web: stalled-stream watchdog merged in #47571 was lost in the v2 client refactor — zombie SSE requires hard refresh #51857

Description

The web UI still loses live updates whenever the SSE event stream dies silently (tab suspended, network flap, proxy idle timeout) and only recovers after a hard refresh. This was already reported in #39030, #47258 and #45860, but there is an important detail those reports miss:

A fix for this exact failure mode was already merged — and the v2 client refactor removed it.

PR #47571 (fix(client): detect stalled event streams and resync on foreground, merged 2026-09-06) implemented a proper stall watchdog in packages/client/src/solid/connection.ts:

• every received chunk (including server.heartbeat keepalives) pushed a stall deadline forward;
• an idle setTimeout aborted the request with "Event stream stalled", triggering reconnect + resync;
• visibilitychange and online events forced a resync;
• a connected | connecting | reconnecting status was exposed to consumers.

On current dev (083ed266e, v1.18.33), packages/client/src/solid/ no longer exists, and the app's live event path is packages/app/src/context/server-sdk.tsx, which contains a naive reconnect loop with none of that liveness logic:

// packages/app/src/context/server-sdk.tsx (~L268-309)
while (!abort.signal.aborted && started && generation === active) {
  attempt = new AbortController()
  try {
    const events = kind === "v1"
      ? (await eventSdk.global.event({ signal: attempt.signal })).stream
      : eventApi.event.subscribe({ signal: attempt.signal })
    for await (const event of events) { /* ... */ }
  } catch (error) { /* logs, then falls through */ }
  await wait(RECONNECT_DELAY_MS) // fixed 250ms — see also #48014
}

The server already emits server.heartbeat every 10s and sets X-Accel-Buffering: no (packages/opencode/src/server/routes/instance/httpapi/handlers/event.ts), so the keepalive exists on the wire — the client just never uses it to detect a dead stream. If the TCP connection half-opens (suspended tab, NAT/proxy drop without RST, cloudflared/reverse-proxy restart), for await never resolves and never rejects: the UI is stale forever.

Additionally, pagehide always calls stop(), but pageshow only restarts when event.persisted === true (resumeStreamAfterPageShow) — and an open streaming fetch makes the page ineligible for bfcache, so persisted is almost always false.

Expected behavior

Port the watchdog semantics that #47571 introduced into server-sdk.tsx (or wherever the app's event stream now lives):

// sketch: stall watchdog driven by the existing 10s server heartbeat
const STALL_TIMEOUT_MS = 3 * HEARTBEAT_INTERVAL_MS + 5_000

let watchdog: ReturnType<typeof setTimeout> | undefined
const arm = () => {
  clearTimeout(watchdog)
  watchdog = setTimeout(() => {
    attempt?.abort(new Error("event stream stalled"))
    // then resync: existing server.connected path already re-bootstraps
    // active directories in server-sync.tsx
  }, STALL_TIMEOUT_MS)
}
// inside the for-await loop: arm() on every event (heartbeats count)

plus a visibilitychange listener that aborts/resyncs when the page returns to the foreground after going quiet — restoring the behavior users had before the refactor.

Steps to reproduce

1. opencode web behind any proxy (or directly), open a session.
2. Suspend the tab (Chrome: chrome://discards, or switch apps on mobile / sleep the OS long enough for the TCP connection to drop without RST).
3. Return to the tab.
4. Trigger a new event server-side → the UI does not update. Ctrl+F5 required.

OpenCode version

v1.18.32–v1.18.33 (dev @ 083ed266e)

Operating System

Observed on Linux server (Nginx + Cloudflare Tunnel in front); reproduction is client-side and platform-independent.

Related

#39030, #47258, #45860, #46733, #49861, #43519 (server-side: heartbeats reset chunkTimeout, so dead streams never expire either), #48014 / PR #48015 (missing reconnect backoff, adjacent), PR #47571 (the merged fix that was lost).

Related work (open issues this would resolve or interact with)

Triage note for maintainers

This is the single highest-impact web reliability defect from the user's perspective: it produces daily "frozen" UIs that only Ctrl+F5 recovers, and it keeps regenerating new reports (#39030 → #47258 → mobile variants). The fix shape is already proven — it existed in packages/client/src/solid/connection.ts — and mostly needs porting to server-sdk.tsx plus coverage to prevent a third regression. Requesting prioritization together with #48014/PR #48015 (backoff) and #43519 (server-side stall reap) so client and server stop trusting half-open connections.