audit

name	audit
description	Audit a feature implementation against its approved plan and implementation report, then produce the architect verdict and follow-up direction.

Skill: /audit

Purpose: Audit Developer's implementation report as Architect and provide final decision plus follow-up direction. Hermetic per feature; agent doc context = feature.

Input Formats

Choose any of the following:

/audit                                   # No args: auto-detect from chat history
/audit feat-2026-01-123                 # With FEATURE_ID: infer plan-slug from feature dir
/audit plan                              # Artifact ref: use current feature's latest plan + implementation
/audit I{n}                              # Implementation ref: audit latest implementation round
/audit feat-2026-01-123 1-refactor      # FEATURE_ID + plan-slug explicitly

When no args: auto-detect from chat — find most recent FEATURE_ID reference and most recent implementation report, or validate if already in feature context.

Instructions for user

• Minimal case: Just say /audit — skill will infer current feature and plan from chat.
• With artifact: /audit plan (use current implementation), /audit I{n} (reference specific round).
• With FEATURE_ID: Provide feature ID if context is unclear; skill will find the latest plan/implementation.
• Implementation report must exist at {feature-dir}/reports/{plan-slug}.implementation.md before audit can proceed.

Instructions for model

Input Parsing

Parse the user's input using this order:

1. Extract FEATURE_ID from:

   • Explicit arg (e.g. feat-2026-01-123)
   • Chat history (most recent feature reference)
   • Current feature context (if already in feature namespace)

   If unable to determine: Prompt user for FEATURE_ID before proceeding.

2. Infer plan-slug from:

   • Explicit arg (second positional after FEATURE_ID)
   • Feature directory structure (find most recent approved plan in plans/ directory)
   • If multiple plans exist: prompt user which plan to audit
   • If exactly one plan exists: use it

3. Extract implementation round reference (optional):

   • I{n} → audit specific Implementation Round n
   • plan → audit latest implementation round for this plan
   • Not specified → audit latest round in implementation report

State Detection (HARD GATE — must read file before proceeding)

After input parsing:

1. You MUST read the actual file {feature-dir}/reports/{plan-slug}.implementation.md from disk. Do NOT assume it exists based on chat history or inference.
2. If missing: output error "Cannot audit: implementation report not found at reports/{plan-slug}.implementation.md. Run /implement first." — STOP immediately. Do NOT run /implement, do NOT fall back to any other action.
3. Load the implementation report and identify the implementation round(s) to audit.
4. Confirm all prerequisites (plan file, contexts if referenced) are accessible.

---

You are the Architect (architect.md). Doc context = this feature (feat doc + plans/ + notes/ + reports/).

Input: Feature context + plan slug (required). Implementation report path = reports/{plan-slug}.implementation.md. Plan path = plans/{plan-slug}.plan.md.

---

Mandatory context load (before audit)

1. Read plans/{plan-slug}.plan.md — sections ## Retrospective and user-only ## Decision Log. Extract all ### v{N} blocks and explicit user choices/corrections. Use as context — especially [risk] entries and user decisions that shaped the plan.
2. Read reports/{plan-slug}.implementation.md — sections ## Retrospective and user-only ## Decision Log. Extract all ### I{n} blocks and explicit user choices/corrections. Use as context — know what the Developer learned, what risks were flagged, and what operations/patterns the Developer identified as candidates for automation.
3. Use both to inform the audit — assess whether known risks materialized, whether Developer learnings are reflected in the outcome, whether unresolved retrospective findings affect implementation quality, and whether the implementation contradicts explicit user decisions.
4. Context documents: If the plan lists contexts in ## Contexts, read those context documents from {feature-dir}/contexts/. Note any requirements defined there — these will be checked in the Audit Process (step 4a below).

---

Audit Process

1. Read implementation report: Extract all ## Implementation Round I{n} sections — step results, issues, out-of-plan work.

