REPLAY AND DEBUGGING

ARTIFICIAL INTELLIGENCE ASSISTANT PROMPTS [START OF PROMPT] Review this recorded session/replay data and tell me whether everything needed for a faithful offline replay was actually captured (tool responses, random seed/temperature, timestamps): [PASTE LAB REPLAY DATA / RECORDED SESSION HERE]. Flag anything missing that would make the replay diverge from the original run, and specify the exact recording change needed. Output format: a completeness checklist with pass/fail per item and fixes for any fail. [END OF PROMPT] [START OF PROMPT] Run a root-cause analysis on this failure trace and determine whether it is a network/tool problem or a model reasoning problem, citing the specific evidence: [PASTE LAB OR PRODUCTION FAILURE TRACE HERE]. Output format: Root Cause Classification (one line), Supporting Evidence (bullets), Recommended Fix (one paragraph), and a suggested new eval/regression test to catch recurrence. [END OF PROMPT] CHAPTER SUMMARY Replay re-runs a recorded session offline to study a failure. Faithful replay needs recorded messages, tool responses, timestamps, and model settings. Root-cause analysis looks for the underlying reason, which may be a stale data source. Telling tool problems from reasoning problems decides which fix to apply.