Run a long-horizon audit-then-implement pipeline: produce a structured audit, cluster findings into PR-sized chunks, and spawn parallel subagents to implement fixes as draft PRs.
Trigger: when the user asks for "audit and fix", wants to clean up a domain, or kicks off an overnight cleanup run. Also triggered directly via /audit-and-fix <domain>.
Arguments: $ARGUMENTS — the audit domain (e.g. "hardcoded business logic", "unused exports in moneyballer/pkg/db", "deprecated Storage API calls"). Required — if empty, ask the user.
Steps
Run the audit. Invoke the /create-audit skill. This produces docs/audits/<topic-slug>-<YYYY-MM-DD>.md with Findings and Issues tables. Commit the audit file on a new branch audit/<topic-slug>-<date> and push.
Cluster findings into PR groups. Read the audit's Issues table and group findings by:
- Same file or package → same cluster
- Same severity AND same fix pattern → same cluster
- Cross-cutting infra changes (test harness, CI config) → their own cluster
Produce 4–8 clusters, not 30 single-finding PRs. Each cluster is one PR. Print the clustering plan as a table before dispatching:
| Cluster |
Findings |
Blast Radius |
Recommended Title |
| 1 |
3 |
low |
"Remove unused exports in pkg/db" |
Initialize the status file.
# Audit-and-Fix Status — <topic> — <YYYY-MM-DD>
Audit: docs/audits/<topic-slug>-<date>.md
| Cluster | PR | Branch | State | Last Update |
| ------- | --- | ------ | ------- | ----------- |
| 1 | — | — | pending | <timestamp> |
Write to docs/audits/<topic-slug>-<date>-status.md.
Dispatch parallel cluster implementers. Use the Agent tool with subagent_type: "general-purpose" in a single message with multiple tool calls — one agent per cluster, cap at 6 concurrent. Each subagent receives:
- The audit file path and the specific findings in its cluster
- Instructions: create a worktree off
main, implement the fixes, write or update tests, open a draft PR with required sections (Summary, Test plan, Spec ID if applicable), wait for CI, address Copilot comments autonomously
- A strict rule: never merge. Leave the PR as draft when done.
- Required output:
{cluster, pr_url, branch, final_state} where final_state ∈ {ready, ci-failing, review-pending, blocked}
Monitor and update the status file every 5 minutes while subagents run. For each cluster, record: PR URL, branch, current state, last-update timestamp. If a subagent fails (worktree disappears, CI times out), mark its cluster blocked and record the reason rather than retrying silently.
Post the final summary table when all clusters reach ready or blocked:
| Cluster |
PR |
State |
Notes |
| 1 |
#201 |
ready |
0 Copilot comments |
| 2 |
#202 |
blocked |
rebase conflict on pkg/foo.go |
Do not merge. The user decides which PRs to merge (use /merge-pr per PR).
Rules
- Never merge. This command produces draft PRs only.
- Never force-push. Never use destructive git operations in subagents.
- If a worktree path disappears mid-run, recreate it — do not fail the whole run.
- Cap parallelism at 6 concurrent subagents. Batch the rest.
- Every draft PR must include the audit file path in its Summary section so reviewers can trace the finding.
- Keep the status file updated even on failures — silence is worse than partial information.
- If clustering produces more than 8 PRs, STOP and ask the user to narrow the audit scope.
Compute scope
The Frontier floor describes the coordinator's consumer requirement. Native skill
frontmatter does not enforce that floor. Keep the existing general-purpose worker
dispatch, inheritance, tools, worktrees and draft-only merge prohibition. This
declaration adds no enforced worker-floor guarantee and does not set a worker's
model or effort. Report coordinator and worker observations separately when they
are available; an unavailable observation remains unknown.
