Plan Audit

Plan Audit

Use this workflow only after explicit selection as $plan-audit or a binding task instruction requiring it. It has two modes:

• plan-readiness: improve plan quality before work starts.
• implementation-audit: review implemented code against the plan after code exists.

The plan-readiness job is to make the plan clearer, safer, simpler, more complete, more code-grounded, and easier to implement without building the wrong thing.

The implementation-audit job is plan-backed code review: check whether the implemented code fits the plan's architecture and quality bar without adding duplicate truth, side doors, drift, bad caller shape, or unnecessary complexity. It is specifically skeptical of name-only completion: code that uses the right labels, wrappers, phase checks, or conventions while the intended unification, simplification, deletion, or behavior change is still false underneath.

The skill audits plans. It does not dictate the user's workflow.

Doctrine-Only Constraint

This is a doctrine-only, prompt-first skill. It ships agent guidance, references, metadata, and examples. It must not become a deterministic harness, runner, controller, rule engine, scorer, checklist executor, grep gate, automated architecture validator, test runner, proof collector, truth arbiter, or script-backed readiness judge.

The checklists are judgment aids. The audit log is a durable Markdown review ledger beside the plan, not a state machine and not a second plan.

Use When

• The user wants an existing plan, PRD, migration plan, architecture plan, checklist, issue-body plan, pasted plan, design doc, or strategy audited.
• The user wants stronger planning before implementation or execution.
• The user asks whether a plan is complete, clear, elegant, safe, simple, code-grounded, or ready.
• The user wants ambiguity, constraint, proof, side-door, delete, depth-first, drift, or existing-pattern risks found before work starts.
• The user wants code already written for a plan reviewed against that plan's promises, phase scope, owner path, side-door closure, drift risks, caller shape, and elegance.

Do Not Use When

• The user wants the plan written from scratch.
• The user wants implementation to start.
• The user wants a generic code diff, branch, or PR reviewed without a plan artifact. Use the host agent's normal review response for that.
• The user asks which workflow or skill to use.

Non-Negotiables

• Adapt to the plan's format. Do not require an arch-suite template, section numbering, or a specific checklist shape.
• Read the plan as written. Do not mentally repair it before auditing it.
• For repo-backed plans, read repo truth and all relevant code before approval.
• For non-trivial file-backed audits, create or update <PLAN_STEM>_PLAN_AUDIT.md beside the plan.
• For inline plans with no file path, return the audit in chat; suggest a persistent audit log only when the user wants repeat-audit tracking.
• Apply ../_shared/agent-orchestration-policy.md whenever the audit uses child agents.
• For broad repo-backed audits, use a coverage-led set of independent review slices when the active host supports them. Start every slice as a new clean same-host native child, keep lenses and path families non-overlapping, and bound fanout by host slots, shared-file or shared-state collision risk, and the parent's ability to integrate every return.
• Choose bounded or full inherited context only when a named decision exists solely in chat. Prefer plan, audit-log, and code paths over inheriting the parent's framing or completion story.
• Use the strongest read-only capability the host exposes, also tell every audit child not to edit or write, and have the parent compare repository status and diffs with the pre-dispatch state before accepting child evidence.
• Children may use their own native sub-agents on their own host at will; they may not start external agents.
• The parent owns child accounting, evidence spot-checking, deduplication, integration, finding scope disposition, audit-log updates, and the final verdict.
• Do not mark ambiguity or constraint questions resolved until a decision owner resolves them and the plan carries the decision through.
• Apply ../_shared/scope-and-convergence.md. Plan-readiness audit must recover the human baseline, initial pre-approval convergence closure, and scope sign-off. It may mark the plan not ready and route a gap back to the initial planning owner, but it must never add an adjacent area to required scope.
• After scope sign-off, a new adjacent path requires a human decision. An agent-authored plan revision, audit finding, review consensus, or Decision Log entry is not approval.
• Findings must include consequence, evidence, and the concrete plan repair.
• In implementation-audit mode, do not run unit tests, integration tests, build commands, lint commands, or CI; do not ask for test logs; do not verify whether tests really ran; and do not investigate whether a completion claim is truthful.
• In implementation-audit mode, accept supplied test-pass claims as context. You may read changed test files as code when they matter to code review, but test execution remains out of scope.
• In implementation-audit mode, treat name-only completion, false simplification, and partial unification as code review blockers when current plan and code evidence show the intended outcome is still false.

