gate-review

gate-review

Exit gate for the dispatching side. Reviews whether deliverables meet the standard, and decides whether to pass or require fixes.

Core principle: Evaluate another agent's deliverables with a critical eye. Agents are not reliable; their reports are for reference only. Review from multiple angles — implementation completeness, code logic, and code style — rather than only checking for errors and unit tests.

Workflow

1. Obtain the content to review from conversation context or $ARGUMENTS
2. Locate the original task/design document (to compare against)
3. Extract the loop field from the tag ([report from ..., loop: true/false])
   • If no loop field is present in the tag, default to false
4. Review item by item
5. Output the review conclusion
6. Decide the next step based on the conclusion and the loop field:
   • loop: true + fixes needed → automatically call /dispatch to send fix instructions (redo)
   • loop: false + fixes needed → output the conclusion and stop, do not automatically initiate fixes
   • Passed → mark as complete (regardless of loop value)

Trigger Scenarios

Scenario A: Cross-Agent Review (received report)

After receiving execution results reported by the receiving side via the report skill, review the deliverables.

Extract from the report message:

• The "original task" section to recover the original design

Scenario B: Local Self-Review

After code implementation is complete and before committing, review your own code against the design document.

Obtain the design document path from $ARGUMENTS or conversation context.

Review Dimensions

Dimension	What to Check
Implementation completeness	Compare item by item against the original design; verify every task is implemented
Code logic	Logic errors, boundary conditions, bugs, security vulnerabilities
Code style	Naming conventions, code organization, interface design
Evaluate feedback	If the receiving side skipped or rejected tasks, judge whether their reasoning is valid

Review Requirements

• Compare item by item against the original design; no omissions allowed
• Verify inputs, outputs, and boundary conditions
• Check error handling, security vulnerabilities, and performance issues
• Check naming conventions, code style, and interface contracts
• Any deviation must be flagged
• If the receiving side skipped certain tasks via gate-evaluate, review whether the skip reasoning is valid

Issue Classification

• [偏离] — Content defined in the design document is inconsistent with the code implementation
• [逻辑] — Code logic errors or flaws
• [缺陷] — Code quality issues: security, performance, maintainability

Severity levels: [CRITICAL] [HIGH] [MEDIUM] [LOW]

CRITICAL means it must be fixed.

Output

After the review is complete, output the conclusion in this format:

## 审查结论

- 总体判断: 通过 | 需要修复
- 修复轮次: {当前轮次}
- 问题统计: CRITICAL: N, HIGH: N, MEDIUM: N, LOW: N

### 问题列表

1. [CRITICAL] [偏离] {问题描述} — {文件路径:行号范围}
   → 修复建议: {具体建议}
2. [HIGH] [逻辑] {问题描述} — {文件路径:行号范围}
   → 修复建议: {具体建议}
3. [MEDIUM] [缺陷] {问题描述}（可选修复）

### 跳过项审查

{如果接收端跳过了任务，逐条审查跳过理由}
- {跳过的任务}: 理由成立 / 理由不成立（需重新执行）

### 结论

{通过：标记完成 / 需修复：列出修复任务清单}

Behavior After Review

Behavior depends on the review conclusion and the loop tag:

loop: true (loop mode)

Conclusion	Behavior
Passed	Mark as complete, document status → done
Fixes needed	Automatically call /dispatch to send the issue list and fix suggestions back to the receiving side (redo)
Receiving side's rejection is justified	Accept the rejection, adjust the task or mark as complete
Receiving side's rejection is unjustified	Automatically call /dispatch to resend, with rebuttal explanation

loop: false (non-loop mode)

Conclusion	Behavior
Passed	Mark as complete, document status → done
Fixes needed	Output the issue list and fix suggestions, do not automatically initiate fixes; let the user decide the next step
Receiving side's rejection is justified	Accept the rejection, inform the user
Receiving side's rejection is unjustified	Inform the user, let the user decide whether to redispatch

Termination Conditions

Stop the loop when any of the following conditions is met:

• Review passes completely (no CRITICAL/HIGH issues)
• Only MEDIUM/LOW issues remain
• Cumulative fix rounds reach 3
• User manually interrupts