| name | audit-agent-plan |
| description | Independently audit SDD specifications and agent-ready plans before freeze, returning findings, readiness verdicts, and required refinements. |
Audit agent plan
Use this skill for independent review of a candidate specification or
agent-ready plan. The default action is audit only.
Inputs
Read the developer overview first, then the specification, plan manifest,
write-set contract, worker DAG, acceptance checklist, evidence plan, budget
profile, and previous review if this is a revision.
Review matrix
Check every item for:
- traceability from requirement to acceptance criterion, workstream, evidence,
and final audit gate;
- SDD tier correctness;
- missing clarification or unsafe assumption;
- architecture, security, performance, data, browser, release, or external
environment risks;
- exact ownership and forbidden writes;
- dependency DAG correctness and parallel safety;
- validation commands that are executable, scoped, bounded, and sufficient;
- context and token budget fit for small and large context hosts;
- deterministic SDD tier resolution evidence or a justified higher-tier
override;
- plan/freeze/run artifact separation;
- adapter neutrality and absence of project-specific assumptions unless the
target project explicitly owns them.
When available, run agent-lifecycle plan completeness-check --manifest <plan.manifest.json> and treat agent-plan-completeness-validation.v1 blockers
as deterministic pre-audit findings. For freeze gates that opt in to structural
enforcement, also check agent-lifecycle plan check --require-completeness.
Verdicts
READY_TO_FREEZE: no open Medium or High finding; freeze inputs are complete.
CHANGES_REQUIRED: the plan can be corrected within the current intent.
BLOCKED: external authority or missing product decision prevents a safe
plan.
REOPEN_REQUIRED: a frozen plan needs contract, ownership, or intent changes.
Rules
- Findings first; do not hide blockers in prose.
- Do not rewrite the plan during independent review unless explicitly asked.
- Do not approve a plan that lacks a developer overview, write ownership,
evidence gates, or final audit route.
- Do not ask for longer prose when the defect is missing structure. Require the
smallest tier-appropriate field or receipt route that closes the blocker.
- Do not accept self-review as independent review.
- Do not downgrade required runtime, security, or performance evidence to
synthetic-only checks unless the plan explicitly narrows the claim.
Output
Return:
- findings ordered by severity;
- coverage matrix;
- missing or excessive scope;
- validation of budgets and context fit;
- freeze verdict;
- exact edits required before the next review.