2. Compare against plan: Check each plan step against implementation round results.

3. Check acceptance criteria: For each criterion in plan — met / not met / partial.

4. Re-evaluate risks: Which plan risks materialized? Which were mitigated?

   • 4a. Context conformance (if contexts exist): If feature has context documents with requirements (loaded in Mandatory context load step 4) — check that the implementation conforms to those requirements. Flag any non-conformance as a finding. Non-conformance with a requirement context is grounds for "needs fixes" or "re-plan" verdict.
   • 4b. Automation suggestions (if present): Review any [automation] entries from implementation Retrospective. Synthesize them — if multiple rounds noted similar repeating patterns, note this as a potential skill candidate in your audit summary (does not affect verdict, but informs retrospective). Examples: "Developer observed manual validation step repeating in 3 places — skill candidate"; "Config pattern X appears in implementation steps — consider parametric skill for similar future features".
   • 4c. User decision conformance (if present): Review user-only ## Decision Log entries from plan and implementation. Check only whether implementation or report output contradicts explicit user choices/corrections. Do not expect priority labels in Decision Log; it is not a backlog and not a source for codebase issues. If a contradiction affects requested scope, artifact format, language, placement, or validation/reporting expectations, include it as an audit finding.
   • 4d. Unresolved retrospective findings (if present): Review ## Retrospective entries from plan and implementation. Check whether unresolved [risk], [codebase], [plan], [process], or [tooling] findings affect implemented behavior, maintainability, validation confidence, or future work. Infer audit severity from evidence and impact; do not require source entries to carry priority labels.

5. Implementation quality check (mandatory): Review the actual code changes listed in the implementation report (## Files Changed). Read the modified files (or relevant diffs) and assess:

---

Output

1. Architect audit report written to {feature-dir}/reports/{plan-slug}.audit.md per audit-format.md.
2. problems.md (if verdict ≠ ready): write {feature-dir}/notes/{plan-slug}.problems.md per the problems.md format in audit-format.md. This is the primary handoff document to the next agent (fixer / replanner). Include one AC block per non-passing criterion.
3. Structured response in chat using Architect Audit Response Format — see response-formats.md.
4. Plan update (if decision is "ready"): Update plan file plans/{plan-slug}.plan.md:
   • Change status from active to implemented
   • Add **Status**: done to each Step that was completed (based on implementation report)
   • Mark all acceptance criteria checkboxes as [x] if they were met
   • Keep plan metadata (version, round) unchanged
5. Feature document update (if decision is "ready"): Update feature document {feature-dir}/{feature-id}.md:
   • Find or create ## Plans
   • Match plan entries by {plan-slug}
   • If plan NOT listed: add new entry with format: - [{plan-slug}](plans/{plan-slug}.plan.md) — {title} · status: implemented
   • If plan IS listed: replace/update existing slug entry with the same format
   • Do not mutate unrelated feature sections or frontmatter
6. Suggested git commit message (if decision is "ready"): Output a message suitable for the application repository (the code repo), not for workflow/docs that live outside git.
   • Subject line: feat: {1-sentence summary} — from what actually changed in the repo + ## Steps Executed / ## Files Changed; max 72 chars
   • Body: bullet list, imperative mood, one bullet per meaningful change (aligned with where it maps to code)

Next-step chooser handling

After emitting Architect Audit Response, halt and wait for the user's next input. The response must end with:

---
**Active context**: role: Architect · feat: {feat-id} · plan: {plan-slug} · audit: A{round} · skill: /audit

When the next user input is exactly one of the offered chooser keys:

