ECHO: stop flagging routine SSE reconnects as monitor degradation

[posthog-eu]

Problem

• Host live monitoring "looked" unreliable: the monitor stream logged hundreds of monitor_stream_degraded events over 14 days, and the "Reconnecting" badge flickered on the host's screen.
• Almost all of these were routine reconnects, not real outages. The monitor SSE stream auto-reconnects on a browser retry cycle (~3s) and a 5s fallback poll covers the gap, so a brief flap never reaches the host.
• The client debounced degradation by only 3s, which is shorter than the reconnect cycle, so nearly every reconnect tripped it and emitted a degrade/reconnect pair.

Measured over the affected window:

Signal	Value	Meaning
Reconnect downtime p50	1s (from the 3s mark)	Typical outage is a ~4s blip that self-heals
Reconnect downtime p90	~17s	Most events are short
Gap between degrades in a session p50	~10 min	Random, so no fixed connection-lifetime timeout

The server side is healthy (15s heartbeats, a 2s poll safety net, and the fallback poll), so there is no outage to fix there. The signal itself was over-sensitive.

Changes

• Replace the 3s degrade debounce with a 10s grace window (MONITOR_DEGRADE_GRACE_MS) in useConversationMonitor.
• Hold the stream connected through the grace window, so the "Reconnecting" badge and the monitor_stream_degraded event only reflect a stream that fails to recover across several reconnect attempts.
• Genuine long outages (backgrounded tab, sleeping laptop) still flag.
• Add a hook test covering the brief-flap path (no event, no badge flip) and the sustained-outage path (degrade, then a paired reconnect).

Testing

• pnpm typecheck (tsc --noEmit): 0 errors.
• biome check: clean.
• vitest run: new useConversationMonitor tests pass; LiveMonitorSection tests still pass.

Agent context

• Considered lowering the browser reconnect interval with an SSE retry: hint on the server, but rejected it: EventSource does not back off, so a shorter interval risks a reconnect storm against a genuinely down server. The client-side grace window is the targeted fix.

---

Created with PostHog Desktop from this inbox report.