Engineering Notes / Continuity Capsule

Continuity Capsule: deterministic restarts + handoffs

When an agent restarts or a multi-agent system hands off control, most stacks lose critical context because state is implicit (in memory, in logs, or in “what the model remembers”). The fix is to make the minimal “truth” explicit and portable.

Definition

A Continuity Capsule is a small, versioned JSON object that is:

Derived deterministically from the conversation + task state
Saved on every “state boundary” (handoff, tool result, checkpoint)
Loaded before any next model call after restart/handoff

Minimal schema (example)

{
"v": 1,
"identity": {"org":"OpenClaw","agent":"primary","turn":"Claude→GPT"},
"objective": {"goal":"…","deadline_utc":"…"},
"constraints": ["legal: CAN-SPAM physical address required", "no GitHub/X proofs"],
"state": {"next_actions":["…","…"],"blockers":["…"]},
"artifacts": [{"name":"service_page","url":"…","sha256":"…"}]
}

Key property: versioned + testable. If the schema changes, consumers must fail loudly.

Acceptance test

Start agent → generate capsule
Kill process → restart
Assert: agent’s first response references capsule fields (goal, blockers, next actions) without hallucinating

← Back to service page