herdr-fleet — скилл для Claude Code: что делает и как установить

---
name: herdr-fleet
description: Run several coding agents in parallel in herdr panes, each under its own clausona profile (another Claude Code or Codex account, or an API model such as GLM or DeepSeek), then watch them, check their work and retire them. Use when the user works in herdr and wants work spread across accounts or models, when one account's limits would slow a fan-out, or says "split this across my accounts in herdr", "run these on DeepSeek in herdr".
---

# herdr fleet on clausona profiles

herdr runs agents in panes and shows their state in its sidebar. clausona puts each pane's agent on
the profile you choose.

## 0. Load the shared rules

Load the `fleet-core` skill first and follow it; this skill only adds the herdr steps. In Claude
Code, invoke it with the Skill tool. If that fails, or you run in Codex, read the first of these
files that exists, where `<base>` is this skill's base directory:

- `<base>/../../../fleet-core/skills/fleet-core/SKILL.md`
- `<base>/../../../../fleet-core/*/skills/fleet-core/SKILL.md` (if several versions match, read the
  highest)

If none exists, stop and tell the user to install it: `claude plugin install fleet-core@clausona`
in Claude Code, or `codex plugin add fleet-core@clausona` in Codex.

## 1. Before anything

Stop and tell the user what is missing if any of these fails.

1. You run inside a herdr pane: `HERDR_ENV` is `1` (`test "${HERDR_ENV:-}" = 1`; in PowerShell,
   `$env:HERDR_ENV -eq '1'`). herdr's CLI talks to the session of the pane it runs in. Do not
   control herdr from outside it. In Codex, `HERDR_ENV` is also unset when Codex hides the
   environment (`fleet-core`, Shells). If it does, do not tell the user they are outside herdr:
   tell them to start Codex again in this pane with `-c shell_environment_policy.inherit=all`.
2. `herdr status` shows a running server. Before a command group's first use, run the group alone
   (`herdr worktree`, `herdr workspace`, `herdr agent`, `herdr pane`) to see its current syntax.
   Never run bare `herdr`: it starts the interface.
3. The repo has a remote: workers push their branches.

herdr commands print JSON. Read ids and paths from it; never guess them. herdr runs on macOS, Linux
and Windows, and the commands below are the same on each.

## 2. Starting a worker

For each task, after `fleet-core` picked its profile:

1. **Worktree.** `herdr worktree create --cwd <repo> --branch fleet/<task> --label "<task> · <profile>" --no-focus`.
   Keep the label to 20 characters or fewer, with a short task word such as `docs · ds-flash`:
   herdr's agents list cuts longer labels, and the profile at the end is what shows which account
   runs where. Record `.result.workspace.workspace_id`, `.result.root_pane.pane_id` and
   `.result.worktree.path`. The worktree lands under `~/.herdr/worktrees/`.
2. **Start.** `herdr pane run <pane> "clausona run <profile> -- <args>"`, with the arguments from
   `fleet-core` section 4. Codex workers keep `--disable hooks`.
3. **Check it started.** Within 20 seconds, `herdr agent list` shows the pane with agent `claude` or
   `codex`. If not, read the pane (`herdr pane read <pane> --source recent-unwrapped --lines 60`).
   If the launch line never ran, for example because the shell was still starting, run it once more.
   Otherwise tell the user what the screen shows.
4. **Name it.** `herdr agent rename <pane> <name>`, with a name such as `w1-docs`: lowercase letters,
   digits, `-` or `_`, starting with a letter, up to 32 characters. Use the name from now on.
5. **Folder trust.** The first worker of a profile in a repo that profile never trusted asks
   whether to trust the folder, and herdr shows it as `blocked`. Read the screen
   (`herdr agent read <name> --source visible`) and answer only that question, and only about
   this repository:
   - Claude Code ("Is this a project you created or one you trust?"): the highlighted answer is
     "No, exit". Send `herdr agent send-keys <name> down enter`.
   - Codex ("Trust this folder?"): the highlighted answer is "1. Trust and continue". Send
     `herdr agent send-keys <name> enter`.

