Claude Code Driver — Development skill for Claude Code

claude-code-driver

**把 Claude Code 当作可派发的执行者来驾驶：来自常驻编排 agent 的无头派发 / 会话续接 / 转录监控实战知识库。**

说「让 Claude Code 帮我把这个仓库发布到 GitHub」——编排 agent 把任务写成任务书 .md，无头派发给 Claude Code，靠转录 JSONL 盯进度、靠文件 mtime 和 git 状态独立验收，全程不需要人坐在 Claude Code 的终端前。

**Drive Claude Code as a dispatchable worker: field-tested knowledge for headless dispatch, session resume, and transcript monitoring from a resident orchestrator agent.**

Say "have Claude Code publish this repo to GitHub" — the orchestrator writes the task into a brief file, dispatches it headlessly, watches progress via the transcript JSONL, and verifies the result independently against file mtimes and git status. Nobody needs to sit at Claude Code's terminal.

Why / 为什么

Claude Code 是为「人坐在终端前」设计的，但当它背后是第三方模型网关、前面是另一个常驻 agent 时，会暴露出一整套文档里找不到的坑：环境变量跨 OS 不透传、cmd.exe 引号分词吃掉多词 prompt、无头模式下 `acceptEdits` 只放行文件工具导致命令全部被拦、auto-mode 分类器在慢思考模型上把每一条 PowerShell 都堵死、`--resume` 绑定启动目录导致旧会话看不见、转录 JSONL 按启动 cwd 分片导致盯错文件。这些坑每一个都值一次白费的排障时间——本 skill 把它们全部收进一张地图，让你（或你的 agent）第一次就绕开。

Claude Code is designed for a human at the terminal — but put a third-party model gateway behind it and a resident orchestrator agent in front of it, and a whole class of undocumented failure modes appears: env vars that don't cross OS boundaries, cross-shell quoting that silently truncates multi-word prompts, headless `acceptEdits` approving file tools only while every command tool blocks, the auto-mode classifier wedging every shell action behind a slow thinking model, `--resume` bound to the launch cwd so old sessions vanish, and transcripts sharded by cwd-slug so you end up watching the wrong file. Each one costs a wasted debugging session — this skill packs them all into one map so you (or your agent) sidestep them on the first try.

内