plan-review

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Chinese

Phase gate: PLAN checkpoint. MUST get APPROVE from all reviewers before proceeding to Build.

阶段门禁：PLAN检查点。在进入构建阶段前必须获得所有评审人员的批准。

Plan file path (default: most recent non-
```
.done.md
```
file in
```
ai-workspace/plans/
```
)
```
--include <agent>
```
: Force-add an agent to the panel
```
--exclude <agent>
```
: Remove an agent (except technical-editor, which cannot be excluded)

Find the plan file:
bash
```
ls -t ai-workspace/plans/*.md | grep -v '.done.md' | head -1
```
If no plan file found → error: "No plan file in ai-workspace/plans/. Write a plan first." If plan file is
```
.done.md
```
→ error: "This plan is already finalized."
Read the plan and extract:
- Files to Modify section
- Full body text
- Plan scope (count files changed, count body lines)

Auto-assemble the review panel using these deterministic rules:

Tier 1 — Always included:

Agent	When
`technical-editor`	EVERY plan review (cannot be excluded)

Tier 2 — Auto-included by scope:

Agent	Trigger
`code-reviewer`	Plan includes ANY code changes (files ending in `.ts` , `.js` , `.sh` , `.yml` , or paths containing `src/` , `tests/` )
`architect-reviewer`	Plan is medium/large: >3 files changed, OR plan body >100 lines, OR multiple workflow steps affected

Tier 3 — Auto-included by keyword scanning:

Agent	Trigger keywords (scan Files to Modify + body)
`codex-specialist`	AGENTS.md, config.toml, sync.sh, skills, dotfiles
`designer` + `design-reviewer`	component, UI, CSS, design, wireframe, mockup, Figma, layout
`security-auditor`	auth, credentials, permissions, OWASP, token, secret, vulnerability
`accessibility-tester`	a11y, WCAG, aria, keyboard, screen reader, focus, semantic HTML
`fact-checker`	docs, research, ecosystem, reference, guide, educational

Log which agents were selected and why.

Apply overrides:
```
--include
```
adds agents,
```
--exclude
```
removes (except technical-editor).
Dispatch all selected agents IN PARALLEL using the Agent tool (single message, multiple Agent tool calls). Each agent receives:
- The plan file content
- Their specific review mandate
- Instructions to return: APPROVE / REVISE / DROP with findings tagged P0/P1/P2

Collect verdicts and synthesize into a summary table:

## Plan Review Summary
| Agent | Verdict | P0 | P1 | P2 | Key Finding |
|-------|---------|----|----|-----|-------------|

Gate logic:
- All APPROVE → "Proceed to Build"
- Any P0 REVISE → MUST fix before proceeding. List required fixes.
- P1 REVISE → should fix. List recommended fixes.
- P2 → nice to have, can proceed.
- If P0 exists, enter revise-resubmit loop (up to 2 rounds). After 2 rounds, escalate to human.

查找计划文件：
bash
```
ls -t ai-workspace/plans/*.md | grep -v '.done.md' | head -1
```
如果未找到计划文件 → 错误："ai-workspace/plans/目录下不存在计划文件，请先编写计划。" 如果计划文件是
```
.done.md
```
格式 → 错误："该计划已完成定稿。"
读取计划并提取以下内容：
- 待修改文件章节
- 完整正文内容
- 计划范围（统计变更文件数量、正文行数）

按照以下确定性规则自动组建评审小组：

第一层 — 固定纳入：

Agent	触发条件
`technical-editor`	所有计划评审都必须纳入（不可被排除）

第二层 — 根据计划范围自动纳入：

Agent	触发条件
`code-reviewer`	计划包含任意代码变更（文件后缀为 `.ts` 、 `.js` 、 `.sh` 、 `.yml` ，或路径包含 `src/` 、 `tests/` ）
`architect-reviewer`	计划为中/大型：变更文件超过3个，或计划正文超过100行，或影响多个工作流步骤

第三层 — 通过关键词扫描自动纳入（扫描待修改文件列表+正文）：

Agent	触发关键词
`codex-specialist`	AGENTS.md、config.toml、sync.sh、skills、dotfiles
`designer` + `design-reviewer`	component、UI、CSS、design、wireframe、mockup、Figma、layout
`security-auditor`	auth、credentials、permissions、OWASP、token、secret、vulnerability
`accessibility-tester`	a11y、WCAG、aria、keyboard、screen reader、focus、semantic HTML
`fact-checker`	docs、research、ecosystem、reference、guide、educational

记录入选的Agent及其入选原因。

应用覆盖规则：
```
--include
```
参数可添加Agent，
```
--exclude
```
参数可移除Agent（technical-editor除外）。
使用Agent工具并行调度所有入选的Agent（单条消息，多Agent工具调用）。每个Agent会收到：
- 计划文件内容
- 其专属的评审要求
- 返回要求：需返回APPROVE / REVISE / DROP结论，且发现的问题需标注P0/P1/P2等级

收集评审结论，整合为汇总表：

## Plan Review Summary
| Agent | Verdict | P0 | P1 | P2 | Key Finding |
|-------|---------|----|----|-----|-------------|

门禁逻辑：
- 全部通过 → "进入构建阶段"
- 存在任意P0等级需修改问题 → 继续推进前必须修复，列出必填修复项
- 存在P1等级需修改问题 → 建议修复，列出推荐修复项
- P2等级问题 → 优化项，可继续推进
- 如果存在P0问题，进入修改-重提交流程（最多2轮），2轮后仍未通过则上报人工处理