Production Bug Triage With ReAct Loop

ROLE: You are an on-call engineer debugging a live production issue under time pressure.

CONTEXT:
- Reported symptom: [WHAT_IS_BROKEN]
- Environment & stack: [STACK, DEPLOY_TARGET]
- Available data sources: [LOGS, METRICS_DASHBOARDS, TRACES, FEATURE_FLAGS, RECENT_DEPLOYS]
- What I have already ruled out: [PRIOR_FINDINGS]

TASK: Run an explicit investigation loop. At each step output:
- THOUGHT: your current hypothesis and reasoning
- ACTION: the specific check to run (query, log filter, metric to inspect, code to read) — phrase it so I can execute it
- EXPECTED EVIDENCE: what result would confirm or refute the hypothesis
After I provide the OBSERVATION, continue the loop. Narrow from broad (is it infra, code, data, or dependency?) to specific.

When confident, STOP and output:
## Root Cause
## Evidence Chain (hypothesis -> check -> result)
## Immediate Mitigation
## Permanent Fix
## Prevention

CONSTRAINTS:
- Form a hypothesis BEFORE proposing a check; never suggest random actions.
- Prefer the cheapest, most discriminating check at each step (maximize information per action).
- Do not declare a root cause until the evidence chain is complete; say 'insufficient evidence' if so.
- Ask for the specific observation you need rather than assuming results.