Pi Compatibility
Pi is a first-class Meta Harness target. Pi can consume the canonical Agent Skills tree directly; no Pi-specific copy of the skill is required.
Paths
- Project skill:
.agents/skills/harness/ - User skill:
~/.agents/skills/harness/ - Optional extensions: Pi’s configured extension/package location
- Durable portable artifacts:
docs/harness/and_workspace/
The installer’s standard layout is the portable starting point. Keep any Pi
extension configuration outside the canonical skill so it remains removable.
Capability profile
| Capability | Base Pi | Pi with pi-safe-agent-team |
|---|---|---|
| Agent Skills and progressive disclosure | Supported | Supported |
| Spawned workers | Extension/runtime dependent | Supported |
| Parent/peer messaging and clarification | Extension dependent | Supported |
| Task state, ownership, and write fencing | Not assumed | Supported by the integration when configured |
| Model/provider/thinking routing | Supported by Pi configuration | Supported per worker where configured |
| Isolated workspaces and lifecycle recovery | Configuration dependent | Supported where the integration exposes it |
pi-safe-agent-team is optional and is not a portable dependency. The guide
describes the semantic boundary; it does not couple Meta Harness to an
internal broker protocol.
Lowering and fallback
If the optional team integration is unavailable, use the portable skill in the
root Pi session. Do not pretend advisory ownership is mechanical: serialize
conflicting writes and persist cross-session results under _workspace/.
Single-agent execution is the correct default for tightly coupled work.
Rippability
Removing Pi extensions or team configuration leaves .agents/skills/,
docs/harness/, and _workspace/ usable. Keep model/provider choices and
recovery heuristics in the optional runtime layer.
See runtime capabilities and the compatibility matrix for shared policy.