orchestrate

Agent Skills › MrModest/reisenotiz › orchestrate

orchestrate

GitHub

作为编排者管理规范问题的子任务，通过轮询脚本监控并启动独立工作器会话并行实现任务，处理合并报告及清理流程。

.claude/skills/orchestrate/SKILL.md MrModest/reisenotiz

Trigger Scenarios

需要并行执行多个子任务协调多个独立开发会话

Install

npx skills add MrModest/reisenotiz --skill orchestrate -g -y

More Options

Non-standard path

npx skills add https://github.com/MrModest/reisenotiz/tree/main/.claude/skills/orchestrate -g -y

Use without installing

npx skills use MrModest/reisenotiz@orchestrate

指定 Agent (Claude Code)

npx skills add MrModest/reisenotiz --skill orchestrate -a claude-code -g -y

安装 repo 全部 skill

npx skills add MrModest/reisenotiz --all -g -y

预览 repo 内 skill

npx skills add MrModest/reisenotiz --list

SKILL.md

Frontmatter

{
    "name": "orchestrate",
    "description": "Run a spec issue's sub-issues as parallel \/implement sessions in herdr panes, until all are closed.",
    "argument-hint": "<spec issue number>",
    "disable-model-invocation": true
}

You are the orchestrator for spec issue #$ARGUMENTS. Each open sub-issue is a ticket that one worker (a Claude session in its own herdr pane and worktree) implements. You don't implement anything, and you don't read the spec's or the tickets' bodies. GitHub's sub-issue and blocked_by links are all you need.

Loop

The polling is next.sh, a plain script, so waiting costs no turns. It changes nothing except a .git/orchestrate-<spec>.seen file that records which merged PRs it has already reported. It exits only when there's something for you to do. Run it with run_in_background, passing every ticket you're leaving alone:

.claude/skills/orchestrate/next.sh <spec> [ticket...]

When it exits, handle each line of its output, then run it again straight away. Don't poll, read panes or report between runs. A run with nothing to do never exits, so your turn simply ends.

Line	Do
`ready <n>`	Run `.claude/skills/orchestrate/launch.sh <spec> <n>` and tell the user in one line. If it refuses because a pane is titled with `#<n>`, a worker was started outside this loop: add `<n>` to the leave-alone list.
`merged <n> <pr>`	Report it to the user. See Cleanup.
`done`	Stop. Report the merged PRs.

Cleanup

Only the user decides when a worker is finished. After a PR merges they may still ask the worker for a follow-up: another branch and PR in the same session, because the session already has the context.

On merged <n> <pr>, tell the user in one line that worker #<n>'s PR #<pr> merged and the worker is ready for cleanup when they say so. A follow-up PR that merges later is reported the same way.

When the user gives the go-ahead for a worker, run:

.claude/skills/orchestrate/cleanup.sh <n>

It closes the worker's pane, removes its worktree and deletes its merged branches. It refuses, changing nothing, if the session is working, the worktree has uncommitted changes, or a branch has an open PR or commits that no merged PR covers. Report a refusal to the user, and leave the worker in place.