orchestrator
Skill2 starsChanged 4 months agoMITView on GitHub
Load when the user wants to execute a planned epic — when they say 'orchestrate', 'execute this epic', 'run the tasks', 'kick off the epic', or 'run the plan'. Also load when the frontier of a beads epic is ready and the user wants autonomous execution. NOT for planning (see planner), NOT for brainstorming (see brainstorm), NOT for code review of a single PR. Requires the `workflow` tool (pi-dynamic-workflows) — stop if it is unavailable.
80 other projects use a version of this.
Adopted by
Most recently taken up by
- BrightGold70/skillstheir own version, first seen Nov 8, 2026
- Veeiy/agent-skillstheir own version, first seen Nov 8, 2026
- Mehul72/skillstheir own version, first seen Nov 8, 2026
- pietz/skillstheir own version, first seen Nov 8, 2026
- kcns008/cluster-skillstheir own version, first seen Nov 8, 2026
- soilmass/motion-design-agenttheir own version, first seen Nov 8, 2026
Used with
Projects with it also use worker, reviewer and 7 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
stotihvSkillworker
Execute beads autonomously within a track. Handles bead-to-bead context persistence via Agent Mail, uses preferred tools from AGENTS.md, and reports progress to orchestrator.
Also in 3 projects that use this
MicrosoftSkillreviewer
ASP.NET Core ハンドブックの記事をレビューする。Use when: 記事をレビューする、ドキュメントを校正する、コードレビューする、PRをレビューする。技術ドキュメントのレビュー、品質チェック、校正。
Also in 4 projects that use this
yakatotechSkilladvisor
Provide improvement suggestions and advice on work content. Use for document review, efficiency suggestions, and best practices
Also in 3 projects that use this
AWSSkillplanning
Discovers user intent and generates a structured, step-by-step plan for model customization workflows. This skill must always be activated alongside any other skill when the user's request relates to model customization — including fine-tuning, training, building, customizing, reviewing data, or getting advice on approach, regardless of domain. Do not skip this skill even if the immediate ask is narrow (e.g., reviewing data format or a single workflow step), because planning discovers the full scope of work needed. Also activate when the user wants to resume, continue, or modify an existing plan.
Also in 3 projects that use this
OpenAISkillobservability
Vercel Observability expert guidance — Drains (logs, traces, speed insights, web analytics), Web Analytics, Speed Insights, runtime logs, custom events, OpenTelemetry integration, and monitoring dashboards. Use when instrumenting, debugging, or optimizing application performance and user experience on Vercel.
Also in 3 projects that use this
scott2bSkillobsidian
PROACTIVELY use for Obsidian vault operations. Fast search using ripgrep, direct file operations (read/write/delete). Works with slipbox in _Slipbox directory. 4-7x faster than API. Use obsidian-api skill ONLY for Dataview queries, periodic notes, or patch operations.
Also in 3 projects that use this
Other versions80 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 80
- nubjs/nub★ 4.4ksame name
- SirAllap/agentglass★ 347same name
- DylXBT/claude-orchestrator-skills★ 114same name
- nwiizo/claudelytics★ 88same name
- nahisaho/MUSUBI★ 74same name
- olivieropinotti/orchestrator-skill★ 70same name
- sanky369/vibe-building-skills★ 30same name
- felvieira/claude-skills-fv★ 27same name
- robertbstillwell/marketing-skills★ 26same name
- kcns008/cluster-agent-swarm-skills★ 25same name
70 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository67 other files it gives its agents.
- AGENTS.mdAGENTS.md
- agent-architectureSkill
- agent-evaluationSkill
- agent-memorySkill
- agent-post-trainingSkill
- agent-self-evolutionSkill
- agent-tool-designSkill
- api-designSkill
- application-architectureSkill
- brainstormSkill
- code-reviewerSkill
- coding-agent-designSkill
- concurrency-designSkill
- context-engineeringSkill
- dockerSkill
- game-audioSkill
- game-designSkill
- game-narrativeSkill
- game-patternsSkill
- game-performanceSkill
- game-renderingSkill
- game-visualsSkill
- git-commitSkill
- git-prSkill
- godot-shaderSkill
- godotSkill
- gpu-rendering-architectureSkill
- grafanaSkill
- helmSkill
- homelab-monitoringSkill
- k3sSkill
- k8s-networkingSkill
- k8s-operationsSkill
- k8s-storageSkill
- k8s-workloadsSkill
- knowledge-baseSkill
- level-designSkill
- llm-inference-servingSkill
- loggingSkill
- logqlSkill
- love2d-fennelSkill
- love2dSkill
- ml-platform-architectureSkill
- ml-post-trainingSkill
- multi-agent-collaborationSkill
- nix-dendriticSkill
- nix-packagingSkill
- nixSkill
- odin-designSkill
- odin-gamedevSkill
- physical-ai-systemsSkill
- plannerSkill
- prd-to-storiesSkill
- prd-writingSkill
- promqlSkill
- pr-reviewerSkill
- python-designSkill
- rag-designSkill
- ray-data-pipelinesSkill
- ray-kubernetes-infrastructureSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
6 rules. 0 are common, 0 are distinctive, and 6 are very specific to this project.
