feat: run chat fully offline with CopilotKit Intelligence optional

[matthewhand]

Summary

CopilotKit Intelligence is no longer a hard requirement. Three runtime modes now exist:

Mode	Trigger	Threads/history	Realtime
Cloud	CPK_INTELLIGENCE_API_KEY set	Intelligence cloud	managed realtime WS (unchanged)
Self-hosted	INTELLIGENCE_API_URL + INTELLIGENCE_WS_URL (must be set together)	your Intelligence endpoint	your WS
Local	no key, no URLs	local thread endpoints, persisted by DurableAgentRunner	in-process SSE

Without any Intelligence configuration the runtime falls back to its official SSE mode: local /threads list/messages/state endpoints, with a new DurableAgentRunner persisting threads, messages, events and state to the store the app already uses — snapshotting before and after each run, merging by runId and message id — so conversations survive restarts.

The client needs no changes: chat.tsx already branches on richThreads, and /api/main-thread serves a stable local id when Intelligence is absent.

Notes

• assertApiDeploymentConfig now validates the shim contract (URLs paired) instead of requiring the cloud key; intelligenceConfigured() is the single switch. The intelligenceKeyRequiredMessage export is kept for the demo entry, which still requires a key by design.
• Cherry-picked with the runtime identity kept as upstream's auth.owner pattern (a deployment-specific intelligenceUserId refactor was deliberately excluded from this PR).

Testing

• Full suite green (277/277) at current main; tsc --noEmit clean.
• Live-verified both directions on a real deployment: local mode boots with no cloud key, streams a real browser round-trip with zero cloud calls, and rehydrates the conversation from the local store after an API restart; cloud mode still reports mode=intelligence and streams as before.

Adapted from sunshaoan0808/openmuse (MIT), planB + planA commits, with attribution.

🤖 Generated with Codebuff