orca-subagents
Skill1 starsChanged 5 days agoMITView on GitHub
Runs a subagent as an Orca worker in another harness's CLI agent (Codex, Grok, Claude Code, …). Covers checking Orca is reachable, each agent's model slugs and efforts, and collecting results. Use when the `/model-tiers` skill resolves a model the harness's built-in subagent tool doesn't offer, or when the user or a calling skill asks for a subagent from another harness.
In this repository22 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .cursor/rules/beads.mdcCursor rule
- beadsSkill
- arenaSkill
- build-ticketSkill
- deep-reviewSkill
- design-it-twiceSkill
- designSkill
- domain-modelingSkill
- grillingSkill
- land-to-mainSkill
- lets-buildSkill
- lets-designSkill
- markdown-for-humansSkill
- model-tiersSkill
- prepare-commitSkill
- ready-ticketSkill
- reflectSkill
- review-diffSkill
- ticket-statesSkill
- to-ticketsSkill
- verify-ticketSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 22
Run scripts/reachable, in this skill's directory.
- Very specific: line 41
Use the cursor agent only for Composer.
- Distinctive: line 44
When adding a row, add its check to the script.
- Very specific: line 50
Stop that worker with worker-stop, and start a new one on the tier's next entry.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Orca subagents
Orca's orchestration skill is the reference for starting and supervising workers.
Load it, or if it isn't installed, fetch the guide matched to the installed CLI with orca skills get orchestration.
Each worker starts with a fresh context window, as a subagent should.
That guide says to prefer your own subagents.
That holds only when they offer the model /model-tiers resolved; otherwise use an Orca worker.
Reachable
Run scripts/reachable, in this skill's directory.
It reports whether Orca is reachable, and which agents in the table below are installed and signed in here.
It exits non-zero if none can run.
Inside an Orca session, $ORCA_CLI_COMMAND names the executable if orca isn't on PATH; the script uses it.
Agents and models
Pick the agent by model family, and pass the model's slug as --model and the tier entry's effort as --effort.
Slugs follow each family's pattern, so infer one not listed here from its name, e.g. GPT-6.2 Terra → gpt-6.2-terra.
These were current at 2026-11; the list may drift.
| Family | --agent | Example slugs | --effort |
| GPT | codex | gpt-6.1-sol, gpt-6.1-terra, gpt-6-astra, gpt-6-luna | low medium high xhigh max; ultra except Luna |
| Grok | grok | grok-4.7, grok-4.7-build-fast (Grok 4.7 Fast), grok-4.6 | low medium high xhigh |
| Claude | claude | claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5, claude-fable-5-1 | low medium high xhigh max |
| Cursor | cursor | composer-2.5, composer-2.5-fast | none; omit --effort |
Cursor's CLI also lists other families' models, but they need prepaid API credits.
Use the cursor agent only for Composer; reach other families through their own agents.
An agent works only if scripts/reachable reports it signed in.
When adding a row, add its check to the script.
Orca's guide says to pass --model only if the user named one: a model resolved from the user's tiers counts as named.
Orca starts a worker with any --model string.
A model the CLI rejects shows up later as an escalation with stopReason: errored, not as a failed start.
Stop that worker with worker-stop, and start a new one on the tier's next entry.
Brief
Give each worker the brief you'd give a built-in subagent, self-contained, plus:
- What it may edit, usually nothing. A reader works in your workspace with
--worktree current. - A
worker_donebody is three sentences. For anything longer, such as findings, have it write the report to a file under/tmpand pass that file's path with--report-path.
Collect results
Workers report asynchronously.
- Once every subagent is started, you can end your turn or wait. Don't poll or sleep.
Orca wakes you with each
worker_doneorescalation; some harnesses deliver them mid-turn, as messages. - Continue the calling skill once every subagent, built-in and Orca, has reported or been given up on.
Finished workers close on their own.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

