implement-issues

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Implement Issues

实现Issues

Turn a set of GitHub issues into open PRs via parallel, worktree-isolated lanes. One lane per issue: its own worktree, its own branch, its own PR. This skill orchestrates and composes

branch

commit

, and

pr

— it never merges. Hand-off to

babysit-pr

babysit-pr --merge

for shepherding and merge.

通过并行的、工作树隔离的任务分支，将一组GitHub Issues转换为已打开的PR。每个Issues对应一个任务分支：独立的工作树、独立的分支、独立的PR。该技能编排并组合

branch

、

commit

和

pr

能力——从不执行合并。如需后续跟进和合并，请转交

babysit-pr

babysit-pr --merge

处理。

Invocation

调用方式

```
/implement-issues #12 #14 #17
```
— implement exactly these issues.
```
/implement-issues
```
(bare) — triage the backlog (
```
gh issue list
```
), prefer self-contained, file-disjoint issues, propose a set and confirm before spawning any lane.
```
--sequential
```
— run one lane at a time (still one worktree/branch/PR each).

```
/implement-issues #12 #14 #17
```
—— 精准实现指定编号的Issues。
```
/implement-issues
```
（无参数）—— 分类处理积压任务（执行
```
gh issue list
```
），优先选择独立无依赖、文件无重叠的Issues，在启动任何任务分支前，先提议一组任务并确认。
```
--sequential
```
—— 依次运行每个任务分支（每个分支仍拥有独立的工作树/分支/PR）。

Inputs & readiness

输入与准备工作

Load repo context first. At start, from the target repo checkout root, read
```
AGENTS.md
```
and/or
```
CLAUDE.md
```
if present. Precedence:
```
AGENTS.md
```
is authoritative when both exist — do not also apply conflicting
```
CLAUDE.md
```
instructions; if only one exists, use that file; if both conflict in a blocking way, stop and ask. Record which filename is authoritative (e.g.
```
AGENTS.md
```
or
```
CLAUDE.md
```
) and treat its house standards, operating agreement, and standing constraints as binding for every lane. Pass that chosen filename into each lane prompt so lanes do not re-pick. See the
```
stand-general
```
skill's Per-repo agent context section for the expected file shape. Missing the file is fine — fall back to chat instructions and
```
stand-*
```
skills.
Take issues from the args, or select from
```
gh issue list
```
favouring self-contained and file-disjoint work.
Every issue must carry an AI Implementation Prompt comment (per the
```
issue
```
skill). Missing it means the issue is not fan-out-ready — write one first, or run that issue as a single lane yourself. Do not hand a bare issue to a sub-agent.
Reproduce against current
main
before implementing — specs can be stale. If the reported problem no longer exists, report back rather than fabricating a change.

先加载仓库上下文。启动时，从目标仓库的 checkout 根目录读取
```
AGENTS.md
```
和/或
```
CLAUDE.md
```
（如果存在）。优先级规则：当两个文件同时存在时，
```
AGENTS.md
```
具有权威性——请勿应用
```
CLAUDE.md
```
中与之冲突的指令；若仅存在其中一个文件，则使用该文件；若两个文件存在阻塞性冲突，请停止操作并询问用户。记录哪个文件具有权威性（例如
```
AGENTS.md
```
或
```
CLAUDE.md
```
），并将其包含的编码规范、操作协议和既定约束视为所有任务分支必须遵守的规则。将选定的文件名传入每个任务分支的提示语中，避免分支重复选择文件。关于文件的预期格式，请参考
```
stand-general
```
技能中的每个仓库的代理上下文部分。若文件不存在也无需担心——可回退至聊天指令和
```
stand-*
```
系列技能。
从参数中获取Issues，或从
```
gh issue list
```
中选择优先处理独立无依赖、文件无重叠的任务。
每个Issue 必须包含AI实现提示评论（遵循
```
issue
```
技能规范）。若缺少该评论，则表示该Issue 不具备拆分执行的条件——请先编写该评论，或自行单独处理该Issue。请勿将未准备好的Issue交给子代理（sub-agent）处理。
在实现前基于当前
main
分支复现问题——需求规格可能已过时。若报告的问题已不存在，请反馈给用户，而非强行修改代码。

Conflict grouping

冲突分组

Build a file-overlap matrix from each issue's declared file targets.
Disjoint file sets → run in parallel lanes.
Overlapping file sets → sequence them, or bundle into a single lane only when the issues genuinely must share files.
Cap at ~10 lanes — beyond that, batch across sessions.