0
Common
Most of the 66 edited versions we read keep these.
0
Distinctive
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 16
Never edit, write, or create files yourself — every change comes from a coder in its own worktree.
- Very specific: line 46
If the epic notes contain orchestrator base=, you are resuming: read BASE from them, run Resume below, then continue with the waves (or …
- Very specific: line 63
Commit message: write one conventional-commit subject per ticket from its title and intent (git-commit skill rules: type(scope): …
- Very specific: line 78
Integration failure: if the ticket's notes already contain integration retry, it is stuck (step 9).
- Very specific: line 86
Clean up: every task of the wave as soon as it is decided (landed, stuck, or sent back for retry): git -C <repo> worktree remove --force …
- Very specific: line 96
Live waves: if any task worktrees exist under <repo>/.pi/worktrees/, ask the user once whether another session is still running a wave for …
“Common” is measured: kept by at least 6 in 10 of the 66 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Orchestrator
Drive a beads epic from its first ready ticket to an audited set of commits on the user's working branch. Coders and
reviewers run as subagents inside a workflow script; you own beads, git integration, verification, and governance.
Each ticket that passes review lands on the working branch immediately, with its review summary in the commit body,
so the user reviews the whole run with git diff origin/<branch>.
<HARD-GATE>
Never edit, write, or create files yourself — every change comes from a coder in its own worktree. Bash is for
`bd`, the git integration commands below, and verification commands from tickets. Never push. `git commit` and
`git reset` go through the user's permission prompts; do not work around them.
</HARD-GATE>Layout
| Thing | Where |
| Working branch | the user's current branch in <repo> (git rev-parse --show-toplevel); passed tickets land here |
| Task worktree | <repo>/.pi/worktrees/<task-id> on branch pi/wf/<task-id>, cut from the working branch HEAD |
| Wave script | saved workflow orchestrator_wave (source: references/wave.js) |
| Audit script | saved workflow orchestrator_audit (source: references/audit.js) |
| Agent profiles | coder, correctness-reviewer, failure-path-reviewer, readability-reviewer, security-reviewer, ticket-auditor, epic-auditor |
Pass absolute paths (<repo>/.pi/worktrees/..., <repo>) as worktreePath / integrationPath; agent cwd must be
absolute. Run both scripts with workflow({ name: "orchestrator_wave" | "orchestrator_audit", background: false, args })
so the result returns in the same turn. Call workflow directly as a tool, never from codemode: nested calls hide the
workflow progress panel (guardrails blocks them). If the saved workflow is missing, pass the reference file's full text
as script instead. Always write beads notes with --append-notes (--notes refuses to overwrite existing notes).
Integration commits are unsigned (-c commit.gpgsign=false) so an unattended run never waits on a signing prompt; the
final summary gives the user the one-line re-sign command.
1. Start
bd show <epic-id>: the description must link a plan or design note. If none, warn the user once and continue — the epic audit then judges against the description and tickets only.- Clean tree:
git -C <repo> status --porcelain --untracked-files=nomust be empty. Otherwise stop and ask the user to commit or stash: cherry-picks into a dirty checkout fail. Tell them not to edit the checkout during the run. - If the epic notes contain
orchestrator base=, you are resuming: readBASEfrom them, run Resume below, then continue with the waves (or straight to Audit if nothing is runnable). - Otherwise
bd ready --parent <epic-id>. Empty frontier: report the state (no children, all blocked, all closed) and stop. - Record
BASE=$(git -C <repo> rev-parse HEAD)andbd update <epic-id> --append-notes "orchestrator base=<BASE>". - Submodule paths, once:
git config -f <repo>/.gitmodules --get-regexp '\.path$'(missing file = none).
2. Each wave
- Candidates:
bd ready --parent <epic-id>plus any tickets Resume returned to the queue, minus the set-aside list. Set aside[human-task],[research],[brainstorming],[prototype]tickets. - Submodule filter: tickets naming paths inside a submodule (from Start step 6) are set aside, once:
bd update <id> --append-notes "edits submodule <path>; run in the main checkout". Set-aside tickets stay in the list for the summary and are never re-noted. No runnable candidates left → go to Audit. - File conflicts: two candidates naming the same file go in different waves (keep the first). Tickets back for an integration retry run at most one per wave: they already collided once, and siblings retried together collide again.
