English | 中文
This repository builds an improved Agent Skill for Herdr, focused on practical multi-agent coordination without heavy waiting or polling.
It is not the Herdr application itself. The deliverable is the installable skill directory at herdr/, whose main entry point is herdr/SKILL.md.
Agents often waste turns polling sibling panes or blocking on vague done status. This skill defaults to a lightweight chat contract:
- discover with
herdr agent list/herdr pane list; - send a short message with reply-to (and optional task);
- treat successful send as delivered, then end the turn;
- do not poll the peer and do not block-wait for replies;
- when the peer injects a reply into your pane, handle that inbound message;
- keep detailed CLI notes behind progressive disclosure.
herdr/
SKILL.md core agent-facing workflow
scripts/herdr-msg thin header + pane-run helper
references/command-reference.md detailed Herdr CLI reference
agents/openai.yaml UI metadata for skill clients
The helper is plain Bash (no Python/jq).
npx skills add yangyang0507/herdr-skillManual: copy herdr/ into a client skill directory such as .agents/skills/herdr.
Use from a Herdr-managed pane where HERDR_ENV=1.
bash -n herdr/scripts/herdr-msgMessages look like:
[herdr-msg reply-to:w1:p1 task:review]
Please review src/api. Reply DONE or BLOCKED to reply-to.
scripts/herdr-msg:
- resolves the target agent/pane;
- autofills
reply-tofrom SELF; - sends via
herdr pane run; - checks the target once; if an idle/done agent did not start, sends Enter and re-checks;
- prints a receipt (
deliveredorunconfirmed, plustarget_status); - exits.
There is no default msg_id, sentinel, --verify, or --wait-reply. Those were useful for machine wait-matching; for day-to-day chat they add weight without helping the natural “push reply into my pane” wake-up.
herdr/scripts/herdr-msg codex --task review <<'MSG'
Review the lightweight protocol.
MSGPost-send protocol: send → check once → delivered/stop, or unconfirmed/Enter-once/stop.