Agent-Skills.md

Agent Skills: Herdr

Control herdr from inside it. Manage workspaces and tabs, split panes, spawn agents, read output, and wait for state changes — all via CLI commands that talk to the running herdr instance over a local unix socket. Use when running inside herdr (HERDR_ENV=1).

UncategorizedID: edmundmiller/dotfiles/herdr

Repository

edmundmiller/dotfiles
edmundmillerLicense: MIT
746

Install this agent skill to your local

pnpm dlx add-skill https://github.com/edmundmiller/dotfiles/tree/HEAD/skills/conditional/herdr/herdr

Skill Files

Browse the full folder contents for herdr.

Download Skill

Loading file tree…

skills/conditional/herdr/herdr/SKILL.md

Skill Metadata

Name
herdr
Description
Control live Herdr workspaces, tabs, panes, agents, worktrees, plugins, and waits. Use when running inside Herdr (`HERDR_ENV=1`) to inspect sibling agents, delegate work—including dedicated Pi workspaces—run services, debug agent detection, or coordinate terminal state.

Herdr

Control the current Herdr session through its high-level CLI wrappers. Prefer semantic agent commands for agent lifecycle and communication; use pane commands for shells, processes, terminal input, and output.

Guard

Check HERDR_ENV=1 before controlling a session. If absent, report that the current pane is not Herdr-managed and stop. Do not infer the focused pane from outside Herdr.

test "${HERDR_ENV:-}" = 1

Treat Herdr IDs as live handles, not durable identifiers. Parse IDs from command responses or use --current; never hard-code example IDs into automation.

Choose the narrowest surface

  1. Harness resource: When herdr:// reads are supported, inspect herdr://status, herdr://snapshot, herdr://workspaces, herdr://tabs?workspace=…, herdr://panes?workspace=…, or herdr://pane/<id>?source=recent&lines=80. This avoids shell parsing.
  2. Agent CLI: Use herdr agent list|get|read|send-keys|prompt|rename|focus|wait|attach|start|explain for detected agents.
  3. Resource CLI: Use workspace, tab, pane, worktree, and integration commands for terminal topology and processes.
  4. Raw API: Use only for protocol clients or event subscriptions. Inspect the installed schema first with herdr api schema --json.

herdr --skill prints the agent skill bundled with the installed binary. Treat it as the version-accurate baseline when installed behavior and this skill disagree.

Read references/cli-map.md for the command map and references/recipes.md for trace-tested coordination and recovery patterns.

Inspect before acting

Start from live state:

herdr agent list
herdr pane current
herdr workspace list

For one agent, gather semantic state, recent output, and detection evidence:

~/.agents/skills/herdr/scripts/agent_context.py <agent-name-or-pane-id> --lines 80

Use agent explain when status is wrong, stuck, or unknown; do not guess from screen text alone.

Coordinate agents semantically

Agent start requires an existing pane sitting at an interactive shell prompt; it never creates layout. Create the pane first, then start the agent in it:

split=$(herdr pane split --current --direction right --cwd "$PWD" --no-focus)
review_pane=$(printf '%s\n' "$split" | ~/.agents/skills/herdr/scripts/extract_ids.py pane)
herdr agent start reviewer --kind omp --pane "$review_pane"
herdr agent prompt reviewer "Review the current changes and report only actionable findings." \
  --wait --timeout 120000
herdr agent read reviewer --source recent-unwrapped --lines 100

Use names only when unique. Otherwise use the pane ID returned by pane split or a fresh agent list response. Agent names match [a-z][a-z0-9_-]{0,31} and must be unique among live agents; use herdr agent rename <pane> <name> to name a manually launched agent.

agent prompt submits the text plus an encoded Enter in one call and honors bracketed paste. Plain --wait waits for the first settled idle, done, or blocked state — that is the default for normal work, so do not restate it as --until done. Use --until only for a state-specific workflow such as waiting for blocked; it requires --wait. From a non-working state, a lifecycle change must be observed within 5s or the call returns agent_prompt_stalled; --wait tracks lifecycle state, not turns.

State meanings: idle = ready for input AND its tab has been seen in the focused UI. done = the same ready state after background work, held until the tab gains focus. Focusing the tab, or targeting the pane/agent with a focus command, marks it seen; CLI reads do not. blocked = an approval or question UI is up. unknown is not a successful completion. A status wait observes agent state, not arbitrary command completion.

Run processes in panes

Use pane primitives for servers, tests, logs, and shells:

SPLIT=$(herdr pane split --current --direction right --no-focus)
# Parse the new pane_id from SPLIT; do not predict it.
herdr pane run <pane-id> "npm run dev"
herdr pane wait-output <pane-id> --match "ready" --timeout 30000
herdr pane read <pane-id> --source recent-unwrapped --lines 40

pane wait-output searches the selected snapshot immediately, including output that already arrived, before it polls — so it is safe for output that may have landed already. Use --match <text> for a literal substring or --regex <pattern> for a Rust regex; omitting --timeout waits indefinitely. Read recent-unwrapped when matching or copying text so soft wraps do not corrupt it. Use --source detection to see the plain-text bottom-buffer snapshot Herdr itself uses for agent detection.