- Claim each candidate (
bd update <id> --claim); skip any that fail. - Commit message: write one conventional-commit subject per ticket from its title and intent (git-commit skill rules:
type(scope): imperative, lowercase, < 72 chars, no ticket IDs).
wave.jsadds the review summary as the body. - Worktrees: for each ticket,
cd <repo> && git worktree add -B pi/wf/<task-id> .pi/worktrees/<task-id> HEAD(-Bresets a branch left by an earlier attempt); note the working-branchHEADSHA as that task'sbase. - Run the wave:
bd update <id> --append-notes "wave started: <commitMessage>"for each ticket, thenworkflow({ name: "orchestrator_wave", background: false, args: { tasks: [{ id, title, description, acceptance, commitMessage, verify, worktreePath, branch, base }] } }).verifyis the command from the acceptance criteria (tests/build), or empty. For integration retries, append the previous failure output todescription. Anullentry inresults(that task's agent call failed outright) maps back to its task by index: treat it as stuck. - Land each
passresult on the working branch, one at a time:git -C <repo> log --oneline <base>..pi/wf/<task-id>must show exactly one commit; otherwise it is an integration failure ("expected one commit, found N").cd <repo> && git -c commit.gpgsign=false cherry-pick pi/wf/<task-id>, then the ticket's verification command in<repo>.- Conflict:
git cherry-pick --abort. Verification fails:git reset --keep HEAD~1(prompts the user). - Integration failure: if the ticket's notes already contain
integration retry, it is stuck (step 9). Otherwisebd update <id> --status open --append-notes "integration retry 1: <reason>"— it comes back in the next wave from the newHEAD, with the reason added to its description. - Success:
bd close <id> --reason "<one line: what landed, reviews passed>"andbd update <epic-id> --append-notes "landed <id> <short sha>"(the audit covers only landed tickets).
- Stuck (
verdict: "stuck"or a second integration failure):bd update <id> --append-notes "stuck: <blocking findings or failure summary>". Do not close; leave it in progress and add it to the set-aside list. Only its dependents wait on it. - Clean up every task of the wave as soon as it is decided (landed, stuck, or sent back for retry):
git -C <repo> worktree remove --force <repo>/.pi/worktrees/<task-id>(untracked build output is expected), thengit -C <repo> branch -D pi/wf/<task-id>— one branch per command (multi-branch deletes prompt). A stuck ticket's findings live in its notes; its branch is not kept. - Loop. No pauses or check-ins between waves.
Stop early only for: infrastructure failure (workflow tool missing, bd/git errors), all remaining tickets stuck or
blocked, or acceptance criteria no reviewer can judge — flag that ticket and continue the rest.
Resume (epic notes contain orchestrator base=):
- Live waves: if any task worktrees exist under
<repo>/.pi/worktrees/, ask the user once whether another session is still running a wave for this epic. Yes → stop. No → remove them and theirpi/wf/*branches (a crash left them). - Set-aside list: rebuild it from notes — tickets whose notes contain
stuck:oredits submodulestay set aside and are not noted again. - Landed: an in-progress ticket whose
wave started: <subject>note is ingit -C <repo> log --format=%s <BASE>..HEADalready landed →bd closeit and notelanded <id>on the epic. The ticket audit re-runs its verification. - Everything else in progress returns to the candidate queue. Coder threads are not journaled, so
resumeFromRunIdgains nothing here.
3. Audit (every run)
workflow({ name: "orchestrator_audit", background: false, args: { integrationPath: <repo>, base: BASE, tickets: [landed
tickets: { id, title, acceptance, verify }], epic: { id, description, planPath } } }) — tickets = those noted
landed <id> on the epic.
- A ticket
not-met: reopen it (bd update <id> --status open --append-notes "audit: <evidence>"). Report it; do not re-run. - A ticket
unknownorepic: null: the auditor produced no verdict — report it as unaudited; never reopen for it. - Epic
missing/outOfScope/contradicted: report; never auto-fix.
4. Summary
The work is already on the working branch. End with:
Epic <id> — <N> waves, <M> commits on <branch> since <BASE>
| Commit | Ticket | Title | Reviews | Verification | Audit |
|---|---|---|---|---|---|
| <sha> | <id> | <title> | pass (retries: n; missing: reviewer?) | pass | met / not-met / unknown |
Stuck: <id — last blocking findings>
Set aside: <id — submodule path / human task>
Epic audit: missing [...], out of scope [...], contradicted [...]
Review: git diff origin/<branch> (or git log --stat <BASE>..HEAD; each commit body carries its review summary)
Re-sign: git rebase --exec 'git commit --amend --no-edit -S' <BASE>Confirm git -C <repo> worktree list shows no .pi/worktrees/ entries and git -C <repo> branch --list 'pi/wf/*' is empty.
Rules
- Never run the
pibinary and never tell a subagent to. - Children: subagents load skill-router, guardrails, audit and rtk, but no MCP, codemode, or web tools. Put everything a coder needs in the ticket text.
- Findings severity:
criticalandimportantblock and trigger a retry (max 2, insidewave.js);suggestionnever blocks —wave.jslists them in the commit body. - Commit subjects describe the change, never the process ("address review findings"). The body is the review summary.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

