Codebase audit

Assess the repository's current state rather than a branch diff. Produce a truthful coverage map, actionable findings, and coherent remediation batches.

Modes and boundaries

The audit is non-remediating by default. It may run safe local tests, builds, servers, browser flows, disposable repros, isolated test data, temporary artifacts, and revert-clean falsification probes. A falsification probe temporarily introduces one targeted wrong behavior to prove the relevant test or gate fails for the right reason. Record the initial tree state, prefer a disposable worktree or copy, restore the mutation immediately, compare the final tree byte-for-byte with the recorded state, and report the mutation and observed failure. Skip the probe and mark the evidence static-only when exact restoration is not safe. The audit must not leave any edit to repository content or create persistent or externally visible side effects. Clean up disposable artifacts and report anything retained.

subagents is an additive execution input. Without it, do not delegate any part of the audit.

A request to save JSON authorizes only the named findings artifact; it does not authorize remediation. Read references/findings-json.md only when JSON output is requested.

There is no fix-all mode. After reporting, offer remediation batches. A fix follow-up begins only when the user selects a coherent batch or finding IDs. It authorizes focused local edits and validation on a focused branch, not commits, pushes, publication, issue creation, deployments, or shared-state mutation.

Map before judging

1. Read applicable project instructions, vision and product documentation, current source, tests, fixtures, configuration, lockfiles, generated interfaces, schemas, public APIs, examples, templates, completion contracts, packaging, CI, scripts, release, deployment, and operational definitions.
2. Map user-visible capabilities, entry points, modules, trust boundaries, data and state flows, background work, external dependencies, tests, tooling, and operational paths.
3. Choose perspectives from the actual codebase:
   • always cover behavior, correctness, architecture and consistency, simplification, self-documentation, test value, and operations;
   • add UI/UX and accessibility when interfaces exist;
   • add security and adversarial-input analysis at trust boundaries;
   • add persistence, migration, transaction, concurrency, and idempotency analysis for stateful paths;
   • add API, CLI, library, packaging, compatibility, deployment, rollback, observability, performance, and discoverability analysis only where those surfaces exist; activate discoverability for a public web surface.
4. For a shallow subsystem, trace its complete path and direct interactions. For a layered codebase, partition work by capability and perspective so later areas do not receive progressively thinner analysis.
5. Build a bounded churn map from repository history. Rank frequently changed files, then inspect function, method, class, or module history where symbols are reliable. Use the repository's declared churn window or 90 days when none exists. Prefer its code-health tool; otherwise use Git file history and git log -L for stable symbols. Follow renames; state the window, touch count, and line churn, and label file-level fallback.

Sub-agent lanes

When the user supplies subagents, the coordinating agent still owns the audit scope, coverage map, capability map, active and skipped perspectives, validation, final findings, remediation batches, and report.

1. Publish a bounded lane map before delegation. Form lanes from capability and perspective intersections so no worker receives an unbounded whole-repository perspective. Give each lane one worker; tightly coupled or individually small perspectives may share a lane. Queue excess lanes when platform capacity is temporarily full.
2. Give each worker its lane ID, assigned capability and perspectives, bounded scope, relevant project instructions, and known evidence. A worker may inspect and run the safe probes allowed by this skill, but it must not edit, create persistent or external side effects, delegate further, assign final finding IDs or severities, propose final remediation batches, or issue an overall conclusion.
3. Require each worker to return its lane ID, assigned capability and perspectives, bounded scope, inspected supporting context, exact probes and observed results, candidate findings with evidence, impact, and smallest remedy, verified claims, limitations, and complete or incomplete status.
4. Validate every candidate against the current checkout, apply the de-duplication model below, reconcile conflicts across lanes, then assign final IDs, severities, categories, remediation batches, and conclusions. Do not repeat a completed lane wholesale.
5. If sub-agents are unsupported, unavailable after any applicable bounded retry, or leave a lane incomplete, complete that lane directly. Report the affected lane and reason as a single-agent fallback. Temporary capacity exhaustion queues work rather than triggering immediate fallback.

Establish current evidence

• Treat documentation, tests, issue text, prior audits, model memory, and generic best practice as leads, not proof.
• Exercise representative behavior through the real UI, API, CLI, library entry point, job, migration, package, or deployment path. Cover success plus the most consequential failure or boundary case.
• Record setup, action or command, input, expected result, and observed result. Mark anything not executed static only.
• Verify test value: relevant wrong behavior should fail; assertions should cover outcomes and meaningful failure paths without excessive mocks, snapshots, or implementation coupling.
• Verify third-party claims against the exact locked version and current official documentation or source. Check whether a current supported version removes custom code, workarounds, adapters, or configuration.
• Compare documentation and declared operational flows with current implementation. A passing unit suite does not validate a broken build, migration, package, deployment, or rollback path.

