Audit an OpenSpec change

Spawn an isolated auditor over one change's artifacts, catch cross-artifact drift before code gets written, and never claim a clean audit the checks did not actually confirm.

Steps

1. Select the change. If not given, run openspec list --json, show active (non-archived) changes with schema, and let the user choose. Never guess.
2. Gather artifacts and convention docs. Run openspec status --change "<name>" --json and openspec instructions apply --change "<name>" --json; collect every artifact path from contextFiles plus .openspec.yaml. If the CLI is unavailable, fall back to filesystem discovery under openspec/changes/<name>/. Locate repo-root convention docs (AGENTS.md, CLAUDE.md, CONTRIBUTING.md) and the schema's context/rules if available from /rhdh-spec-driven-schema.
3. Enforce the apply-ready minimum before doing anything else. Build the union of the schema's apply.requires (or CLI applyRequires) and {proposal, design, specs, tasks}. Preferred: every ID in that union must show status: "done" in the status JSON. Filesystem fallback: read .openspec.yaml for the schema name, read that schema's apply.requires, and check each mapped path exists and is non-empty (proposal->proposal.md, design->design.md, tasks->tasks.md, specs->at least one specs/**/spec.md). If the required set cannot be determined, or any required artifact is missing/not done: report CRITICAL, state "Audit blocked (CRITICAL findings remain)", write/update audit.md with the current UTC timestamp, suggest /openspec-continue-change, and stop — do not build the ownership map or spawn the auditor.
4. Build a cross-change ownership map. Resolve sibling changes (reuse openspec list --json from step 1, or filesystem-list openspec/changes/* excluding this change and archive/). For each sibling, read only its proposal.md Canonical Touchpoints section and design.md Decisions section, and build a lightweight capability/namespace/claim -> [change names] map. Pass the map (or condensed claim lines), not full sibling trees, to the auditor. If sibling discovery fails entirely, pass an empty map and note category E ran degraded in the final report — do not fail the whole audit for this alone.

5. Spawn an independent auditor subagent (Task tool, subagent_type: general-purpose). Give it only: change name, absolute artifact paths, the ownership map, convention doc paths, the checklist below, and the required JSON return schema. Explicitly forbid passing authoring rationale or "what we meant" — this is prompt isolation, not a sandbox; instruct it to limit reads to the given paths. Require it to return:

   {"findings": [{"category": "A|B|C|D|E|F|G|H", "severity": "CRITICAL|WARNING|SUGGESTION", "file": "openspec/changes/<name>/design.md", "line": 1, "quote": "short excerpt", "recommendation": "actionable fix", "autofixable": true}]}

   file must be repo-root-relative; normalize any absolute paths before writing audit.md. Fail-closed: if the subagent fails, times out, or returns malformed findings, treat it as CRITICAL — write/update audit.md with that CRITICAL and the current UTC timestamp, state "Audit blocked (CRITICAL findings remain)", report the failure, suggest retrying, and stop. Only a parseable {"findings": [...]} — including an empty array — may proceed to the fix loop.

6. Auditor checklist (all categories evaluated):

   ID	Category	Catches	Autofixable
   A	Entity propagation	Name/value fixed in design.md (or proposal.md) not identically propagated to sibling artifacts	Yes, if one unambiguous spelling
   B	Enum / vocabulary	Enumerated sets with different membership/count across files or vs openspec/specs/<capability>/	Yes, if superset unambiguous
   C	Semantic contradiction	Design fails to satisfy spec WHEN/THEN; Non-Goals vs Decisions; non-deterministic outcomes; Verify scoped differently than the task	No
   D	Codebase & convention grounding	References to APIs/tickets/modules that don't exist; mechanisms violating loaded house rules	No
   E	Namespace & cross-change ownership	Unresolved openspec/changes/<id>/ refs; claim shared with a sibling with no reconciliation note	No
   F	Template/copy-paste residue	Literal paths/prefixes from a different project than this repo	Yes, if correct target unambiguous
   G	Extended coherence	Helper signature vs call sites; shared exports not tasked into consumers; lifecycle semantics missing/inconsistent across artifacts	Yes when a name/signature is already fixed and merely missing in siblings; no for any WHEN/THEN or product-policy choice
   H	Security lint	Mutating/admin routes with no authorize story; ambiguous DENY outcomes; hidden-count exposure; secrets in examples	Only if design already named the permission and wiring is obvious

   Never hardcode product-specific banlists — categories D and H load constraints from project convention files and schema context. When uncertain, prefer SUGGESTION over WARNING, WARNING over CRITICAL.

7. Fix-and-reaudit loop, up to 3 passes. Collect autofixable findings with one unambiguous canonical value (prefer the value in design.md Decisions; else the most frequent spelling agreeing with the proposal's Capabilities names). If none, stop the loop. Otherwise list the proposed mechanical fixes and ask the user: apply this batch, or skip autofix (report-only for remaining passes). On confirm, apply edits and re-spawn the auditor (same isolation and fail-closed rules); on skip, break and proceed to reporting with current findings. Never auto-apply non-autofixable (judgment-call) findings, and never apply any fix without explicit confirmation.
8. Generate the audit report. Write openspec/changes/<name>/audit.md (and show the same content in chat) using references/audit-report-template.md, keeping its section structure. Always set Last audited to the current UTC time when writing. Every finding needs a repo-root-relative file reference and an actionable recommendation. A severity section with no findings must contain exactly - None.
9. Report the outcome. Any CRITICAL remaining: state "Audit blocked (CRITICAL findings remain)" and list blockers — never claim a clean audit. No CRITICAL: state "Audit clean (no CRITICAL)" (WARNINGs may still be listed) — never say "Ready for implementation." Always close with: "Fixes to one file can introduce new drift elsewhere — re-run the audit after resolving remaining findings." This is advisory output; it does not gate /openspec-apply-change or /openspec-archive-change.

Guardrails

• Never skip the independent auditor subagent for a full audit.
• Fail-closed on subagent failure — never claim a clean audit from a failed or malformed response.
• Never invent product-specific rules in this skill's own body — load them from project files.
• Never pivot into design-decision grilling for unresolved findings — report them only.
• Never apply autofixes without explicit confirmation; cap confirmed autofix loops at 3 passes.
• Always run the apply-ready minimum check before the ownership map or subagent — an incomplete change is CRITICAL and stops the audit early.

Completion

Complete when audit.md reflects the latest auditor pass (or the latest CRITICAL blocker) with a current UTC timestamp, every finding carries a repo-root-relative file reference and recommendation, and the user has been told explicitly whether the audit is blocked or clean — never left to infer it.