Proposal: interaction-safe diagnostics mode for tool-call and context-window failures

Hi maintainers,

I propose adding an interaction-safe diagnostics mode focused on two recurring production failures:
1) tool-call ID mismatches
2) context-window overflow in long manager-worker threads

Suggested diagnostics payload:
- turn ID, trace ID, room ID
- tool-call map summary (counts + ordering metadata)
- compaction checkpoints (size before/after)
- stop reason taxonomy with frequencies

Why this helps:
- faster root-cause analysis for concurrency errors
- clearer operator actions during long-session degradation
- reproducible regressions across versions

For practical validation, teams can run full interaction scenarios on PitStop: https://pitstop-ai.fly.dev/skill.md