The above is written by OH-agent, my “secretary”, who peaked at all trajectory via API, and • found Sonnet-4 in the newest toolcall metadata • and GPT-5 in all previous toolcall metadata. I set it to check trajectory it when I actually saw myself that the model is Sonnet-4 and felt that’s weird, because I was sure the model in that conversation used to be GPT-5.

The agent committed the .json and some summary it made

Does the Cloud code actually force the model assignment to persist? OpenHands doesn’t.

But that doesn’t make it less weird, because … I don’t think I changed anything, and it’s the Cloud code that changes / chooses models 🤔

In the commit above, you may want to see from this line (“model switching”) https://github.com/enyst/playground/pull/94/commits/3cb81e39af2d6da90cd4a3f074a6409eb9b54fd4#diff-6c9cc3cd5599e7bb6d76fe07261354ef7c9fb59abed3ec912eb200b66aee7b36R195

Another conversation where the model was switched: 4c2948df77424deaa23b6a09ada34259

In this case, it was 2h old, and FWIW I don't think the runtime stopped completely in this time

I didn't ask my secretary about this one, but it's obvious between Sonnet and GPT-5

Actually, I made a new secretary - ended up with GPT-5, who says:

• Full trajectory: I downloaded all events and saved them to cache/trajectories/conversation_4c2948df77424deaa23b6a09ada34259_events.json for offline analysis.

• Model findings:

• Unique models observed:

▪︎ litellm_proxy/claude-sonnet-4-20250514 (first seen at event 9, last at 383)

◦ gpt-5-2025-08-07 (first seen at 392, last at 434)

◦ Prefix discrepancy in this conversation: Claude shows the litellm_proxy/ prefix, GPT-5 does not.

• Comparison to prior model-switch note in the MD (Conversation #23):

◦ There, both models were recorded with the litellm_proxy/ prefix (e.g., litellm_proxy/gpt-5-2025-08-07 and litellm_proxy/claude-sonnet-4-20250514).

◦ So yes: Previously both had the prefix; in the new conversation only Claude has the prefix while GPT-5 appears without it.