跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
开发与测试

herdr-issue-coordinator

Coordinates a batch of GitHub issues through separate Claude Code sessions running in visible Herdr tabs, with tested pull requests, review loops, and gated merges. Use when the user asks one Claude session to sequence several coding sessions, or to complete a parent issue, milestone, or issue batch inside Herdr.

owainlewis · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add owainlewis/agent-skills --skill herdr-issue-coordinator

技能说明


name: herdr-issue-coordinator description: "Coordinates a batch of GitHub issues through separate Claude Code sessions running in visible Herdr tabs, with tested pull requests, review loops, and gated merges. Use when the user asks one Claude session to sequence several coding sessions, or to complete a parent issue, milestone, or issue batch inside Herdr." user-invocable: true argument-hint: "<parent issue, milestone, or issue list>"

Herdr issue coordination

Use this session as the coordinator. Give each dispatched GitHub issue its own git worktree, branch, Herdr tab, and Claude Code session. Sequence the work. Keep implementation context in workers and all durable state in GitHub.

This skill is experimental. It depends on the herdr CLI and its API changes between versions. Verify the commands below against the installed binary before relying on them.

Explicitly naming /herdr-issue-coordinator, or explicitly telling the coordinator that agents may merge, authorizes in-scope workers to merge their own pull requests after every merge gate below passes. An implicit skill match does not grant merge authority: run in no-merge mode unless the user's words grant it. Neither mode authorizes deployment, release publication, destructive operations, branch-protection bypasses, or changes outside the supplied batch.

Preconditions

Stop if any of these fail. Do not substitute hidden subagents for workers.

  1. test "${HERDR_ENV:-}" = 1. You must be running inside a Herdr pane.
  2. herdr --version reports 0.8.0 or later. Earlier versions have no agent prompt and no agent start --pane, and this skill will not work.
  3. herdr agent lists prompt, and agent start accepts --kind and --pane. The binary is the authority, not this document.
  4. The working directory is a git repository with a GitHub remote and gh authenticated.
  5. Read the repository instructions, parent issue, sub-issues, declared dependencies, project board, issue state, and linked open or merged pull requests. A merged pull request is only evidence of completion when it closes or explicitly references the issue and its final change and proof satisfy the issue scope. Reconcile that state instead of dispatching again.
  6. Make every issue a complete task before dispatch. If a missing product or technical decision could change behavior, data, security, compatibility, operations, cost, or proof, mark that issue as needing human input.
  7. Build a dependency graph. Stop for human correction on a dependency cycle.

State rules

The coordinator is the only writer of batch state. These rules exist because a sequenced batch runs for hours and this session's context is not durable.

  • GitHub holds issue, pull request, review, and merge state. Re-read it at the start of every cycle. Never cache it across cycles.
  • Never create a second tracker: no local state file, no notes file, no in-context list treated as authoritative.
  • Workers never write batch state. They own one issue and report back.
  • Workers never write their own tab status. The coordinator pushes it.
  • If your own context is compacted mid-batch, re-derive the whole graph from GitHub before dispatching anything.

Concurrency and tabs

  • Default to 1 active worker. Use 2 only when two issues are independent and the user asked for speed. Use another limit only when the user asks.
  • One tab per active issue. Close it after its merge is verified.
  • Never leave finished tabs open. A long batch must not accumulate tabs.
  • Keep the user's focus in the coordinator tab.

Coordinator workflow

  1. Name this tab for the batch.

    herdr tab rename "$HERDR_TAB_ID" "Coordinator"
    

    tab rename has no --clear. Record the original label and restore it verbatim when the batch ends.

  2. Dispatch only issues whose prerequisites are merged. For each one, create the worktree first, then the tab, then the agent.

    # 1. worktree, from the updated remote default branch
    git worktree add -b <branch> <worktree-path> origin/<default-branch>
    
    # 2. tab, in this workspace, rooted in the worktree
    herdr tab create --workspace "$HERDR_WORKSPACE_ID" \
      --cwd <worktree-path> --label "#<issue> <short title>" --no-focus
    # read .result.tab.tab_id and .result.root_pane.pane_id from the response
    
    # 3. agent, in that tab's root pane
    herdr agent start issue-<issue> --kind claude --pane <root-pane-id>
    
    # 4. reclaim focus; agent start has no --no-focus and takes it
    herdr tab focus "$HERDR_TAB_ID"
    

    Agent names must match [a-z][a-z0-9_-]{0,31} and be unique among live agents, so #123 is not a legal name. Use issue-123 as the agent name and #123 <short title> as the tab label. Parse every id from the JSON response. Do not predict ids.

    If the issue already has an open pull request, use that pull request's exact head branch instead of creating a new one, and pass the branch and pull request URL in the worker prompt.

  3. agent start returns only once Herdr sees the agent ready. It reports agent_status and interactive_ready. Do not add a separate wait.

  4. Send the worker its task and wait for it to settle.

    herdr agent prompt issue-<issue> "<worker prompt>" --wait --timeout <ms>
    

    agent prompt submits text and Enter atomically. --wait returns on the first settled idle, done, or blocked state. Do not use agent send followed by a separate Enter.

  5. Read a worker only when it needs help or reports completion.

    herdr agent get  issue-<issue>
    herdr agent read issue-<issue> --source recent-unwrapped --lines 200
    

    agent read returns plain text, not JSON. If a completed response is longer than the pane can return, ask the worker to write its full response to a temporary file and reply with only the path, then read that file.

  6. On blocked, inspect before answering. Send a deci

阅读完整来源 ↗