[renkelvin]

Gap fix B. Since #657, a connected Service's context is saved into the World every hour. Agents can now search and read that saved context with one World tool instead of calling each Service live.

• The tool, world_context (targets context/search and context/read):
  • search takes a query and matches every word against the saved records of every Service the Agent may read. It returns at most 20 results, each with an ID <provider>:<recordId> and a short snippet.
  • read takes one of those IDs and returns the saved record.
  • Results are marked as untrusted content, and the tool is in the untrusted-turn list (turn-trust.ts). It reads only the saved snapshots, so a search makes no request to the Service.
• Gated by each Service's own read Grant.
  • The context targets come with every Service's "Can read" Grant (serviceGrantTargets), so granting a Service also covers its saved context.
  • Each result is kept only if the Agent's own Grants allow read_service on that Service (standingCallDecision). A Service the Agent may not read is never searched for it.
  • Only the Agent's own Grants count, never a Fox turn's (OWN_GRANT_TARGETS). Fox itself is refused with a pointer to read_service.
• Where it lives:
  • Pure Core: core/accounts/services/context.ts (serviceContext, limits and reasons).
  • Snapshots: the host's serviceContext in modules/sources/service-sync.ts, over the current non-Notion sync jobs.
  • Dispatch: world-tool-bridge.ts.

Docs: INTEGRATIONS.md § Searching saved Service context, GRANTS.md, README.md and README.zh.md.

Checks:

• New: scripts/service-context-check.ts, added to test:core.
• Updated for the context targets on a Service Grant:
  • agent-grants-check
  • github-service-check
  • harness-sessions-check
  • mcp-service-records-check
  • world-actions-equivalence-check
• npm run check:pr, npm test and npm run test:ui:pr pass locally.

---

Generated by Claude Code