Herdr voice dispatcher
SkillAI & modelsDispatches spoken or written prompts to live Herdr agents and reports their status or results. Use when the user says to list, check, ask, tell, prompt, steer, or wait for an OMP, Codex, or other agent running in Herdr.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the Herdr voice dispatcher skill
What this skill tells your AI
The instructions your AI receives, as published by edmundmiller/dotfiles in skills/catalog/herdr-voice-dispatcher/SKILL.md and read by ahel’s review.
Translate conversational requests into bounded Herdr CLI actions. This skill is
an external controller for Codex Voice or chat, so it intentionally works
without HERDR_ENV=1.
Connect
-
Run the concise inventory without native session references:
python3 ~/.agents/skills/herdr-voice-dispatcher/scripts/list_agents.py -
If Herdr is unavailable, report the exact error.
- For
Operation not permitted, ask the user to run the Codex task with Full Access; a skill cannot bypass the sandbox.
- For
-
Do not dump raw
herdr agent listoutput. It includes noisy native session references.
Resolve the target
-
Use an exact live agent name or pane ID.
-
If the user names a project, title, or directory, filter the inventory:
python3 ~/.agents/skills/herdr-voice-dispatcher/scripts/list_agents.py --query dotfiles -
Treat labels such as
omporcodexas agent kinds, not unique targets, when several matches exist. -
If more than one candidate remains, ask one concise clarification. Never guess a pane.
Act
List or check status: use the inventory helper. Use herdr agent get "$target" only when more detail about one resolved agent is necessary.
Dispatch work: preserve the user's requested outcome and constraints, then submit without waiting:
herdr agent prompt "$target" "$prompt"
Report only that Herdr accepted the prompt. If the target was already
working, explain that Herdr lifecycle state does not correlate completion to
an individual queued turn.
Read a result: check status, then read bounded output:
herdr agent read "$target" --source recent-unwrapped --lines 80
If recent output is incomplete because the agent uses an alternate screen, use
--source visible. Do not claim success from unknown state or truncated
output.
Wait when explicitly requested: keep Voice responsive with a bounded wait:
herdr agent wait "$target" \
--until idle --until done --until blocked \
--timeout 60000
If the agent becomes blocked, read and relay its question. Do not answer it
for the user.
Guardrails
- Prompt only the target and intent the user authorized.
- Never start, stop, focus, interrupt, send keys, or answer an approval unless the user explicitly requests that exact action.
- Treat dispatch as an external state change; a zero exit status proves submission, not task completion.
- Do not expose agent session paths or unrelated terminal content.
- Never wait indefinitely. Return control after 60 seconds and offer a fresh status check.
Spoken examples
- “Which Herdr agents need attention?” → list and summarize
blocked,working, anddone. - “Tell the dotfiles reviewer to inspect the current diff without editing.” → resolve one target, dispatch once, and return immediately.
- “What did the reviewer say?” → check its state, then read bounded output.
- “The reviewer is blocked—what does it need?” → read and relay the question without answering it.
Signals
- GitHub stars
- 80
- Forks
- 6
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Gateway key
herdr-voice-dispatcher- Source
- github.com/edmundmiller/dotfiles