or weak tests for tests

Do not run broad checks unrelated to a focused lens. If a useful command is missing, report that as a gap. Do not invent a pass or claim that a focused review covered the other lenses.

Step 3 - review the code

For all lenses, ground findings in reachable code and project-specific expectations. Apply only the selected lens or lenses:

• Quality: duplicated logic, dead or unused code, unreachable paths, oversized modules, abstractions that do not pay for themselves, risky missing abstractions, speculative dependencies, services, configuration surfaces, compatibility layers or security machinery, inconsistent patterns, and drift from the standards or spec. Untuned stack-specific template defaults are not established requirements.
• Security: missing authentication or authorization, client-controlled ownership, injection, unsafe parsing or deserialization, sensitive-data exposure, secret handling, insecure defaults, and trust-boundary mistakes. Inspect existing dependency or scanner output when available, but never imply that local manifest inspection is a current vulnerability scan.
• Performance: N+1 queries, repeated network or database work, unnecessary rendering, blocking work on hot paths, unbounded loops or collections, memory growth, oversized payloads, missing pagination, and unsafe concurrency. Mark hypotheses as unverified when runtime or profiling evidence is missing.
• Tests: important logic without coverage when a test command exists, weak assertions, tests that only mirror implementation, excessive mocking, shared state, time or order dependence, skipped or focused tests, placeholder tests, swallowed failures, and missing browser or integration evidence where behavior crosses a real boundary. Never invent a coverage percentage.

Do not nitpick harmless style differences unless they signal drift from the local patterns. Prefer a short list of real findings over a broad list of guesses.

For a proportionality finding, state in Suggested fix what can be deleted, which existing, standard-library, native-platform, or installed mechanism replaces it, and which current requirement would be lost. Use None when no current requirement would be lost. If the suggested fix removes or changes shipped behavior, require an explicit user decision and never describe it as an automatic repair.

Do not broaden a focused pass because another category might be interesting. Do not report or call out non-critical concerns from omitted lenses, even as suggestions for a later audit. If an obvious P0 is directly encountered outside the selected lens, report and record it as an out-of-lens critical risk, but do not continue searching that other lens.

If a possible secret is found, never quote its value, paste the matching source line, or include raw command output containing it. Report only the redacted secret category, file, line, risk, and remediation. Redact sensitive values from all audit evidence before responding.

Step 4 - update the findings ledger

blueprint/context/findings.md is the durable record of findings. Chat reports do not survive a context clear; the ledger does. It is the only file this skill writes. If it is missing (an older install), create it with a # Findings heading first.

The ledger never scopes the review. Review the code fresh in Step 3, then record what the review found. Working from the open findings as a checklist and verifying only those is the exact failure this file exists to prevent: a repair can introduce a new defect that no existing entry points at.

One block per finding. The header line is the machine-readable contract and must keep this exact shape; the prose below it is for humans and may vary:

### F-03 [P0] open - Retained auth volumes carry the run label

**File:** ops/agent-proof/compose.yaml:86
**Found:** 2026-07-21 by /audit (scope: current; lens: security)
**Why it matters:** ...
**Suggested fix:** ...
**Resolution:**

IDs are sequential within the ledger (F-01, F-02, ...), never reused and never renumbered while their entries live here, even after a finding closes. Bare IDs are scoped to the live ledger: /complete archives resolved entries under a work-item prefix (feature 12's first F-03 becomes 12/F-03, and its second build's becomes 12-build-2/F-03). The build attempt comes from the verified spec/history proof, not arbitrary filename text; fix and rollback prefixes stay their archive filenames. That prefixed form is the permanent reference. A later ledger that has emptied and reset starts at F-01 again without colliding. Severity reuses the P0-P3 scheme from Step 5; only P0 and P1 block /complete. Status is one of:

Status	Meaning	Blocks P0/P1 at /complete
unverified	Suspected, no confirming evidence yet	No
open	Confirmed, not yet repaired	Yes
fixed	Repaired, not yet re-reviewed	Yes
closed	Repaired and re-reviewed against the new code	No
accepted	Not fixing, by the user's explicit decision; reason recorded in Resolution	No
invalid	Re-examination proved the finding wrong; evidence recorded in Resolution	No