1---2# dotbabel:generated projection=sha256:1d563fb0cc146487036ea6bedd894815cafec7a24c87e109f39fa44b9fd23d79 source=skills/audit-and-fix/SKILL.md — do not edit directly3argument-hint: "[domain]"4description: |5 Run an audit-then-implement pipeline: produce an audit, cluster findings into PR-sized chunks, and spawn parallel subagents to implement fixes as draft PRs. Triggers on: "audit and fix", "clean up domain", "overnight cleanup".6name: audit-and-fix-37metadata:8 source: https://skillmd.com/skills/kaiohenricunha/audit-and-fix-39---1011Run a long-horizon audit-then-implement pipeline: produce a structured audit, cluster findings into PR-sized chunks, and spawn parallel subagents to implement fixes as draft PRs.1213Trigger: when the user asks for "audit and fix", wants to clean up a domain, or kicks off an overnight cleanup run. Also triggered directly via `/audit-and-fix <domain>`.1415Arguments: `$ARGUMENTS` — the audit domain (e.g. "hardcoded business logic", "unused exports in moneyballer/pkg/db", "deprecated Storage API calls"). Required — if empty, ask the user.1617## Steps18191. **Run the audit.** Invoke the `/create-audit` skill. This produces `docs/audits/<topic-slug>-<YYYY-MM-DD>.md` with Findings and Issues tables. Commit the audit file on a new branch `audit/<topic-slug>-<date>` and push.20212. **Cluster findings into PR groups.** Read the audit's Issues table and group findings by:22 - Same file or package → same cluster23 - Same severity AND same fix pattern → same cluster24 - Cross-cutting infra changes (test harness, CI config) → their own cluster2526 Produce **4–8 clusters**, not 30 single-finding PRs. Each cluster is one PR. Print the clustering plan as a table before dispatching:2728 | Cluster | Findings | Blast Radius | Recommended Title |29 | ------- | -------- | ------------ | --------------------------------- |30 | 1 | 3 | low | "Remove unused exports in pkg/db" |31323. **Initialize the status file.**3334 ```markdown35 # Audit-and-Fix Status — <topic> — <YYYY-MM-DD>3637 Audit: docs/audits/<topic-slug>-<date>.md3839 | Cluster | PR | Branch | State | Last Update |40 | ------- | --- | ------ | ------- | ----------- |41 | 1 | — | — | pending | <timestamp> |42 ```4344 Write to `docs/audits/<topic-slug>-<date>-status.md`.45464. **Dispatch parallel cluster implementers.** Use the `Agent` tool with `subagent_type: "general-purpose"` in a **single message with multiple tool calls** — one agent per cluster, cap at 6 concurrent. Each subagent receives:47 - The audit file path and the specific findings in its cluster48 - Instructions: create a worktree off `main`, implement the fixes, write or update tests, open a **draft** PR with required sections (Summary, Test plan, Spec ID if applicable), wait for CI, address Copilot comments autonomously49 - A strict rule: **never merge.** Leave the PR as draft when done.50 - Required output: `{cluster, pr_url, branch, final_state}` where `final_state ∈ {ready, ci-failing, review-pending, blocked}`51525. **Monitor and update the status file** every 5 minutes while subagents run. For each cluster, record: PR URL, branch, current state, last-update timestamp. If a subagent fails (worktree disappears, CI times out), mark its cluster `blocked` and record the reason rather than retrying silently.53546. **Post the final summary table** when all clusters reach `ready` or `blocked`:5556 | Cluster | PR | State | Notes |57 | ------- | ---- | ------- | ----------------------------- |58 | 1 | #201 | ready | 0 Copilot comments |59 | 2 | #202 | blocked | rebase conflict on pkg/foo.go |6061 Do not merge. The user decides which PRs to merge (use `/merge-pr` per PR).6263## Rules6465- Never merge. This command produces draft PRs only.66- Never force-push. Never use destructive git operations in subagents.67- If a worktree path disappears mid-run, recreate it — do not fail the whole run.68- Cap parallelism at 6 concurrent subagents. Batch the rest.69- Every draft PR must include the audit file path in its Summary section so reviewers can trace the finding.70- Keep the status file updated even on failures — silence is worse than partial information.71- If clustering produces more than 8 PRs, STOP and ask the user to narrow the audit scope.7273## Compute scope7475The Frontier floor describes the coordinator's consumer requirement. Native skill76frontmatter does not enforce that floor. Keep the existing general-purpose worker77dispatch, inheritance, tools, worktrees and draft-only merge prohibition. This78declaration adds no enforced worker-floor guarantee and does not set a worker's79model or effort. Report coordinator and worker observations separately when they80are available; an unavailable observation remains unknown.