SSE client drops every ~16s on stable 1.92.1 — reconnect backoff runaways to 120s (no keepalive, no 429 handling)

[liuahiqun8888]

SSE client drops every ~16s on stable @evomap/evolver@1.92.1 — reconnect backoff runaways to 120s (no keepalive, no 429 handling)

Summary

On the current stable @evomap/evolver@1.92.1 (published 2026-07-16), the daemon's SSE connection to the hub drops roughly every 16 seconds, and the client's reconnect backoff then runs away exponentially (5s → 10s → … → 120s) and never recovers — so the event stream is effectively dead until something external resets it.

I reproduced and root-caused this locally. This is filed as an Issue (not a PR) because src/gep/a2aProtocol.js in this repo appears to be the obfuscated/minified release artifact (the file I fetched decodes to _0x… obfuscated code, not source), so I can't open a clean PR diff against it. See "Repo state note" below.

Environment

• Package: @evomap/evolver@1.92.1 (latest dist-tag, 2026-07-16)
• SSE transport: eventsource package (v4.x) via require('eventsource') inside src/gep/a2aProtocol.js → hubOpenEventStream() / startEventStream()
• Hub: https://evomap.ai/a2a/events/stream?node_id=…&duration_ms=…
• OS: win32 (Node 22)

Symptoms

1. [SSE] Event stream connected appears, then ~16s later the socket is killed (intermediate proxy/firewall drops idle long-lived streams).
2. onerror fires; reconnect is scheduled via _sseReconnectMs which doubles each error and is never reset on a successful (re)connect.
3. Over a few minutes the reconnect gap climbs to the 120s cap and stays there → SSE is "fake-dead".
4. Hub rate-limits SSE connections hard (reproduced frequent HTTP 429), which the client does not honor (no Retry-After handling, no jitter) → tight retry storm makes the 429s worse.

Root causes (all reproduced locally)

1. No client-side keepalive / idle probe. When the hub is idle it sends zero bytes, so the proxy kills the idle socket at ~16s. A client ping/heartbeat on the SSE socket is impossible because SSE is server→client only — the fix must come from the hub sending periodic : keepalive comment frames.
2. _sseReconnectMs only grows, never resets. The backoff starts at 5000ms and doubles on every error, but _sseReconnectMs is not reset to base on a successful onopen, so it saturates at 120000ms and stays there.
3. No 429 / Retry-After handling, no jitter. Rapid reconnects trigger hub 429s; the client ignores them and retries immediately, creating a reconnect storm across nodes.

Proposed fixes

Client (a2aProtocol.js, once source is available):

• Reset _sseReconnectMs to a base (_sseReconnectBaseMs, e.g. 5000) inside onopen on every successful (re)connect — this is the core runaway fix.
• In onerror, honor HTTP 429 / Retry-After, apply ±30% jitter, and cap at _sseMaxReconnectMs to avoid synchronized storms.

Server (hub, /a2a/events/stream) — required for 100% uptime:

• Send a : keepalive\n\n comment frame every ~10s (comment frames don't trigger onmessage but keep the TCP socket non-idle, dodging the 16s proxy kill).
• Add X-Accel-Buffering: no + Cache-Control: no-transform so nginx/CDN don't buffer the stream.
• Return Retry-After on 429 to coordinate with the client backoff.

Note: the existing POST /a2a/heartbeat is a separate TCP connection (node registration) and does not keep the SSE socket alive — it's a different path.

Local diagnostics

• curl -N against /a2a/events/stream returned HTTP 429 in ~1.6s with no keepalive bytes in 22s → confirms hub sends nothing when idle and rate-limits aggressively.
• Observed pre-patch reconnect gaps: 7s → 5s → 12s → 48s → 118s (runaway to the 120s cap). Post local client-side patch (bounded backoff + onopen reset): stable 2–4s, no 429 storm, no crash.

Repo state note (why Issue, not PR)

When I fetched src/gep/a2aProtocol.js from main it decodes to obfuscated code (const _0x1f36f7=_0x447a;…), so there is no readable source to diff against for a PR. Meanwhile, test files in this repo (test/heartbeatResilienceRound3.test.js, test/issue600EventSourceFallback.test.js) suggest the SSE client has been reworked on a 2.0.0 line (fetch-based stream + a round-3 audit that reset backoff on wake).

Questions for maintainers:

• Is the fix already shipping in the 2.0.0-beta series (latest 2.0.0-beta.5, 2026-07-21)? If so, can stable 1.92.1 get a backport, or should users move to beta?
• Can the hub emit : keepalive frames so idle SSE survives proxies?

Local workaround (for others hitting this now)

A non-invasive preload that wraps the eventsource EventSource with a bounded-backoff reconnect facade, keeping the daemon stable on 1.92.1 until an official fix lands. Applied via NODE_OPTIONS="--require sse-idle-watchdog.js". Effectively bounds reconnect to 2–4s with no 429 storm.

[owenshen0907]

🤖 @bjw9808 这条 issue 在你负责的仓库,麻烦认领或跟进一下。(agent-control 自动提醒;未自动接入)

[liuahiqun8888]

这是来自QQ邮箱的假期自动回复邮件。   你好，我最近正在休假中，无法亲自回复你的邮件。我将在假期结束后，尽快给你回复。