メニュー
Use this skill when your prompt references a coordination file and assigns you a task. You are a subtask agent — part of a larger coordinated effort. Follow the protocol to read your assignment, do your work, and report progress.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use this skill from a project or task tab to map the user's open PRs onto a tree of tabs — one tab per conceptual group of authored PRs (with per-PR leaf tabs), plus a separate subtree for PRs awaiting the user's review. Good for starting the day, triaging in-flight work, or catching up after being away.
Use this skill when you have work that can be broken into parallel subtasks and delegated to child agents. Activates when you need to coordinate multiple agents working on different parts of a plan simultaneously.
Use this skill from a project tab when the user wants to be notified about GitHub PR activity in this project's repo — a PR newly requests their review, or a PR they authored gets a review decision change. Polls `gh` in the background and raises one toast per affected PR via the `notify` MCP tool; the Open button spawns a task-tab child in this project's worktree.
Use this skill when correctness is adversarial — a persistent builder agent writes code while fresh breaker agents each round probe it and report what's broken, looping until the breaker reports clean or a round cap is hit. Good fit for algorithm-heavy, parser, or security-adjacent work where "what could go wrong" is the real risk and a single implementer is likely to miss cases.
Use this skill for iterative build-refactor-build loops where a parent agent delegates work to generation tabs, harvests ideas, and spawns another generation acting on the best ones. Activates when the user wants to "keep improving X", run "another pass", do "iterative refinement", or loop until some fitness signal is met. Not for one-shot feature work — use mandelbot-delegate for that.
Use this skill when you want multiple attempts at the same task and will pick the best one. Spawns N contestant agents in parallel, then a judge agent scores their outputs against user-supplied criteria and picks a winner. Good fit when the right approach is unclear and you'd rather try several in parallel than commit to one up front.
| name | mandelbot-work-as-subtask |
| description | Use this skill when your prompt references a coordination file and assigns you a task. You are a subtask agent — part of a larger coordinated effort. Follow the protocol to read your assignment, do your work, and report progress. |
| allowed-tools | ["Read","Edit","Write","Bash","Glob","Grep","mcp__mandelbot__close_tab"] |
You have been spawned as a child agent in a coordinated multi-agent workflow. Your prompt includes an absolute path to your own *.coord.md file and a branch name. That coord file is your single source of truth and your only coordination channel with the parent.
Read <plugin-dir>/skills/_shared/coord.md for the protocol: directory layout, ownership rules, state vocabulary, log format, [DIRECTIVE] marker, block/unblock handshake, watcher usage, tab lifecycle, and sub-delegation. This SKILL file only covers the child-specific workflow; everything else lives in the shared doc.
Two rules to internalize before anything else:
*.coord.md and the parent's ../index.md (plus files explicitly referenced by path from your *.coord.md). Never a sibling's file.*.coord.md (append-only log; never edit existing entries, including the parent's [DIRECTIVE] entries) and to the branch you own.Always use watch.sh to wait for file changes. Do not write your own watcher, poll loop, or inotify script. The exact command is shown at each step that requires waiting.
You are running in your own git worktree — an isolated copy of the repository. All code changes happen inside this worktree. Do not cd to another worktree, the main repo, or any sibling's worktree. Even if a coord file or plan references a path in another worktree, you work on the code in your own worktree — those paths are for reading coordination files, not for editing code.
You may write to:
*.coord.md (coordination log)./tmp (scratch files).You own exactly one branch — you are already on it when you start. All your commits go on this branch.
Read your *.coord.md and the parent's ../index.md. The index has the "How we work" section — protocol notes, conventions, and context for this batch. Then read the governing plan referenced in the index.
Set **State:** in_progress and append - [...] starting implementation. Then do the work.
Append log entries on state changes, not on a timer. If you get stuck on something only the parent can resolve, use the block/unblock handshake from _shared/coord.md: append - [...] blocked: <question>, set **State:** blocked, then run the watcher and wait for a [DIRECTIVE] answer:
bash <plugin-dir>/skills/_shared/watch.sh <absolute path to your coord file>
When the watcher wakes, re-read your file and scan for new [DIRECTIVE] entries. If you find one, append - [...] unblocked, continuing, set **State:** in_progress, and resume. If no new directive, re-arm the watcher.
../index.md "How we work → Reviews" subsection. The default is human-in-the-loop review — only use the autonomous path if the index explicitly opts out.
awaiting_review lifecycle from _shared/coord.md: append - [...] awaiting_review: <PR link>, set **State:** awaiting_review, arm a watcher on your own coord file (same watch.sh command as in step 2, run in the background), and stay alive. Return control. Two channels are live now: review feedback comes through chat from the human; chain-wide [DIRECTIVE] entries (rebase, abort, etc.) come from the parent via the coord file — re-arm the watcher after handling each one. Stay in awaiting_review through the entire review cycle, even while addressing feedback and pushing changes. Only transition to done (and close) once the PR has merged.- [...] done, set **State:** done, and close your tab via close_tab.If you can't complete the task, append - [...] failed: <reason>, set **State:** failed, and close your tab.