Audit-Fix Loop

Converge on verified fixes for defects within the requested scope.

Contract

Identify the target diff, feature, paths, or whole-project scope from the request and repository instructions. Preserve existing authorization and user-owned changes. Audit/review alone produces findings; apply fixes only when requested. Continue already authorized fixes without a new permission question for each round.

Inspect source, callers, tests, and build configuration before choosing audit slices or verification commands. Do not add cleanup, commits, PRs, remote messages, memory updates, or documentation projects to the user's request. Update documentation directly invalidated by a fix as part of that fix.

Audit and triage

Audit the full requested scope first. Report concrete defects with location, trigger or reproduction, expected versus actual behavior, impact, and evidence. Trace cross-file contracts far enough to establish the failure. Distinguish reproducible bugs from stylistic preferences, uncertain hypotheses, and unrelated existing debt; do not turn coverage percentages into bug findings.

Deduplicate findings that share a root cause. Prioritize by impact and confidence grounded in evidence, without a numeric confidence cutoff. An in-scope supported defect remains work even when its severity is low; a nit or speculative redesign does not become required merely because a reviewer mentioned it.

If parallel work is useful and permitted, delegate independent audit slices or fix clusters with explicit boundaries. Give writers disjoint ownership and keep dependent fixes sequential. Choose available capabilities for the actual difficulty; no prescribed agent roster, model, or compressed return limit. Require enough evidence to integrate and verify each result.

Fix, verify, recheck

Implement the smallest change that resolves the root cause and preserves contracts. Test the demonstrated failure mode using the repository's existing tools and fixtures. Add a regression test when it meaningfully protects that behavior; choose unit, integration, or UI coverage from where the defect lives.

Examples of useful verification:

Do not require both unit and E2E tests for every fix. Reversible wording or format changes usually need inspection or the relevant linter. Required project checks and project coverage contracts still apply.

After a fix, run the focused check and required project verification. Investigate failures; distinguish failures caused by this diff from existing or environmental failures. Do not delete a valid failing test to claim green. Reuse passing evidence until new changes or unresolved concerns justify repeating it.

Re-audit the changed files, their callers, affected contracts, and new tests. Expand only when evidence shows wider impact. The initial requested scope does not need a fresh whole-codebase audit after each local fix.

Convergence and output

Finish when supported in-scope defects are resolved, affected rechecks find no remaining defects, and required verification supports the acceptance criteria. Report the checked scope and limitations rather than claiming an entire codebase is defect-free.

If the same failure recurs without new evidence, change the hypothesis, reproduce it more directly, or identify the missing dependency. Follow any user-set round or time budget; an exhausted budget or external blocker is an incomplete outcome, not convergence. Ask only for missing authority or a material decision that cannot be inferred, after completing independent work.

Keep progress updates brief: supported findings, fixes verified, and the next uncertainty. Deliver a concise result with changed files, evidence, and any unresolved finding with its blocker. Store a report only when requested or the project's workflow calls for one.