build

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

/build

Execute plan tasks via subagents, validate, report.

Announce: "Starting /build — executing plan with subagent dispatch."

通过子Agent执行计划任务，进行验证并生成报告。

公告: "开始执行/build——通过子Agent调度执行计划。"

Prerequisites

前置条件

Plan approved by
```
/review
```
(Step 4), OR one-sentence scope / hotfix (no plan)
Working in a linked worktree (not primary)

计划已通过
```
/review
```
（步骤4）审批，或为单句范围需求/热修复（无计划）
在关联工作树中操作（而非主工作树）

With Plan

有计划场景

1. Load

1. 加载

Read

ai-workspace/plans/<name>.md

. Extract tasks into ordered batches:

Tasks under the same step or marked parallel = one batch
Tasks with dependencies on prior tasks = separate sequential batch
Note file paths, test expectations, and acceptance criteria per task

读取

ai-workspace/plans/<name>.md

文件。将任务提取为有序批次:

同一步骤下的任务或标记为并行的任务 = 一个批次
依赖于前置任务的任务 = 单独的顺序批次
记录每个任务的文件路径、测试预期和验收标准

2. Execute Batches

2. 执行批次

Batches run sequentially. Tasks within a batch dispatch in parallel via subagents.

Agent selection per task:

Plan structure	Agent	Notes
Separate "write test" + "implement" tasks	`test-writer` (RED) then `implementer` (GREEN)	Sequential within the pair
Combined task ("implement X with tests")	`implementer`	Handles TDD internally
Docs-only task	`doc-writer`

Parallel safety: only dispatch tasks in parallel when they touch different files.

Subagent prompt — provide each agent:

You are implementing Task N: [task name]

批次按顺序运行。批次内的任务通过子Agent并行调度。

按任务选择Agent:

计划结构	Agent	说明
拆分的「编写测试」+「实现」任务	`test-writer` (RED) then `implementer` (GREEN)	任务对内部按顺序执行
合并任务（「实现X并编写测试」）	`implementer`	内部处理TDD流程
仅文档类任务	`doc-writer`

并行安全性: 仅当任务涉及不同文件时，才并行调度任务。

子Agent提示词——为每个Agent提供:

You are implementing Task N: [task name]

Task Description

[FULL TEXT from plan — paste it, don't make the subagent read the file]

Context

[Where this fits, what earlier batches produced, architectural decisions]

Working Directory

[Absolute path to worktree]

Project Conventions

TDD: write failing test first, make it pass, refactor
Commit after task completes (conventional commits)
TypeScript strict, ESM, Biome, Vitest, Pino logger, no barrel exports
Self-review before reporting back

TDD: write failing test first, make it pass, refactor
Commit after task completes (conventional commits)
TypeScript strict, ESM, Biome, Vitest, Pino logger, no barrel exports
Self-review before reporting back

Report: DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT


**Handle responses:**

| Status | Action |
|---|---|
| DONE | Mark task complete |
| DONE_WITH_CONCERNS | Read concerns — if correctness/scope, address before continuing |
| NEEDS_CONTEXT | Provide missing context, re-dispatch |
| BLOCKED | Assess: more context? more capable model? break smaller? escalate to human? |
| Task touches protected file | Stop immediately — never dispatch |

Never retry BLOCKED without changing something.

**Inter-batch check** — after each batch, verify no regressions before starting the next:

```bash
pnpm typecheck && pnpm test


**响应处理:**

| 状态 | 操作 |
|---|---|
| DONE | 标记任务完成 |
| DONE_WITH_CONCERNS | 查看问题——若涉及正确性/范围，需解决后再继续 |
| NEEDS_CONTEXT | 补充缺失的上下文，重新调度 |
| BLOCKED | 评估：是否需要更多上下文？更换更强大的模型？拆分为更小任务？升级给人工处理？ |
| 任务涉及受保护文件 | 立即停止——绝不调度 |

未做出调整前，请勿重试BLOCKED状态的任务。

**批次间检查** — 每个批次完成后，在启动下一批次前验证无回归问题：

```bash
pnpm typecheck && pnpm test

3. Validate (Hard Gate)

3. 验证（硬性关卡）

After ALL batches complete:

bash

pnpm validate

Read the output. Never claim success without evidence. If it fails:

Diagnose the failure
Fix (dispatch subagent or fix directly for small issues)
Re-run
```
pnpm validate
```
Max 3 fix-validate cycles. If still failing, stop and report.

Step 6 (

/review

) CANNOT begin until validate passes.

所有批次完成后：

bash

pnpm validate

读取输出结果。无证据绝不宣称成功。若验证失败:

诊断失败原因
修复（调度子Agent或直接修复小问题）
重新运行
```
pnpm validate
```
最多进行3次修复-验证循环。若仍失败，停止并报告。

验证通过前，不得启动步骤6（

/review

）。

4. Report

4. 报告

Build complete.
Tasks: <completed>/<total>
Validate: PASS
Next step: /review (Step 6)

Include any DONE_WITH_CONCERNS notes. Update

.branch-context.md

with decisions made during build.

Build complete.
Tasks: <completed>/<total>
Validate: PASS
Next step: /review (Step 6)

包含所有DONE_WITH_CONCERNS的问题记录。将构建过程中做出的决策更新至

.branch-context.md

文件。

Without Plan (Hotfix / One-Sentence Scope)

无计划场景（热修复/单句范围需求）

Implement directly — no subagent orchestration for small scope.

TDD: write failing test (RED), implement fix (GREEN), refactor
Minimal change: fix the defect, nothing else (hotfixes)
Root cause: document in commit message body (hotfixes)
Validate:
```
pnpm validate
```
— same hard gate, same 3-attempt max

直接实现——小范围需求无需子Agent编排。

TDD: 编写失败测试（RED阶段），实现修复（GREEN阶段），重构
最小改动: 仅修复缺陷，不做其他修改（热修复场景）
根本原因: 在提交信息正文中记录（热修复场景）
验证:
```
pnpm validate
```
— 同样的硬性关卡，最多尝试3次

Codex / Cursor Fallback

Codex / Cursor 降级方案

No subagent dispatch. Execute tasks sequentially in main session. Same validation gate applies.

不调度子Agent。在主会话中按顺序执行任务。同样适用验证关卡。

Integration

集成关系

Relation	Skill
Called by	`/task` (Step 5)
Depends on	`/plan` (Step 3) + `/review` plan approval (Step 4)
Followed by	`/review` build review (Step 6)
Replaces	`executing-plans` , `subagent-driven-development` (superpowers)

关系	技能
被调用方	`/task` (Step 5)
依赖项	`/plan` (Step 3) + `/review` 计划审批 (Step 4)
后续步骤	`/review` 构建评审 (Step 6)
替代项	`executing-plans` , `subagent-driven-development` (superpowers)