Audit checks

De-duplication

Apply four separate checks across the repository and its delivery surface:

• Implementation: find repeated code, logic, tests, fixtures, generated forms, schemas, APIs, configuration, workflows, documentation, examples, templates, release or deployment paths, operational scripts, dependencies, custom tooling, and competing representations of one concept.
• Work: find repeated investigations, questions, decisions, findings, tickets, remediation, reruns, and recurring repair patterns across current issue and pull-request history and decision records. Reuse and revalidate existing work.
• Evidence: coalesce the same event reported by multiple checks, logs, ledgers, or tools so it is counted once while retaining every source.
• Output: combine candidates with the same cause, impact, and remedy into one finding, preserve provenance, and explicitly reconcile contradictory evidence.

The default scope may inspect direct consumers and sibling components to test a public contract or competing representation. Report a cross-repository finding only when the user included that scope or current public-contract evidence makes the external impact part of the repository finding.

Technical checks

• Trace boundary inputs, authorization, state changes, failures, partial success, retries, concurrency, idempotency, deletions, and side effects where relevant.
• Check dependency direction, module cohesion, ownership, public interfaces, and whether multiple patterns or representations compete for the same concept.
• Search the live repository before proposing anything new. Prefer deletion, reuse, direct control flow, consolidation, the standard platform, and existing dependencies over new layers.
• Report dead code, duplication, speculative abstractions, needless wrappers, one-use indirection, stale compatibility paths, and dependencies that no longer earn their cost.
• Require names, types, boundaries, and interfaces to communicate intent. Match the surrounding comment density; comments should preserve rationale and constraints, not narrate syntax.
• Treat repeated changes to the same symbol or file as an architectural-risk signal, not a defect by itself. Raise an ARCHITECTURE_RISK finding when the measured churn coincides with mixed responsibilities, recurring fixes or reverts, competing representations, broad blast radius, unstable interfaces, or weak regression coverage. Cite the window, touch count, granularity, and co-signal.
• Suppress generic advice when repository evidence documents a deliberate alternative. Every best-practice finding must cite current project and authoritative-source evidence plus the concrete simplification or risk.

Discoverability

For a public web surface, verify crawl and index controls, canonical and descriptive metadata, internal discovery paths, structured data that matches visible content, semantic content structure, rendering, and material web performance. Assess conventional search and AI-assisted discovery together, while keeping crawler access, search inclusion, and model-training controls distinct. Use current official search-engine and publisher guidance; do not invent special AEO markup, keywords, or guarantees.

Report

Search the complete mapped scope for evidence-backed candidates before filtering the report; do not stop after the first or highest-severity issue.

Lead with the highest-value current conclusion. Include:

• a coverage map of inspected, executed, sampled, static-only, and unreached capabilities;
• active and skipped perspectives with reasons;
• when subagents was supplied, the capability-and-perspective lane map, completed and incomplete lanes, and every coordinator-completed fallback with its reason;
• the churn window, symbol/file coverage, and architectural-risk hotspots;
• exact probes and project gates with observed results;
• de-duplication coverage across implementation, work, evidence, and output, including coalesced sources and reconciled conflicts;
• actionable findings as [CA-N][BLOCKING|IMPORTANT|IMPROVEMENT][BEHAVIOR|QUALITY|ARCHITECTURE_RISK| OPERATIONS|DISCOVERABILITY] file:line: evidence, impact, smallest remedy;
• grouped remediation batches ordered by risk reduction, dependency, and reviewability;
• limitations and retained probe artifacts.

BLOCKING is a current exploitable, corrupting, or deployment-blocking defect. IMPORTANT has material correctness, security, operability, test-value, maintainability, simplification, or comprehension cost. IMPROVEMENT is a verified worthwhile simplification or current-practice alignment. Omit praise, inventory narration, style nits, and findings without concrete impact.

Fix follow-up

After the user selects a batch or IDs, create or reuse a focused branch and implement the smallest complete remedies. Do not absorb unrelated findings. Promote useful probes into regression tests, remove disposable artifacts, run affected behavioral probes and project gates, and report fixed and unresolved IDs with observed evidence. The coordinator makes every edit. Do not redispatch completed lanes after fixes; re-engage a worker only to resolve incomplete or contradictory evidence. Use the repository's separate git and PR workflows only when the user requests publication.