What happened

When using openai-codex / gpt-5.5 as the primary model, requests frequently stall until the 90s stale timeout is hit, then retry and stall again. A single simple message ("报数") took 401.8s to get a response — 5 API calls, 4 of them hitting the 90s stale limit.

Other servers calling the same openai-codex / gpt-5.5 respond quickly (3-7s). Network latency to chatgpt.com is fine (~0.32-0.39s TLS).

Observed pattern

Not every request is slow — it is fast-slow alternating:

• Fast: 3.1s, 4.4s, 5.3s, 6.5s, 7.1s
• Slow: 96.8s, 98.4s, 99.2s, 103.2s, 150.2s, 208.6s, 290.2s

Input context was ~24k-28k tokens per retry (no growth between retries).

Environment

• Hermes Agent v0.14.0
• macOS (Mac Mini)
• Provider: openai-codex / gpt-5.5 via chatgpt.com/backend-api/codex
• ~20 active sessions, ~24 cron jobs concurrent
• Telegram gateway

The question

Is this a compatibility issue between Hermes and the openai-codex backend? Or is it purely a provider-side issue? Not sure — but other servers using the same model do not show this behavior.

Temporary workaround

Switched primary model to a different provider (mimo-v2.5-pro via custom), which resolves the issue immediately. Same context size, same workload, no timeouts.