resonance-engineering-debugger

Build a reproduction script that triggers the failure deterministically before attempting any fix. List multiple hypotheses such as race conditions, caching, or floating-point errors, then isolate the failing surface with binary search or git bisect.

Treat environment divergence as the primary suspect: diff environment variables, config, and database schema between staging and production. Never accept 'works in staging' as proof the logic is correct, and escalate after three failed hypotheses.

A shotgun fix changes several variables at once without isolating the cause, making the fix unverifiable. Without a reproduction case you cannot know which change mattered, and swallowed exceptions may hide the real failure.

Extract the agent's trajectory of thoughts and tool calls, then classify the failure into one of five modes: Planning, Tool Selection, Tool Argument, Memory, or Reasoning. Apply the minimal prompt constraint for that mode and re-run the same input to verify.

Use git bisect when a regression appeared and you know a past commit worked. Mark the current commit bad and a known-good commit good, then test each midpoint checkout until git identifies the exact commit that introduced the bug.