• [t] / t (ready only): run /retro {feature-dir} in the current Architect context. Do not launch a subagent by default.
• [x] / x (ready only): finish and stop.
• [i] / i (needs-fixes only): launch a Developer subagent (developer.md) with /implement {feature-dir}/plans/{plan-slug}.plan.md A{round} in refinement mode. Pass the audit report, problems.md, implementation report, and active audit round as context. If subagent launch is unavailable, print only the copyable /implement command from the response template and stop.
• [q] / q (needs-fixes only): launch QA Automation subagent (qa-automation.md, subagent_type="qa-automation") for validation/test-evidence analysis of audit concerns. QA returns assessment only; Architect decides whether it changes the audit follow-up direction.
• [r] / r (plan-fix or re-plan): stay in Architect context and run /replan {feature-dir}/plans/{plan-slug}.plan.md A{round}. If the host cannot execute /replan inline, print only the copyable /replan command from the response template and stop.
• [d] / d (plan-fix or re-plan): stay in current Architect context and clarify whether this is a local plan fix or broader re-plan before editing artifacts. Do not edit artifacts until the user confirms direction.
• [v] / v (needs-fixes, plan-fix, or re-plan only): restart /audit as a skill invocation on the same implementation artifact after external changes. Do not continue with free-form Architect audit analysis in place. Append a new audit round; do not overwrite the existing round. If the implementation round is not explicit, read the implementation report and use the latest implementation round. If the host cannot execute /audit inline, print only the copyable /audit command from the response template and stop.
• Any key not offered for the current verdict: reprint the valid chooser keys for that verdict and stop.

Do not treat arbitrary custom user messages as chooser input. If the user writes a normal instruction instead of a chooser key, handle it as normal Architect/audit context and do not auto-launch workflow agents.

Post-audit loop continuation

CRITICAL — Execution context: When the user invokes /audit directly (inline mode, no "loop" or "auto"), execute the audit in the current context — you ARE the Architect auditor. Do NOT launch a subagent for the audit itself. After producing the audit artifacts, show the verdict-aware next-step chooser and stop.

Loop/auto mode only: If the user invoked this command in loop mode (said "loop" or "auto", or called /loop), continue automatically after writing the audit artifacts:

Verdict	Action
ready	Report DONE. Output commit message. Stop.
needs-fixes	Spawn Developer subagent (refinement mode) using problems.md. Wait. Then spawn fresh Architect subagent for re-audit. Max 2 needs-fixes iterations total.
plan-fix	Report BLOCKED. Do not auto-continue — plan repair still requires user decision before /replan or any later implementation.
re-plan	Report BLOCKED. Do not auto-continue — re-plan requires user-assisted scope decision unless user explicitly delegated full authority.

Fresh-session constraint: The re-audit subagent must be a different subagent instance from the Developer that just ran. Never re-audit in the same context that just implemented.

---

Cross-document links and trailing refs: see cross-document-links.md.

Validation (before completing)

• Workflow-doc validation (Validator subagent): After reports/{plan-slug}.audit.md is written on disk, launch a Validator subagent in a separate context. Use the handoff prompt in /validate skill § Validator subagent (delegation): absolute {ABS_PATH_TO_VALIDATE_SKILL} and absolute path to the audit file. Do not complete until exit 0. PostToolUse hook warnings ≠ PASS — see hook-integration.md.
  • (Concrete tooling — if “spawn a Validator subagent” is unclear in your host) Start a separate delegated task (e.g. Cursor Task) so validation runs outside this Architect (audit) thread — commonly subagent_type="generalPurpose" or the same type your /loop skill uses for one-shot handoffs. Delegated body = only the filled handoff prompt from /validate skill § Validator subagent; do not paste audit verdict narrative, findings, or implementation excerpts — only validation instructions.
  • When /loop already documented successful validation for this revision, you may skip nested Validator — state that.
  • Inline fallback (no subagent): open the validate skill and complete one run as defined in that skill; label inline fallback.
• Verify audit report includes conformance assessment (done | partial | no)
• Verify implementation quality check is present (pass | concerns | fail) with concrete findings if concerns/fail
• Verify acceptance criteria are checked (met count/total)
• Verify risk re-evaluation is included
• Verify decision (ready | needs-fixes | plan-fix | re-plan) with reasoning
• Verify report follows audit-format.md
• If decision ≠ ready: Verify problems.md is written to

See also: Use /loop skill for automated post-audit routing.