| title | Scope | |
|---|---|---|
| status | active | |
| last_updated | 2026-06-15 | |
| owner | maintainers | |
| related_aps |
|
|
| related_adrs |
|
v1 commits to all three pillars on a minimal, custom harness (chosen over OpenHands/LangGraph so the contribution — the mechanism and the contract — stays unobstructed and the ablations stay trivial):
- Unified distillation — one mechanism writing the Continuation State for both compaction and checkpoint.
- Re-grounding resume — the RGR protocol.
- Effect-safety — write-ahead effect ledger + idempotency, evaluated on a small tool set.
v1 targets a single agent and a controlled/code+web evaluation.
- Multi-agent recovery — recovering a swarm; shared coordination state; partial failure of one agent.
- Cross-model / cross-version resume — checkpoint on model A, resume on model B (RGR's structural advantage over replay).
- Human-in-the-loop recovery for the
never-retryeffect class. - Contract adoption — the boundary contract used by frameworks as a standard recovery interface.
- Formal effect-safety guarantees / verification.
- Very-long-horizon agents (days–weeks) — recovery as the enabling subsystem for persistent autonomous workers.
- Being a full agent framework.
- Training or serving models.
- General-purpose workflow orchestration.
Cairn is the recovery layer — nothing more, deliberately.
Cairn assumes (does not provide) a Code Harness that emits code-as-action and a Runtime that provides a sandbox, filesystem, and tool access. The minimal harness built in Phase 3 exists to exercise recovery, not to be a general agent platform.