Skip to content

Latest commit

 

History

2 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

orchestrator

A Claude Code skill that drives N parallel coding tasks to completion through a strict per-task lifecycle: implement → /secure → /review → fix → verify → fix → done.

Full docs: skills/orchestrator/SKILL.md.

Install

Copy into your Claude Code config directory (~/.claude/ for global, or <project>/.claude/ for project-local):

skills/orchestrator/  → ~/.claude/skills/orchestrator/
commands/secure.md    → ~/.claude/commands/secure.md
commands/review.md    → ~/.claude/commands/review.md
agents/security-reviewer.md → ~/.claude/agents/security-reviewer.md

Or clone this repo and symlink each piece.

What's bundled vs. optional

Bundled — required for the core loop:

  • /secure — security review (dispatches to the security-reviewer agent, also bundled)
  • /review — general code review (correctness, security, performance, maintainability, test coverage)

Optional — the skill checks for these and degrades gracefully if they're missing, rather than assuming:

  • /agent-browser — used by frontend verifiers to drive an actual browser. Without it, verifiers fall back to a static diff review and note the reduced coverage explicitly.
  • /audit — a UI/accessibility/theming auditor from a separate design-skill library. It is not a general code-quality tool (that's /review, bundled) and depends on sibling skills this repo doesn't include. Only relevant if you already have that library installed.

Multi-model delegation

The orchestrator isn't limited to one model. Per Phase 2 of SKILL.md, it enumerates whatever agent types are actually available in your harness each session and assigns each worker/verifier by role — default, strong-reasoning, fast/cheap, specialized — rather than hardcoding names. Verifiers default to a different model family than the worker that built the task, when the roster has one, since a different model is more likely to catch what the builder rationalized away.

A run with several model families installed might look like this:

● main
○ sonnet       Checking server.py diff stats
○ kimi         Running full pytest suite
○ opus         Adding _note_unresolved_failure to treasury_webhook.py
○ codex (+1)   Reading WS5 harness output

Four parallel-safe tasks, four different models — chosen for fit (a hard fix went to the strong-reasoning tier, a full test run went to a fast tier, verification landed on a different family than its worker) not availability of just one vendor.

Getting more than one model family in your roster. By default your harness only exposes whatever's natively configured, which for most people is one vendor. CLIProxyAPI wraps other providers' CLI-based models (Codex, Gemini/Antigravity, Grok, Claude Code) behind an OpenAI/Gemini/Claude-compatible API, so they can be registered as additional agent types your harness dispatches to. Once configured, orchestrator picks them up automatically — Phase 2 discovers the roster live, no changes to this skill needed. Without it, orchestrator still works fine on a single-family roster; it just says so in its plan instead of silently skipping cross-family verification.

License

MIT — see LICENSE.

About

The orchestrator skill I use for long-running tasks

Resources

Stars

8 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors