Moved. This skill now lives in
OpenSWE/skills, with its full history. This repo is archived and gets no updates. Install withnpx skills add OpenSWE/skills --skill herdr-advisor, or/plugin install herdr-advisor@opensweafter/plugin marketplace add OpenSWE/skills.
Pair a Herdr worker with one advisor — a stronger model, running in its own pane, that answers the worker's questions on the user's behalf to unblock it and leads it to its next task.
The worker does all the work and makes every change. The advisor is a read-only leaf: it never
edits, it tells the worker what to change. A worker that stops with "still waiting on your go"
gets go and keeps moving — which is the point. The pair runs while you are away.
Invoked by the user (/herdr-advisor); the agent invokes it itself in one case only: a grilling
session ends, the user confirms a plan for it to execute, and no advisor pair is live.
npx skills add OpenSWE/herdr-advisor
/plugin marketplace add OpenSWE/herdr-advisor
/plugin install herdr-advisor@herdr-advisor
Eight places in the skill hardcode ~/.agents/skills/herdr-advisor/ — the watchdog's launch line,
the advisor's handoff prompt, and four harness rows that pass a config file by absolute path. The
skill only works when it lands at exactly that path, under exactly that name. Both routes above
produce it, and so does agent-sync adopt <clone>/skills/herdr-advisor if you would rather track a
git clone. If you vendor it anywhere else, rewrite those paths.
| Herdr | The terminal multiplexer this drives. HERDR_ENV=1 must be set — outside Herdr the skill stops. |
The herdr skill |
A separate skill this one reads before operating on agents: npx skills add herdrdev/herdr. Not bundled here. |
python3 |
The watchdog and its self-test. Standard library only — no packages to install. |
| macOS | Only for the watchdog log at ~/Library/Logs/herdr-advisor.log. Everything else is portable. |
| A second CLI, logged in | The advisor runs in its own harness. 16 CLIs are mapped in HARNESS-CLIS.md. |
All under skills/herdr-advisor/:
| File | Reader | Purpose |
|---|---|---|
SKILL.md |
the worker | Choose the advisor, split the pane, plant the watchdog, hand off, stop the pair. |
ADVISOR.md |
the advisor | Read-only rules, the polling loop, decision precedence, next-task sources, end conditions. |
MODELS.md |
worker | 29 models ranked by DeepSWE v1.1 Pass@1, plus who is eligible to advise whom. |
HARNESS-CLIS.md |
worker | 16 coding-agent CLIs: Herdr kind, effort rung, exact launch arguments, and per-CLI gaps. |
watchdog.py |
— | Keeps the advisor's loop alive from outside its harness: re-prompts any turn that ends while the pane still holds an advisor name. 8 blocks/hour cap; exits when the name is cleared. |
watchdog_selftest.py |
— | 8 offline scenarios against a fake herdr on PATH. No Herdr needed. |
harness/ |
— | Per-CLI read-only permission configs (Devin, Kimi, Kiro, oh-my-pi, OpenCode) and launch-test.sh. |
Offline, no Herdr required — this is the regression gate:
cd skills/herdr-advisor && python3 watchdog_selftest.py # prints: okIt takes the path to watchdog.py as an optional argument, defaulting to the current directory.
Live, one HARNESS-CLIS.md row end to end (needs a running Herdr session and that CLI logged in) —
it splits a pane, plants the watchdog, starts the agent, verifies a re-prompt, clears the name,
verifies the watchdog exits, then closes the pane:
sh harness/launch-test.sh <root-pane> <kind> [--env K=V ...] -- <native args>A row in HARNESS-CLIS.md without a † is one this has passed.
DECISIONS.md — 31 entries, Q22 onward, recording every consequential call behind
this skill, kept per the log-decisions skill. Start
there before changing a rule; most of them are load-bearing in a way that is not obvious.
Extracted from soulmachine/skills with its full commit
history. Configs ship with placeholders (<model-id>, <your-devin-org-id>) — fill them in before
first use.
MIT