CSE Roadmap

Common tier · about 3 hours · 12 parts

Part 6 of 12 · 1 min read

What the simulation measured

Resume behaviour, idle timeouts, buffering and backpressure, each with a reading guide.

How to read the first table: both rows start from the same drop after 20 of 50 events; the difference is only whether the reconnect carries Last-Event-ID.

Scenario (simulated) Outcome
First signal, silent vs streamed (60 s run) 60 s vs 0.4 s
Drop after 20 of 50: blind reconnect 70 received, 20 duplicates
Drop after 20 of 50: Last-Event-ID 50 received, 0 duplicates, 0 missing
Silent tool call 4 s to 95 s, 60 s proxy closed at 64 s; kept open by 15 s heartbeat
First byte: no buffering / 4 KiB / 16 KiB 0.05 s / 5.15 s / 10 s
Stalled client, queue 10 drop deltas: 89 dropped; block: 89 refused; disconnect: closed

What to notice: blind reconnect looks like it "worked," since 70 events arrived. But 20 of them are repeats the UI would show twice. Carrying Last-Event-ID is what turns a drop into a clean resume instead of a glitch.

How to read the second table: it compares the two transports on the same three questions, not on raw speed.

WebSockets SSE + context management
Direction two-way server to client; input by POST
Infrastructure session affinity standard load balancing
Waiting for a human for hours limited by connection timeouts works, no resources held

What to notice: the deciding factor for a long-running agent is usually the third row. An approval that waits for hours is simple to support when the context is saved and reloaded, rather than held open on one in-memory connection.

{{svg:stream|Resumable stream with ids, an event store and Last-Event-ID, plus what breaks it.}} {{svg:results|Resume, heartbeat, buffering and backpressure.}}