Workspaces, worktrees, and layouts

Use a workspace for a project context, a tab for a subcontext, and a pane for one process. Prefer Herdr worktree commands when isolation is part of the task:

herdr worktree list
herdr worktree create --help
herdr workspace create --cwd /path/to/project --label api --no-focus

Inspect installed help before using less-common worktree/plugin flags; these evolve faster than the core commands. Worktree commands are JSON-only and no longer advertise --json, though the flag is still accepted.

workspace create also creates the first tab and its root pane — parse .result.root_pane.pane_id and use that pane before splitting further. Likewise tab create returns .result.root_pane.

There is no layout export/apply in 0.8.0. Capture herdr pane layout output for reference and script workspace create/pane split sequences that parse returned IDs:

herdr pane layout --current

Input rules

  • Use agent prompt for agent prompts (submits text plus encoded Enter in one call; honors bracketed paste).
  • Use pane run for shell command text followed by Enter.
  • Use pane send-text plus pane send-keys … enter for literal TUI input when no agent adapter applies.
  • Pass key combos such as ctrl+h, shift+tab, or named punctuation. Do not pass configuration strings such as prefix+] to send-keys.
  • Use agent send-keys for agent UI interaction (esc, up, enter, ctrl+c; escape aliases esc). Pane input targets the terminal regardless of occupant; agent input is rejected if the agent no longer controls the pane.

Recovery

  • Re-read IDs after closes, moves, or reconnects.
  • If a command reports an unsupported flag, inspect installed --help; do not continue from stale examples.
  • If config changed but behavior did not, run herdr server reload-config and inspect diagnostics.
  • If agent status is wrong, run herdr agent explain <target> --json before changing detection rules.
  • If an output wait times out, read recent output and agent state before retrying.
  • If a helper cannot reach Herdr, verify HERDR_SOCKET_PATH and HERDR_ENV; never scan unrelated sockets.

Known pitfalls (from session traces)

  • Nonexistent guessed commands: herdr agents, herdr agent status, herdr pane stop, herdr worktree sessions, herdr wait, herdr layout — run herdr <group> --help instead of guessing. Bare herdr is not discovery; it launches or attaches the TUI.
  • Opening a workspace or worktree auto-creates panes; run herdr agent list and reuse an existing idle agent before agent start, to avoid duplicate agents in one workspace.
  • On agent wait timeout (exit 1, JSON on stderr): run herdr agent read <t> --source recent-unwrapped --lines 80 and herdr agent explain <t> --json before retrying; never re-issue the same wait blind. Always pass --timeout — waits are indefinite by default.
  • Full-screen agents may render on the alternate screen, so those rows never enter scrollback: if a larger --lines adds no text, read --source visible after scrolling in the agent, or (fallback only) ask the agent to write its full answer to a temp file and read that.
  • Pane and agent reads set truncated: true when older rows were dropped. Do not answer from a truncated read: raise --lines (capped at 1000; there is no offset flag, so omitted older rows are unrecoverable). If it is still truncated, read a durable file/log or ask the agent to restate the result.
  • After herdr pane rename, also run herdr tab rename — stale tab labels misdirect later targeting; trust agent list cwd/session fields over labels.
  • integration install omp fails with "Pi and OMP resolve to the same extension directory" when run from a wrapped OMP that exports PI_CODING_AGENT_DIR=~/.omp/agent. Clear that variable and let PI_CONFIG_DIR place OMP: PI_CODING_AGENT_DIR= PI_CONFIG_DIR=.omp herdr integration install omp. Pi's own install passes an absolute PI_CODING_AGENT_DIR="$HOME/.pi/agent" and needs no change.

Bundled resources

  • references/cli-map.md — high-level command selection and response rules.
  • references/recipes.md — delegation, service, layout, and failure-recovery recipes.
  • references/pi-workspace.md — dedicated Pi workspace delegation and handoff workflow.
  • scripts/start_pi_workspace.py — create a workspace, launch Pi, and submit a prompt file.
  • scripts/send_prompt_to_pane.py — submit a prompt file to an existing pane.
  • scripts/write_handoff_prompt.py — generate a structured child-agent prompt.
  • scripts/monitor_pane.py and scripts/extract_ids.py — inspect delegated work and parse live IDs.
  • scripts/agent_context.py — bounded JSON snapshot of one agent's metadata, recent output, and detection explanation.
  • herdr --skill — the skill bundled with the installed binary; authoritative for the running version.
  • Upstream reference: https://herdr.dev/docs/agent-automation/ — official automation primitives for herdr 0.8.0.