feat(go): real-time SSE message streaming + instant "Thinking…" indicator

[baryhuang]

Summary

Makes the OpenAgents Go clients feel real-time: messages arrive push-driven over SSE instead of on a polling timer, and there's instant feedback the moment you send (no dead gap before the agent's first frame). Brings packages/go to parity with the workspace/frontend SSE reference.

Changes

SSE-first message streaming (Swift app + packages/go/web)

Consume /v1/events/stream for push updates, fall back to adaptive polling when the stream is unavailable. While the stream is healthy there is no steady message polling.

• Swift (WorkspaceAPI.swift): new streamEvents(channel:) — a URLSession.bytes SSE client that yields a signal per data: frame and throws on connect failure.
• Swift (WorkspaceStore.swift): the message loop is now SSE-first — each frame triggers an incremental forward poll, reusing pollNewMessages' dedup / optimistic-reconcile / notification path rather than a second decode path. Adaptive-polling fallback (1.5s active / 3s idle) on SSE failure; reopens on session switch.
• Web (packages/go/web): adds getSSEUrl() + the EventSource-first effect in use-polling.ts — now byte-identical to the workspace/frontend reference.

Instant "Thinking…" indicator

The ~1–3s gap after sending is genuine agent latency (nothing has been emitted yet). Show a client-side affordance immediately:

• Flag-based (awaitingFirstResponseSessionIds) — not a placeholder message, which would reorder against the user's real message during reconciliation.
• Set on send (gated on the channel having an agent), cleared when the agent emits any output, with a 45s safety timeout.
• The send button swaps to stop immediately on send.

Backend dependency

SSE delivery uses the existing stream_events Redis pub/sub path. Where Redis isn't configured (REDIS_URL unset), subscribe_events returns immediately and clients transparently fall back to adaptive polling — so this change is safe regardless of backend Redis availability.

Verification

• xcodebuild OpenAgentsGo_macOS → BUILD SUCCEEDED.
• Verified end-to-end against a Redis-backed deployment: [sse] stream open held open (no reconnect churn), idle message polling drops to zero, ~150ms push latency; "Thinking…" appears instantly on send and is replaced by the agent's first frame.

Test plan

• Open a chat with an agent, send a message → "Thinking…" appears instantly, replaced by the agent's response on its first frame.
• With Redis configured: confirm [sse] stream open stays open and no message polling while idle.
• Without Redis: confirm graceful fallback to adaptive polling.

[vercel]

@baryhuang is attempting to deploy a commit to the Raphael's projects Team on Vercel.

A member of the Team first needs to authorize it.

[zomux]

Clean, well-designed SSE streaming + thinking indicator. Signal-based approach on Swift (reusing pollNewMessages) is solid, graceful polling fallback on both clients, proper lifecycle/cancellation handling. LGTM.