This seems like good feedback. Curious if anyone on the agent / SDK team has any thoughts here?

opened github.com/OpenHands/software-agent-sdk/pull/4012 but @modern-fireman-99755 also opened a PR! should be fixed very soon

Thank you Simon and Vasco!

Hey @strong-ice-84145 I was looking inside the two PRs. What I found: The two PRs chase the same symptom ("agent-server freezes while waiting on an LLM") but fix two different locks: - PR 4012 (simonrosenberg): the process-wide litellm.modify_params lock. A blocking with around await freezes the whole event loop when a sync call (e.g. the condenser) holds it. This is the actual mechanism behind the reported "freezes during condensation / local inference". - PR 4009 (yours): the per-conversation state FIFOLock held across astep. It blocks other ops on the same conversation (send_message, WS snapshots) during a model turn — a real but different problem. It does not fix the reported condensation freeze. So: not duplicates — complementary. They don't overlap in code. What to do 1. Land 4012 first — it directly fixes the reported freeze, small and low-risk (1 file). Flag two nits: possible lock leak if the coroutine is cancelled between run_in_executor acquire and the try, and confirm nothing re-enters after RLock→Lock. 2. Keep 4009, reframed as "allow message interjection / snapshots during a model turn" — a separate improvement, not a competitor. It's the better-tested PR but more invasive (new FIFOLock hand-off primitive). 3. Don't close either as a dup of the other.

fixed!

Accepted 🙂 perhaps trim the docstring and comments because are huge 😅 otherwise LGTM

The PR review bot has a section just to fight those opus 4.8 huge comments 😕 too bad we aren't using it

No problem. It was a nit ;)