First Move

1. Resolve the plan artifact: file path, pasted plan, issue text, doc section, PRD, checklist, or named planning artifact.
2. Read local instructions and repo context when the plan is repo-backed.
3. Resolve the audit log path for non-trivial file-backed audits and read the existing audit log if present.
4. Decide the mode from the user's ask:
   • plan-readiness for pre-implementation plan quality audit.
   • implementation-audit for plan-backed code review after code exists.
5. For implementation-audit, read references/implementation-audit-mode.md and follow that mode contract.
6. For plan-readiness, read references/progressive-audit-order.md.
7. Read ../_shared/agent-orchestration-policy.md before creating or resuming any child.
8. Read the smallest additional references needed for the scope:
   • references/architecture-quality-canon.md for the quality bar
   • references/review-lenses.md for lens definitions
   • references/audit-log-contract.md for sidecar ledger rules
   • references/proper-audit-checklist.md before a readiness verdict
   • references/child-prompt-contract.md before native subagent prompts
   • references/output-contract.md before final output

Workflow

If the user asks for implementation-audit, use references/implementation-audit-mode.md. Do not use the plan-readiness workflow below.

1. Audit the artifact in its native format.
2. Extract the North Star, done-state requirements, requirements, non-requirements, constraints, non-constraints, assumptions, and complexity sources.
3. Identify real ambiguity and constraint questions that can change the built outcome.
4. For repo-backed plans, map and read all relevant code: owner paths, callers, comparable patterns, legacy paths, side doors, contracts, tests, docs, prompts, and generated artifacts. When children improve coverage, use a proportional set of new clean native read-only slices with distinct lenses or path families and integrate every result in the parent.
5. Run the required lenses from references/review-lenses.md.
6. Challenge the plan for simpler architecture, fewer live concepts, better ownership, depth-first proof, delete work, and drift-proof coupling.
7. Update the audit log when applicable.
8. Return a findings-first verdict using references/output-contract.md.

Output Expectations

Return a concise, findings-first audit:

• plan-readiness verdict: ready, not-ready, blocked-on-decision, or inconclusive
• implementation-audit verdict: approve, not-approved, or scope-inconclusive
• plan artifact and audit log path when applicable
• required repairs first
• observations or out-of-scope follow-ups only when they matter
• real ambiguity and required decisions
• relevant code read and relevant code not yet read
• proper-audit checklist status
• the smallest next plan repair

Every material finding must include one scope disposition: authorized, approved-convergence-required, new-scope-needs-human, out-of-scope, or unauthorized-built-scope. Only the first two may be required plan or code repairs without another human decision. In implementation-audit mode, unauthorized built scope is a blocking IMP-* subtraction finding even when a later agent-authored plan revision includes it.

Do not approve a plan while required code is unread, the audit log is stale, a required lens could not inspect its scope, or an outcome-changing ambiguity or constraint question remains unresolved or uncaptured in the plan.

Reference Map

• references/architecture-quality-canon.md - strict plan-quality and code-quality doctrine
• references/review-lenses.md - required and conditional audit lenses
• references/progressive-audit-order.md - ordered pass for using the skill
• ../_shared/scope-and-convergence.md - scope authority, pre-approval convergence, finding dispositions, and scope-cycling rules
• references/implementation-audit-mode.md - plan-backed code review mode after implementation
• references/audit-log-contract.md - sidecar audit log shape and loop rules
• references/proper-audit-checklist.md - final "was this audited properly" check
• references/child-prompt-contract.md - native subagent prompts for broad audits
• ../_shared/agent-orchestration-policy.md - transport, starting context, continuation, isolation, topology, and parent-integration policy
• references/output-contract.md - final verdict and finding format
• references/examples.md - examples and anti-examples that teach judgment