karlorz/agent-skills/ranked-audit

Dev-Loop Ranked Audit

Run a bounded cross-project lifecycle audit without changing repositories or work-item status. The scan may publish one managed evidence report; it must not apply lifecycle corrections.

Hard Rules

1. Treat repositories and work items as read-only evidence sources.
2. Do NOT change work-item lifecycle status, checkboxes, evidence, plans, logs,

project indexes, branches, tags, releases, or deployments.

1. Do not prompt. Ranked audit must run under interactive, /goal, scheduled,

or satellite contexts with the same mechanical behavior.

1. Bound the scan with --top <n>; default to 20.
2. Use repository evidence only through the host-aware project-repos.yaml

resolver. Never borrow one current checkout as evidence for every project.

1. Publish only the audit report through SkillWiki's managed publisher. If the

publisher is unavailable, leave the draft outside the vault and report the blocked publication.

Inputs

/dev-loop ranked-audit
/dev-loop ranked-audit --top 20

Resolve:

• <vault> through the configured SkillWiki backend or skillwiki path.
• <project-repos> from

projects/llm-wiki/architecture/project-repos.yaml, with the selected project's architecture fallback when the coordinator file is absent.

• <host-id> from SkillWiki fleet context or the configured host identity.
• <repo-user> from the runtime user.

Scan

Run the deterministic helper from the dev-loop plugin root:

node scripts/ranked-audit.js \
  --vault <vault> \
  --top <n> \
  --project-repos <project-repos> \
  --host-id <host-id> \
  --repo-user <repo-user>

The helper reuses preflight-inventory.js, preserves its priority/status ordering, resolves repository evidence per selected project, and emits:

• delivered-close-candidate
• active-code-work
• verification-only
• human-gated
• stale-or-superseded
• unverifiable

When signals overlap, classification precedence is verification-only, human-gated, delivered-close-candidate, unverifiable, stale-or-superseded, then active-code-work.

Require read_only: true and writes_executed: false. Stop if either field is missing or false.

Evidence Report

Create a complete typed-page draft outside the vault. Target:

meta/YYYY-MM-DD-ranked-audit-top-<n>.md

Include:

• command inputs, host identity, and repository-resolution degradations;
• the ranked candidate table with project, lifecycle status, priority,

classification, reasons, evidence files, and matching commits;

• counts by classification;
• explicit confirmation that no work-item or repository mutation ran;
• the attended continuation command:

/dev-loop office-hours <audit-report> --reconcile.

Publish through skillwiki page publish: dry-run, review the state-bound token, then write with the same draft and token. Do not edit root index.md or log.md directly.

Exit

Print the report path and classification counts. If no candidate needs human reconciliation, say so and do not recommend office-hours merely to create work.