根据每个Issue声明的目标文件构建文件重叠矩阵。
无重叠的文件集 → 以并行方式运行任务分支。
有重叠的文件集 → 按顺序执行，或仅当Issues确实必须共享文件时，才将其合并到同一个任务分支中。
任务分支数量上限为约10个——超出该数量时，分批次处理。

Lane claiming

任务分支占用

Before spawning, claim each lane — assign the issue to yourself and/or add a working label. This prevents a second session from opening a duplicate PR for the same issue (a race that has bitten this workflow before).

在启动任务分支前，占用每个分支——将Issue分配给自己和/或添加"处理中"标签。这可避免另一个会话为同一个Issue打开重复的PR（此前该工作流曾遇到过此类竞态问题）。

Per lane

每个任务分支的处理流程

Orchestrator creates the worktree per the
```
branch -w
```
convention: a sibling directory, branch
```
type/NN-slug
```
cut from a freshly fetched
origin/main
. Never place a worktree under
```
.claude/
```
— lint configs commonly exclude that path and will silently scan nothing.
Sub-agent prompt is pointers, not content. Instruct the lane to:
- Work only inside the lane worktree path the orchestrator created.
- Read issue #N and its AI Implementation Prompt comment.
- From the lane worktree root (absolute path), read the authoritative context file the parent already chose (
```
AGENTS.md
```
  or
```
CLAUDE.md
```
  ) when present. Do not resolve context from the parent checkout or any other directory; do not re-pick between the two files. Treat its standards/constraints/contract as binding; follow applicable
```
stand-*
```
  skills.
- Run the full lint and full test suite before every commit — never a subset.
- Use the
```
commit
```
  skill, then the
```
pr
```
  skill with
```
Closes #N
```
  .
- Never merge. Report the PR URL on completion.
Orchestrator verifies each PR exists once the lane reports done.

编排器按照
```
branch -w
```
约定创建工作树：在同级目录中创建分支
```
type/NN-slug
```
，分支基于刚拉取的
origin/main
。切勿将工作树放置在
```
.claude/
```
目录下——代码检查配置通常会排除该路径，导致扫描无结果。
子代理提示语为指向性内容，而非具体细节。向任务分支下达以下指令：
- 仅在编排器创建的任务分支工作树路径内工作。
- 查看Issue #N及其AI实现提示评论。
- 从任务分支工作树的根目录（绝对路径）读取父级已选定的权威上下文文件（
```
AGENTS.md
```
  或
```
CLAUDE.md
```
  ，如果存在）。请勿从父级checkout目录或其他目录解析上下文；请勿重新选择两个文件中的一个。将该文件中的规范/约束/协议视为必须遵守的规则；遵循适用的
```
stand-*
```
  技能。
- 在每次提交前运行完整的代码检查和完整的测试套件——绝不执行子集检查。
- 使用
```
commit
```
  技能，然后使用
```
pr
```
  技能并添加
```
Closes #N
```
  标识。
- 绝不合并。完成后报告PR的URL。
编排器在任务分支报告完成后，验证每个PR是否存在。

Guardrails

防护规则

Signing pre-flight before spawning. Confirm commit signing works now — a locked signing key must be surfaced to the user before you go AFK, not discovered mid-run.
A lane that cannot stay green STOPS and reports. Never weaken or delete tests, never open a knowingly broken PR.
Final report lists every PR URL and every blocked lane with the reason.

启动前确认签名功能正常。先确认提交签名功能当前可正常工作——若签名密钥锁定，必须在开始执行前告知用户，而非在执行过程中才发现问题。
若任务分支无法保持测试通过状态，则停止并报告。绝不能弱化或删除测试，绝不能打开明知有问题的PR。
最终报告需列出所有PR的URL以及所有阻塞的任务分支及其原因。

Hand-off

转交后续处理

Shepherding and merging are not this skill. Once PRs are open, hand off to

babysit-pr

(drive to merge-ready) or

babysit-pr --merge

Harness mechanics (Claude Code)

底层机制（Claude Code）

Run lanes as background sub-agents so the main thread stays free.
Model per lane by risk:
- docs / config / mechanical →
```
sonnet
```
- logic / refactor / security / auth / CI-touching →
```
opus
```
- never
  haiku
  for code.
- An explicit user instruction overrides these defaults.
Elsewhere (harnesses without sub-agent spawning): run the lanes sequentially using the same worktree-per-lane layout.

将任务分支作为后台子代理运行，使主线程保持空闲。
根据风险为每个任务分支选择模型：
- 文档/配置/机械性修改 → 使用
```
sonnet
```
  模型
- 逻辑重构/安全/认证/涉及CI的修改 → 使用
```
opus
```
  模型
- 绝不为代码任务使用
  haiku
  模型。
- 用户的明确指令可覆盖上述默认规则。
其他场景（不支持子代理启动的底层框架）：使用相同的"每个任务分支对应一个工作树"布局，按顺序运行任务分支。