[contexto]: Prototype Phase 2 — sub-agent context infra on OpenClaw

[sm86]

Background

Phase 2 of the agent handoff roadmap is sub-agents: distributing structured context horizontally so each sub-agent receives only what it needs and composes results back without losing constraints.

OpenClaw's plugin SDK already exposes the surface for this (prepareSubagentSpawn, onSubagentEnded on ContextEngine). Today both are no-ops in AbstractContextEngine (packages/contexto/src/engine/base.ts).

Goal

A design plan + working PoC that turns those hooks into real behavior end-to-end against an OpenClaw sub-agent flow.

Design plan should cover

• Context packet shape — what slice of mindmap state travels with the spawn (filtered episodes? branch subset? path ancestors?).
• Scoping — how to constrain a sub-agent to its task without leaking unrelated branches.
• Return composition — how sub-agent outputs (and any new constraints surfaced during its run) get folded back into the parent mindmap on onSubagentEnded.
• Traceability — how a parent retrieval can later explain "this came from sub-agent X working on Y".
• Backend contract — what (if any) new endpoints the hosted backend needs (e.g. mindmap.subgraph, mindmap.merge).

PoC scope

• Implement prepareSubagentSpawn + onSubagentEnded in DefaultContextEngine (sliding-window can come later).
• One concrete OpenClaw demo: parent agent decomposes a task, spawns one sub-agent with a scoped context packet, sub-agent's episode is ingested under a traceable parent link, parent retrieval can surface it.
• Keep changes additive — don't break existing single-agent flows.

Acceptance criteria

• Design doc committed under docs/explorations/ (or similar) and linked from this issue
• PoC branch with the demo flow runnable via a documented script
• Notes on what didn't work / what the backend needs before this can ship for real