pi-herdr-live-agents

Visible Pi (sub)agents that run as real, live Pi sessions in Herdr panes.

A parent session spawns each agent and receives its result, and that is where the hierarchy ends. Everything else is an ordinary Pi TUI in its own pane: you can watch it work, read the whole conversation, scroll its history, type into it, and keep the pane open long after the delegated task finished.

The parent still orchestrates through tools, so the model delegates exactly as it would with a headless extension.

┌─────────────────────────────────┬─────────────────────────────────┐
│ parent Pi session               │ agent · review-auth             │
│                                 │                                 │
│ > spawn_agent(review-auth)      │ > Review src/auth.ts and report │
│   pane w2:p4 · claude-fable-5   │   every token-expiry bug.       │
│ > wait_agent(review-auth)       │   read src/auth.ts              │
│   waiting...                    │   ...                           │
│                                 │                                 │
│ agents  1 running               │ ~/app > claude-fable-5 > high   │
└─────────────────────────────────┴─────────────────────────────────┘

Requirements

• Pi 0.84.2 or newer
• Herdr 0.8.0 or newer
• Pi started inside a Herdr pane, so HERDR_ENV=1 is set

This package calls the Herdr CLI directly. @ogulcancelik/pi-herdr is optional and only makes blocked-agent detection instant instead of taking up to 5 seconds.

Install

pi install npm:pi-herdr-live-agents

Try it without installing permanently:

pi -e npm:pi-herdr-live-agents

How a delegation runs

1. The parent model calls spawn_agent with a self-contained task message it wrote itself.
2. The extension opens a sibling pane, or an agents · <session> tab when a split would leave any pane smaller than 72x20.
3. Herdr starts a normal Pi session there with the chosen provider, model, and thinking level.
4. A private mailbox hands the task to the child, which delivers it through pi.sendUserMessage().
5. When the child finishes, its reporter writes the final response to disk and the parent receives it as a structured result.
6. The pane stays open until you or the model closes it.

Nothing is scraped from the terminal, and no hidden prompt is injected into the child. The child keeps your AGENTS.md, skills, extensions, and working directory, and starts with an empty conversation.

Tools

| Tool | Purpose | | --- | --- | | spawn_agent | Start an agent in a new pane with a task message | | send_message | Send another turn to a running agent | | wait_agent | Wait for the next agent to finish | | wait_all_agents | Wait for several agents to finish | | list_agents | List agents, their status, and their panes | | read_agent_response | Page through a response larger than the delivered slice | | interrupt_agent | Stop the current turn and keep the session usable | | focus_agent | Move the Herdr focus to an agent pane | | close_agent | Close an idle or finished agent pane |

Only the root session receives these tools. Child sessions load the reporter and /return-to-parent, so agents cannot spawn further agents.

Results are delivered automatically up to 64 KiB, with a result_id in every payload. Longer responses stay complete on disk, and read_agent_response returns the rest from any byte offset without splitting a UTF-8 character.

Commands

/agents                         list agents and focus the selected pane
/subagents                      alias for /agents
/subagent <name>                focus one pane
/agents close-done              close every idle or finished pane
/agents close <name>            close one idle or finished pane
/agents close <name> --force    force a close, from the user only
/agents purge                   delete metadata for runs already closed

A widget above the editor shows running, blocked, and undelivered agents while any run is open.

Taking over an agent

Type in an agent pane whenever you want. Manual turns are yours, and the parent never sees them. When you want to hand the conversation back, run this in the child:

/return-to-parent [optional note]

That forwards the last final response, plus your note, to the parent. Typing during a turn the parent requested steers that turn without breaking the link, so the parent still receives the result.

Model profiles

Profiles choose only provider, model, and thinking. They never add a persona, system prompt, skill list, or tool list, so an agent behaves like the Pi you already configured.

Configure them globally in ~/.pi/agent/pi-herdr-subagents.json, or per project in <project>/.pi/pi-herdr-subagents.json, which is read only after you trust the project.

{
  "defaultProfile": "general",
  "profiles": {
    "general": { "provider": "openai-codex", "model": "gpt-5.6-sol", "thinking": "high" },
    "explore": { "provider": "openai-codex", "model": "gpt-5.6-luna", "thinking": "xhigh" },
    "review": { "provider": "anthropic", "model": "claude-fable-5", "thinking": "high" }
  },
  "layout": { "minPaneWidth": 72, "minPaneHeight": 20 },
  "limits": { "maxConcurrentAgents": 4, "maxOpenPanes": 8 },
  "retention": { "deliveredDays": 7, "undeliveredDays": 30 }
}

Without a profile, the child inherits the provider, model, and thinking level the parent is using.

Config files, storage paths, and environment variables still use the earlier subagent spelling. They stay unchanged so existing setups keep working.

Layout and limits

Splits happen only when both resulting panes stay at least 72x20. Otherwise the extension creates or reuses a dedicated agents tab. Pane creation never steals your focus.

Defaults allow 4 agents starting, working, or blocked at once and 8 open panes. There is no hidden queue: past the limit, spawn_agent fails and tells the model to wait or close a pane.

Agents share your working tree. This extension does not create worktrees, so parallel write tasks should touch separate files.

Data and retention

Coordination state lives under ~/.pi/agent/pi-herdr-subagents/runs/<parent-scope>/<run-id>/, readable only by you. The mailbox uses versioned JSON, atomic writes, capability tokens, and acknowledgements, and each run belongs to one exact parent session.

• Inbox files are deleted once Pi confirms the delegated input arrived, or after a failed acknowledgement when Pi never confirms a delivered input.
• The capability token never travels through the pane environment or Herdr argv; the child reads it from the private run.json.
• Active runs and open panes are never cleaned up.
• Closed runs with delivered results are kept for 7 days, and closed runs with an undelivered result for 30 days.
• Each run keeps the retention values that applied when it started, and 0 disables deletion for that category.
• Pi transcripts are never deleted by this extension.
• Resuming a different session never adopts agents from another one.

Project configuration is read only from trusted projects. Trust and permission prompts are never answered for you: an agent waiting on one is reported as blocked so you can decide in its pane.

Development

npm install
npm run check

The Herdr test opens a real pane and runs a child Pi against a live model, so it is skipped unless you ask for it from inside a Herdr pane:

PI_HERDR_SMOKE=1 npx vitest run test/smoke.herdr.test.ts

Set PI_HERDR_SMOKE_PROVIDER and PI_HERDR_SMOKE_MODEL to choose the child model.

Credits

Herdr is built by Can Celik, and this extension exists because of it. Its @ogulcancelik/pi-herdr extension showed how Pi and Herdr fit together, and reading it shaped how this package talks to the Herdr CLI. Thank you.

License

MIT