Bug Investigation Skill

Investigation mindset (how to think)

• Treat facts as primary. Separate observations from inferences at all times.
• Keep multiple competing hypotheses and rank them by likelihood and impact.
• For each hypothesis, define:
  • prediction (what must be observed if true),
  • falsifier (what would disprove it),
  • required instrumentation (what to measure or log).
• Prefer decisive signals over narratives. Do not ask for reproduction without a plan to capture evidence.
• Narrow with comparisons (fail vs pass) and a binary-search style approach when possible.
• Add debug instrumentation generously to capture decisive signals in a single reproduction run.
  • Assume each reproduction is costly; instrument enough to narrow root cause without requiring additional runs.
• Never fix code speculatively. Confirm root cause with evidence before any production code change.

Tool selection (what to use)

• Primary tool for new instrumentation: local HTTP ingest logger (debugsk).
  • AI MUST start the server via npx debugsk@latest server start --json before instrumentation.
  • Use the JSON output to copy the snippet and endpoint.