/harness-adapters
**The router and reference set for harness-specific behavior.**
Skill type: Agent-only
Upstream path:.agents/skills/harness-adapters/SKILL.md
The router and reference set for harness-specific behavior.
What it does
harness-adapters is the single routing owner for operations that vary by agent harness.
It currently carries verified facts for Claude, Codex, OpenCode, Pi, Pi-signed, Grok, Kimi, Cursor, Gemini, and Muse.
The files under its references/ directory are supporting resources of this one skill, not separate skills.
When to reach for it
Load it before spawning or recovering workers, handling trust dialogs, invoking harness-specific skills, interrupting or exiting agents, resuming exited agents, selecting model/effort behavior, or verifying a new adapter.
How it behaves
The skill first chooses the operation, then loads the required common references plus exactly the selected harness reference.
It refuses to dispatch work on an unverified adapter.
For recovery and control, the recorded task harness= is authoritative; Firstmate does not guess from the model provider.
Lifecycle control is delivered through the guarded control plane rather than typed as chat text.
Common questions
Is Pi-signed a separate reference file? It routes through the Pi reference while preserving its distinct verified identity.
Can Muse or Gemini be used everywhere? Their verified role is narrower; the skill records those support boundaries.
It's working if
- Only verified adapters receive dispatches.
- Each operation loads the minimum relevant common and harness references.
- Recovery uses recorded harness identity.
- Harness-specific control never gets accidentally sent as ordinary chat.
Where it fits
This skill is the portability layer that lets Firstmate keep one lifecycle while supporting multiple coding-agent runtimes.
Upstream: .agents/skills/harness-adapters/SKILL.md