   Then `herdr agent wait <name> --until idle --until done --timeout 60000`. A plain
   `agent wait` right after the keys returns the old `blocked` state.
6. **Brief.** Write it as `fleet-core` section 5 says, then send and wait in one command:
   `herdr agent prompt <name> "Read .fleet-brief.md in this folder and do what it says." --wait --timeout 1800000`.
   In Claude Code, run it as a background task (section 3). In Codex, use the Codex form in
   section 3 instead.
7. **Record** the name, pane, workspace, worktree, branch and profile in the table.

Start workers on the same Claude Code profile one at a time, each after the previous one shows its
first output.

The user can switch to any worker's pane in herdr, read along and type into it at any time.

## 3. Waiting and reading

- **Send and wait:** every brief and follow-up goes out as
  `herdr agent prompt <name> "<one line>" --wait --timeout 1800000`. It waits until the worker has
  started its turn and then settles: `done` or `idle` when the turn ended, `blocked` at a prompt
  or question. Do not send with a plain `agent prompt` and then call `agent wait`: a worker that
  has not picked up the prompt yet is still `idle`, and the wait returns at once.
- **Until it works again** (a worker the user is answering):
  `herdr agent wait <name> --until working --timeout 1800000`, then
  `herdr agent wait <name> --until idle --until done --until blocked --timeout 1800000`.
- **The reply:** herdr keeps no record of what a worker said, and Claude Code in full-screen mode
  leaves nothing above the screen to read. So for a Claude Code worker, read its last reply with
  `fleet-core`'s `last-reply.mjs` and the worktree path from section 2 (`fleet-core` section 8),
  before its screen.
- **Read:** `herdr agent read <name> --source recent-unwrapped --lines 120`. If the reply scrolled
  away, use `--source visible`.
- **Follow-up:** write the file, then
  `herdr agent prompt <name> "Read .fleet-followup-<n>.md in this folder and do what it says." --wait --timeout 1800000`.
  `agent prompt` refuses an agent at a prompt (`agent_blocked`); read its screen first.
- **After a `timeout`:** read the screen. If the worker still works, wait again with
  `herdr agent wait <name> --timeout <ms>`, which matches `idle`, `done` or `blocked`. Never send
  the line again just to wait: the worker would get it twice.
- **In Codex** (one foreground wait at a time, three minutes each): give every worker its brief
  before you wait on any. Send each with
  `herdr agent prompt <name> "<one line>" --wait --until working --until blocked --timeout 20000`,
  which returns once the worker has picked the line up. Then wait on each unfinished worker in turn
  with `herdr agent wait <name> --timeout 180000`, and read it when the wait returns.

A `timeout` or `agent_prompt_stalled` does not prove a prompt was lost. Read the screen before
sending again.

## 4. When an account runs out

1. `herdr pane split <old pane> --direction down --cwd <worktree> --no-focus`, and read the new
   pane id.
2. Start the new profile there, as in section 2, steps 2 to 6, with `fleet-core` section 9's brief.
   Name the new agent with a new suffix, such as `w1-docs-2`: the old one still holds its name.
3. `herdr pane close <old pane>`.
4. `herdr workspace rename <workspace> "<task> · <new profile>"`, with the workspace id from
   section 2, so the sidebar names the new account. Keep the label to 20 characters or fewer, as in
   section 2.

## 5. Retiring

After `fleet-core` section 11's checks, run `herdr worktree remove --workspace <id>`. Never add
`--force`. herdr refuses a worktree with changes (`dirty_worktree_requires_force`); then keep the
worker and report why. When it succeeds, it closes the workspace and its panes and keeps the
branch. Confirm the worktree directory is gone.

Never close panes, tabs or workspaces you did not create. The workspace herdr opened for the repo
itself stays.