Review Plan — implementation plan quality audit

Audit an existing plan.md in two layers:

1. Mechanical via scripts/plan-lint.py — required sections, per-task field

presence, dependency resolution, no vague language. Deterministic; no LLM judgment.

1. Judgment via the risk-assessor agent — anti-pattern scanning, architecture

concerns, rollback quality, checkpoint placement, TDD strategy.

This is the quality gate between planning and execution. It does not write durable docs and does not modify the plan unless explicitly authorized (Step 5).

Workflow

1. Locate the plan

If a path was provided, read it. Otherwise detect the plan in order: docs/plan.md, docs/planning/current-execution.md, docs/planning/current.md, root PLAN.md. Read it completely.

If none exists, report that there's no plan to review and recommend /rad-planner:plan.

2. Mechanical lint

Use python3, or python on Windows:

python3 ${plugin_root}/scripts/plan-lint.py <plan-path> --json

Parse the JSON. It reports missing required sections, tasks missing any of the six fields (Objective, Files, Depends on, Done when, Validate, Rollback), unresolved or cyclic dependencies, and vague language. Exit 1 = issues, 0 = clean. These findings are deterministic and bypass the risk-assessor — they feed straight into the report.

3. Delegate to risk-assessor

Use the Agent tool with the substituted template from references/subagent-prompts/risk-assessment.md. Pass the plan content, iteration_number: 1, max_iterations: 1 (review-plan is a single pass; the iterative loop belongs to /rad-planner:plan), and a note that the mechanical layer already ran — focus judgment on what scripts can't check (anti-patterns, architecture, rollback quality, checkpoint placement, TDD strategy).

If durable context docs exist and are relevant (e.g. a docs/prd.md, docs/architecture.md, a decision log), read them read-only and pass them as supporting context so the assessor judges the plan against stated product/architecture intent. Do not require them — many projects won't have them.

Validate the agent's JSON:

echo "$AGENT_OUTPUT" | python3 ${plugin_root}/scripts/validate-json.py \
  ${plugin_root}/references/subagent-prompts/risk-assessment.schema.json - --extract-from-markdown

Re-prompt once on schema failure, then fall back to markdown parsing per agents/risk-assessor.md. Key fields: verdict (APPROVE | REVISE | RETHINK), blocking_issues[], advisory_issues[], escalation_required.

4. Present results

# Plan Review Report

**Plan:** [path]
**Mechanical lint:** [PASS | N issues]
**Risk-assessor verdict:** APPROVE | REVISE | RETHINK
**Overall recommendation:** [actionable next step]

## Mechanical issues (from plan-lint.py)
- [deterministic — no judgment debate]

## Critical issues (block execution until fixed)
- [issue with specific fix]

## Improvements (recommended before execution)
- [issue with fix]

## Optional enhancements
- [nice-to-have]

## Escalation
[only when verdict=RETHINK — recommend /rad-brainstormer:design-sprint]

Apply production-grade standards by default: every task needs both Validate and Rollback; auth/payment/data tasks need a security checkpoint; external integrations need error handling in the task spec. Treat gaps in these as blocking.

Horizon rule: task-level detail is required only for the "Now" horizon. A Release map whose "Next"/"Later" entries are outlines and themes is correct by design — do not flag them as under-specified. A ## Shipped section is history — out of review scope entirely.

5. Offer fixes (only when authorized)

If verdict is REVISE, offer to fix issues directly in the plan file. For each fix, explain the change, show before/after, and modify only plan.md — never create implementation files, never write durable docs. Mechanical issues (missing field, vague phrase) usually have unambiguous fixes; offer to apply those directly.

If verdict is RETHINK, do not offer fixes — the architecture has fundamental issues task-level edits won't resolve. Recommend /rad-brainstormer:design-sprint, then re-review.

What this skill does NOT do

• Does not modify the plan unless authorized in Step 5.
• Does not write durable docs (PRD, architecture, decision log) — if the review surfaces

a durable change, name it in the report for the user to apply.

• Does not test that the plan works — that's the executor's job.
• Does not re-run the iterative risk loop — that belongs to /rad-planner:plan.