herdr-orchestration

Orchestrate Claude and Codex worker TUIs from a controller agent over the herdr socket CLI.

2.7k|171|Updated Mar 28, 2026
One-click install
npx skills add https://github.com/compozy/compozy --skill herdr-orchestration
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: herdr-orchestration
Source: https://github.com/compozy/compozy/tree/main/.agents/skills/herdr-orchestration
Command: npx skills add https://github.com/compozy/compozy --skill herdr-orchestration

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Coordinating multiple AI coding agents in parallel is error-prone: workers stall on permission prompts, headless runners never report status, and unverified patches get merged. This Skill gives a controller agent a disciplined protocol to launch, monitor, verify, and retire Claude and Codex worker TUIs through herdr's socket CLI.

Core Features & Use Cases

  • Named-tab worker management: Each worker TUI gets its own labeled tab and unique agent name, launched via herdr agent start with unattended flags like --dangerously-skip-permissions and --yolo.
  • Native status waits: Block on agent wait --until done/blocked in 3-5 minute check-in intervals instead of screen-scraping, with agent explain for debugging detection.
  • Plan-first delegation: Optional plan-mode flow for Claude (--permission-mode plan) and Codex (shift+tab), with exact acceptance-menu mechanics documented in references/plan-mode.md.
  • Verification and retirement: Worker output is treated as untrusted until the controller re-runs commands and reviews diffs, then the worker's tab is closed and disposition recorded.
  • Use Case: A controller agent splits a bug fix into two slices, launches an Opus worker and a GPT-5.5 worker in isolated worktrees, waits on their native status, verifies both diffs, and retires the tabs.

Quick Start

Ask the agent to use the herdr-orchestration skill to split this task into slices and delegate them to Claude and Codex worker TUIs in named herdr tabs.

Frequently Asked Questions about herdr-orchestration

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I orchestrate multiple Claude and Codex agents in parallel?

Launch each worker as an interactive TUI in its own named herdr tab using herdr agent start, then delegate task slices as self-contained packets. The controller tracks workers via agent wait status checks and verifies their output before retiring each tab.

How do I run Claude Code and Codex workers unattended?

Launch Claude with --dangerously-skip-permissions and Codex with --yolo so workers never stall on permission prompts. Pass the delegation packet as launch argv so the session starts working immediately.

Why does my herdr worker never report done or blocked status?

Status detection only works for real interactive TUIs launched through herdr agent start. Headless runners like claude -p or codex exec stream JSON events and never report state, so relaunch the worker through agent start and check detection with agent explain.

How do I use plan mode with Claude Code or Codex workers?

Plan-first is opt-in and requires explicit user activation. Claude launches with --permission-mode plan, while Codex enters Plan mode via shift+tab after a bare launch; the exact acceptance menus and key sequences are documented in references/plan-mode.md.

Can herdr workers edit overlapping files safely?

Only when the controller integrates and resolves conflicts immediately. For parallel edits on overlapping files, create an isolated worktree per worker with herdr worktree create so each worker operates in its own checkout.