After the review:

• Append each new confirmed finding as open with the next sequential ID, one past the highest ID present in the ledger (entries carried forward from earlier work count; a fresh ledger starts at F-01).
• Record an unverified risk worth tracking as unverified. It is a lead, not a defect, and never gates a merge.
• Update the entries this pass re-examined: correct the status or severity and note the evidence in Resolution.
• Move a fixed finding to closed only when all three hold: this pass's reviewed set included the finding's file, re-examining the repaired code confirmed the original defect is gone and the repair introduced no new one, and the report names the finding as closed. An unrelated new finding in the same file gets its own entry and does not keep the repaired one open. Never close a finding implicitly.
• Set accepted only on the user's explicit decision in the current session, and record their reason. Never accept a finding on their behalf.
• Set invalid only when re-examination shows the finding was wrong, and record that evidence in Resolution. It is a review verdict (or the user's explicit call), never a shortcut past the gate for blocked work.

fixed blocking /complete is deliberate: a repair is not done when the code changes, it is done when a review has looked at the result. /implement marks repairs fixed; only a review pass moves them to closed.

Step 5 - report findings

Lead with findings, ordered by severity, using the IDs the ledger assigned:

F-04 [P1] Title
File: path:line
Why it matters: ...
Suggested fix: ...

Severity:

• P0 - data loss, security break, or code that cannot ship
• P1 - likely bug, broken contract, missing guard, or high-risk duplication
• P2 - maintainability issue worth fixing before the feature closes
• P3 - small cleanup, consistency issue, or follow-up candidate

Use P0 or P1 only when a concrete code path, violated contract or security boundary, failing command or test, or reproducible behavior confirms the risk. If the evidence is incomplete, list it under Unverified risks with the missing validation instead of presenting it as a confirmed high-severity finding. An otherwise pure proportionality finding is P2 or P3. Raise it to P0 or P1 only when the unnecessary machinery causes a concrete reachable defect or violates an established security or data-integrity boundary.

If there are no findings, say that clearly for the selected lens and name any remaining risk or missing signal, such as "no test command declared" or "browser flow not audited."

Then include:

• ledger changes: findings added, updated, or closed this pass, by ID
• commands run and results
• selected scope
• selected lens or lenses
• base branch, merge base, and commit range for current, when available
• files or directories reviewed
• generated, third-party, or otherwise excluded paths
• applicable standards checked
• browser or runtime evidence inspected, when relevant
• skipped, focused, or placeholder tests found, when the tests lens was selected
• checks that were unavailable or could not run
• suggested repair order
• independent receipt status and target, when independent mode ran

For full, say whether coverage was complete or partial. Never label a partial review as a full-project audit.

Rules

• A normal audit writes only the findings ledger. Independent Phase A may also create the exact local spec snapshot and pending request. Phase B may write only blueprint/context/findings.md and blueprint/context/review.md. Never edit, format, install, commit, merge, push, or delete anything else.
• Never let a builder complete its own independent request in the same session.
• Never silently substitute another reviewer adapter or model.
• A stale receipt is no receipt. Re-review the complete new checkpoint.
• A focused lens is not a broad audit. State what was not reviewed and never imply that omitted lenses passed.
• The ledger reports status; it never defines what the review looks at. Do not turn open findings into the review checklist.
• Never fetch, pull, or run network-backed audit tools without explicit approval.
• Never reproduce secrets or sensitive values in findings or command output.
• Findings first. Keep summaries short.
• Ground every finding in a file path and line number when possible.
• Avoid speculative rewrites. Recommend the smallest fix that removes the risk.
• Respect existing project patterns over generic advice.
• Do not require perfection. The goal is code that is understandable, consistent, testable where it matters, and safe to keep building on.

Formatting

Format the output to match the project's conventions in blueprint/context/ai-interaction.md: concise, scannable markdown, with lists for enumerations and tables for matrices rather than dense paragraphs.