implement-audit-findings

Implement confirmed Ergopti AHK or Hammerspoon audit findings in the dedicated worktree, one local commit with regression coverage per fix. Use for "implémente l'audit", "corrige l'audit AHK", or "cor

SKILL.md

GitHub

Implement driver audit findings

Turn a validated audit manifest into independently reviewable local commits. The manifest is the queue; do not create a second narrative plan unless the user asks for one. Never push, merge, stash, clean, reset, move a worktree, or modify the main checkout's unrelated files.

Select and validate the audit

Use the report the user named, otherwise the newest strict dated directory for the requested scope:

node tools/audit/workflow.cjs validate-report --report <findings.json>
node tools/audit/workflow.cjs preflight --scope <ahk|hammerspoon>
node tools/audit/workflow.cjs status --report <findings.json>

Canonical worktrees are sibling directories named from the repository: <repo>-fix-ahk and <repo>-fix-hs. preflight recognizes an already registered worktree and reports it as resumable. It never creates or modifies one. If the canonical path is absent but another active worktree owns a branch for that scope, stop; do not create a duplicate or migrate it implicitly.

When preflight returns missing, first verify that its exact expected_path is absent and that the intended branch name does not already exist. Then create that one worktree from the manifest's audited_sha with an OS-neutral Git command:

git worktree add -b fix/<ahk|hammerspoon>-audit-<YYYY-MM-DD> <expected_path> <audited_sha>

Run preflight again and continue only when it reports ready. An existing branch, directory, competing worktree, or different base is user state: inspect and stop rather than deleting, resetting, force-creating, or silently reusing it.

An existing dirty canonical worktree is not disposable clutter. Inspect its branch and diffs and map them to a finding before editing. If ownership is ambiguous or changes span findings, report that state and stop. Never use stash as a shortcut: stashes are repository-global and these worktrees are active.

Implement one finding

Read windows-toolchain, verify-change, ship-fix, commit-and-push, and the matching driver skill. Add cross-driver-parity, i18n, logger, or meta-test only when the finding touches those surfaces.

For the next open finding:

  1. Extract only that record with workflow.cjs extract; do not repeatedly load the complete report.

  2. Reproduce and re-derive it against current code and tests. The audit severity is not proof. Record a refutation instead of forcing a stale fix.

  3. Add the narrow regression test first and observe it fail for the claimed root cause before changing production. If resumed dirty source already contains a partial fix, do not rewrite or stash it merely to manufacture fail-before evidence; state the evidence gap explicitly.

  4. Fix the root cause and every sibling site governed by the same invariant.

  5. Run the failing test again, then every gate selected by verify-change.

  6. Stage exact paths and commit exactly this finding plus its regression test. Use an English Conventional Commit message and one trailer:

    Audit-Finding: AHK-001
    
  7. Run workflow.cjs verify-commit --report <...> --id <...> --commit HEAD. Then refresh status, which derives completion from Git rather than editing the audit manifest.

Repeat sequentially. One agent owns one worktree; do not parallelize commits or let multiple agents mutate the same branch. Durable new knowledge goes in docs/memory/; final documentation or audit retirement is a separate commit.

Completion boundary

Stop after all still-valid findings have a verified commit and the selected gates pass. Report refuted/stale findings, environment deferrals, and real-driver validation still required. Leave commits local. Do not push or merge even when the queue is empty.