review-plan（审查计划）

Run a fresh reviewer round against the current task plan, revise the <plan> section after every in-scope VERDICT: REVISE, and repeat until the plan reaches VERDICT: APPROVED, the loop hits 5 rounds, or the task must return to /apex:research or /apex:plan because the inputs are too weak for safe review.

Target: 2-3 rounds. Hard max: 5 rounds.

You can find tasks in ./apex/tasks/ or run with: /apex:review-plan [identifier] Load task file, extract the current plan, and start round 1.

1. Create a compact round summary for the reviewer:

- Current plan snapshot - Contract and research constraints - Prior round findings and what changed since then

1. Spawn exactly one fresh subagent for the round
2. Instruct the subagent:

- This is a read-only review round - Do not edit files - Review the plan as an adversary, not a collaborator - Inspect repo files if needed to verify feasibility - Focus on: - mandatory artifact completeness - architecture soundness - contract coverage - contract version / amendment acknowledgement - security concerns - edge cases and failure modes - concurrency, state, ordering, and race risks - pattern provenance and trust alignment - documentation drift - non-functional requirement coverage - validation quality and realism - implementation sequencing and dependency order - research-to-plan mismatches

1. Require the subagent to return this structure:

ROUND: [N]
VERDICT: [REVISE|APPROVED]
NEXT ACTION: [CONTINUE|RETURN_TO_RESEARCH|RETURN_TO_PLAN|IMPLEMENT]

MUST FIX
- ...

NON-BLOCKING
- ...

SUGGESTED PLAN EDITS
- ...

APPROVAL SUMMARY
- ...

Rules for the verdict:

• APPROVED only when there are no blocking issues left
• REVISE when any blocking issue remains
• Reviewer must not use softer top-line verdicts like PROCEED or RETHINK

Rules for NEXT ACTION:

• CONTINUE means the plan is reviewable and can be revised in place
• RETURN_TO_RESEARCH means the evidence base is too weak for safe review
• RETURN_TO_PLAN means the plan is below threshold or requires scope / contract rework
• IMPLEMENT is only valid with VERDICT: APPROVED

1. Extract the blocking findings
2. If NEXT ACTION is RETURN_TO_RESEARCH:

- stop the loop - do not edit <plan> - instruct the user to run /apex:research [identifier]

1. If NEXT ACTION is RETURN_TO_PLAN:

- stop the loop - do not edit <plan> - instruct the user to run /apex:plan [identifier]

1. Otherwise continue with in-scope plan fixes
2. Update the task's <plan> section directly to resolve them
3. Keep edits scoped to the plan:

- preserve all 5 mandatory artifacts - preserve architecture decision, builder handoff, and contract validation - acknowledge contract version and any amendments - preserve or restore docs drift, pattern provenance, and NFR coverage checks - refine architecture choice - add missing mitigations - cover uncovered acceptance criteria - tighten sequencing - replace vague validation with concrete commands - clarify ownership of edge cases or concurrency handling

1. Do not edit <task-contract> or <research>
2. If a blocking issue cannot be fixed without changing contract scope, assumptions, or acceptance criteria:

- stop the loop - report VERDICT: REVISE - set next action to RETURN_TO_PLAN - instruct the user to return to /apex:plan [identifier]

1. After editing the plan, validate the mutation before the next round:

- reread the task file - confirm the <plan> block still parses and is isolated from surrounding sections - confirm the edit stayed scoped to <plan> - confirm required plan artifacts were not accidentally removed - if the task file is malformed, repair it before continuing; if that is not possible, stop and return to /apex:plan [identifier]

1. After each revision, summarize in chat:

• what the reviewer flagged
• what changed in the plan
• what remains to validate next round

1. Start the next round with a fresh reviewer subagent

If round 5 ends with VERDICT: REVISE:

• Stop the loop
• Set next action to RETURN_TO_PLAN
• Do not continue beyond 5 rounds
• Present the unresolved blockers
• Recommend returning to /apex:plan [identifier] for a larger rework

Round Status

• Round [N]: [REVISE|APPROVED]
• Repeat for each executed round, up to 5

Final Verdict

VERDICT: [REVISE|APPROVED]

Next Action

[RETURN_TO_RESEARCH|RETURN_TO_PLAN|IMPLEMENT]

Blocking Findings

• [Finding + exact fix, if any remain]

Plan Changes Made

• [High-signal summary of plan revisions applied during the loop]

Final Notes

• [If IMPLEMENT]: Ready for /apex:implement [identifier]
• [If RETURN_TO_RESEARCH]: Return to /apex:research [identifier]
• [If RETURN_TO_PLAN]: Return to /apex:plan [identifier]

• Missing mandatory plan artifacts or thin placeholder artifacts
• Missing or contradictory contract coverage
• Missing contract version / amendment acknowledgement
• Architecture that does not match research evidence
• Security-sensitive gaps with no mitigation or explicit exclusion
• Concurrency/state/order-of-operations risks left implicit
• Edge cases handled with wishful thinking
• Pattern claims that do not trace back to research or codebase evidence
• Documentation drift identified in research but absent from the plan
• Non-functional requirements without validation coverage
• Validation gates that are vague, manual-only, or not runnable
• File or dependency assumptions that do not match the repo