[BUG?] Multiple requests to different models per 1 simple "Hello" message

[KernelBypass]

1. I've just setup codex-lb with several paid business accounts, and it works, but it's weird - for 1 simple "hi" chat message sent to gpt-5.5-xhigh, I'm seeing 4 requests with different models (gpt-5.5-xhigh and gpt-5.4-mini-low) and efforts (xhigh and low) - see the screenshot:

Is this a normal behavior? As far as I can tell, there should have not been any tool calls.
The two 5.5-xhigh requests were both pretty big too.

2. Is it possible to examine the payloads from within the logs tab of the web interface?

[KernelBypass]

I noticed when switching to supports_websockets = false in config.toml, which effectively switches request from WS to HTTP mode, the number of requests in the log tab of codex-lb goes from 4 per 1 turn (in WS mode) to 2 per one turn (HTTP mode).
So, WS mode is 4 per turn, http is 2 per turn.

I used mitmweb to intercept and examine the traffic. For http mode it's 1 main prompt and 1 chat summary (that's what gpt-5.4-mini was).

Interestingly, for WS mode, mitmweb also shows only 2 requests. Not sure why codex-lb is showing 4. Could it be receiving them as WS and then downgrading to HTTP before relaying to OpenAI and displaying both the original and the downgrade as 4 total?

I wish I could examine the payloads from within codex-lb. If there is a way, it's not obvious to me. The UI seems to not support it.

[KernelBypass]

This corresponds to the last 3 turns in the screenshots above:

• 1 turn (2 http calls)
• 1 turn (2 http calls)
• 1 turn in WS mode (2 calls in mitmweb, yet 4 calls in codex-lb log section)

[KernelBypass]

Probably has nothing with Codex-lb, but I'm also still curious why GPT-5.4-mini (low) gets called on old, already-named chats. Maybe it's memory summarization.