mirror of
https://github.com/headroomlabs-ai/headroom.git
synced 2026-08-27 14:17:10 -04:00
c482d75a73
Reproducibly exercises the multi-agent Codex reconnect/retry storm
(origin §"Latest Correction") against a local proxy: concurrent Codex WS
sessions + parallel Anthropic /v1/messages replays, with /livez probed
every 250ms to detect event-loop starvation. Exits non-zero if /livez
p99 exceeds the configured threshold.
- scripts/repro_codex_replay.py — CLI harness (asyncio + websockets +
httpx, no new pip deps).
- scripts/fixtures/{anthropic_replay_body,codex_response_create_frame}.json
— hand-crafted, fully synthetic fixtures shaped like real traffic.
- scripts/README.md — "Reproducing the reconnect storm" section.
- tests/test_scripts/test_repro_codex_replay_smoke.py — FastAPI/uvicorn
mock server; runs the harness end-to-end in < 3s.
- .gitignore — allowlist the new checked-in scripts (existing rule
previously ignored everything under scripts/ except install.sh/.ps1).
52 lines
2.1 KiB
JSON
52 lines
2.1 KiB
JSON
{
|
|
"type": "response.create",
|
|
"response": {
|
|
"model": "gpt-5.3-codex",
|
|
"stream": true,
|
|
"store": false,
|
|
"parallel_tool_calls": true,
|
|
"tools": [],
|
|
"input": [
|
|
{
|
|
"type": "message",
|
|
"role": "system",
|
|
"content": [
|
|
{
|
|
"type": "input_text",
|
|
"text": "You are Codex, a proxy-aware coding agent. The upstream connection is a WebSocket proxied through a local Headroom process. Keep responses incremental and stream-friendly; the proxy compresses large tool outputs on the way back and expects well-formed response.created / response.output_text.delta / response.completed envelopes."
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"type": "message",
|
|
"role": "user",
|
|
"content": [
|
|
{
|
|
"type": "input_text",
|
|
"text": "We're reproducing a reconnect storm against the local proxy at 127.0.0.1:8787. Open a WebSocket session, send one response.create frame, and stay connected. The goal is not to generate useful output — it's to make sure the proxy's WS lifecycle (accept, upstream connect, relay-task wiring, registry register/deregister) survives N concurrent sessions plus a parallel burst of large /v1/messages POSTs without /livez going dark. Do not expect a real model response; upstream auth will fail and the proxy should gracefully close. Just hold the slot for the storm duration."
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"type": "message",
|
|
"role": "assistant",
|
|
"content": [
|
|
{
|
|
"type": "output_text",
|
|
"text": "Acknowledged. I'll keep this session open, avoid generating extra frames, and let the harness measure /livez latency and registry drain time. If the upstream closes with an auth error I'll expect a clean response.completed or error frame from the proxy, not a dangling task."
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"type": "message",
|
|
"role": "user",
|
|
"content": [
|
|
{
|
|
"type": "input_text",
|
|
"text": "Confirm the session. No further work required — the harness is driving."
|
|
}
|
|
]
|
|
}
|
|
]
|
|
}
|
|
}
|