Remaining tasks

• Do a msbench run, compare with completions, check results of different thinking levels
• What stats to look at before rolling out?
• Allocate context window space for reasoning
• Test fix for summaries returned from endpoint