/firstmate-orca
**The operator checklist for the experimental Orca runtime backend.**
Skill type: Agent-only
Upstream path:.agents/skills/firstmate-orca/SKILL.md
The operator checklist for the experimental Orca runtime backend.
What it does
firstmate-orca explains how to safely spawn, supervise, recover, and smoke-test Firstmate work using Orca.
The important model is that Orca is a runtime backend, not an agent harness. Orca owns the task endpoint and worktree; a harness such as Claude, Codex, OpenCode, Pi, Grok, or Kimi still runs inside it.
When to reach for it
Load it when selecting Orca, spawning Orca-backed work, debugging an Orca task, reconciling Orca metadata, or verifying backend lifecycle behavior.
How it behaves
The skill prefers Firstmate's guarded helper scripts instead of raw Orca operations. Existing task metadata remains the routing identity. Switching backend configuration affects future spawns only; already-running tasks keep their recorded backend.
Recovery begins from recorded metadata and preserves the existing task copy rather than manufacturing a replacement task.
Common questions
Does this replace harness-adapters?
No.
Use harness-adapters for harness-specific launch, interrupt, trust, resume, and invocation behavior.
Can raw Orca deletion be used to clean a broken task? The normal path is guarded Firstmate teardown and recovery, not manual backend surgery.
It's working if
- Metadata records Orca as the backend and retains the necessary Orca/worktree identity.
- Peek, send, state, watcher, and teardown operate through Firstmate helpers.
- Recovery preserves existing task work.
- Backend smoke tests stay isolated from unrelated feature work.
Where it fits
This skill is the backend-specific operating layer between Firstmate lifecycle commands and Orca.
Upstream: .agents/skills/firstmate-orca/SKILL.md