Give your coding agent a team of workers in Herdr. Ask one agent to coordinate a task; it can delegate investigation or implementation to Pi, Codex CLI, or Claude Code workers, collect their results, and check the work. Workers open in their own Herdr panes, while your controller stays in its original tab.
Use the core skill for delegation, or add the optional pstack bundle for workflows such as independent review and multi-step implementation. You can use one configured worker or several, including workers from different harnesses. No changes to your application code are needed.
- Herdr 0.9.3 or later, with your controller agent running in a Herdr pane (
HERDR_ENV=1). - Python 3.11 or later with POSIX file locks. Linux has been tested; macOS is not yet verified.
- Node.js and npm for the installer.
- The agent CLIs you want to use—
pi,codex, orclaude—installed and authenticated. - For Pi workers, install Herdr's integration:
herdr integration install pi.
Run one of these commands from your project. For delegation alone, install the core skill:
npx skills@1.7.1 add conscientiousness/herdr-pstack-orchestrator -s herdr-orchestratorFor pstack workflows as well, install the full bundle:
npx skills@1.7.1 add https://github.com/conscientiousness/herdr-pstack-orchestrator/tree/main/skills --skill '*'Choose your agent harnesses and project or global scope in the installer, and review any same-name skill conflicts. Then start a new agent session inside Herdr so it can load the skills. See the installation guide for manual installation, updates, and removal.
In your new Herdr session, ask your agent to set up its workers:
Set up the herdr-orchestrator skill with me. Find the available harnesses,
ask me which workers, models, and permissions to use, then check the setup.
After setup, give it a task. For example:
Use herdr-orchestrator to fix the login redirect bug in this repository.
Delegate the implementation and have a fresh worker review the result before
reporting back.
The controller creates a separate process for each task, waits for the result, and checks the work. You can inspect the worker panes as it goes.
Your agent helps you create ~/.config/herdr-orchestrator/workers.toml (or the
equivalent under $XDG_CONFIG_HOME). It records each worker's harness, model,
and permission settings; new tasks use the current definitions. One worker is
enough to start, and you can add more later.
Workers run with the permissions you choose. horch does not enforce a sandbox,
and separate worktrees only keep edits apart. See setup and configuration
for the available options and permission behavior.
The full bundle includes selected original pstack skills and a Herdr adapter. After installing it and configuring workers, try:
Use interrogate to review this branch.
Or ask for poteto-mode when you want a broader workflow. The controller can
help configure the worker roles that a workflow needs. For the included skills,
role mapping, and compatibility details, see the pstack guide.
Some original workflows require tools this adapter does not provide, including Cursor hooks or cloud execution. A workflow cannot complete a required step when its tools are unavailable. Full pstack playbook coverage has not been verified; see the playbook inventory for current results.
To update, rerun the install command you chose, select the same
harnesses and scope, and start a new Herdr session. These commands install from
this repository's main branch. The installation guide also
explains how to update a manual clone or remove the skills.
Ask your agent to run horch detect to check available harnesses and Pi
integration. If a worker stops or fails, inspect its Herdr pane before closing
it. The setup guide and task
recovery guide
cover common problems.
Built for Herdr, with an independent adapter for Lauren Tan's pstack. The banner and workflow diagram have their own asset notes and diagram notices. Licensed under the MIT License.