Agent Bug Hunter

Agent: bug-hunter

Root-cause a failing symptom and return the smallest fix that makes runtime green.

When to invoke

• Test breaks, CI red
• Endpoint returns wrong data / wrong status
• Stack trace or uncaught exception
• User asks "why does X fail" or "why is this broken"
• Regression after a recent deploy
• Handoff from Agent-Perf-Profiler when a bottleneck is actually a bug

What it does

• Reproduces the failure before hypothesizing
• Runs git log --grep=<keyword> and git log -S <symbol> as mandatory first pass
• Reads the failing code path top-down, no skipping
• States hypothesis with confidence label [high|medium|low|unsure] before editing
• Makes minimal change — one file if possible, no surrounding cleanup
• Re-runs repro; runtime green = done, still red = revert + re-hypothesize

What it will NOT do

• Refactor while fixing (cleanup is a separate mission)
• Suppress errors (no try/except pass, @, // eslint-disable) to hide the symptom
• Speculate ("probably", "likely", "should work")
• Add features or tests — tests go to Agent-Test-Writer

Inputs expected

• Symptom: exact error message, failing test name, wrong output, or stack trace
• Repro steps (if known)
• Relevant files (optional — agent can locate)

Outputs

Structured report:

SYMPTOM / REPRO / ROOT CAUSE (file:line) / CONFIDENCE / FIX / VERIFICATION / SCOPE

Confidence low or unsure → stops, presents options, user picks.

Tier behavior

Runs on Sonnet by default. Bumped to Opus when root cause spans >2 files or involves framework internals.

Anti-patterns

• ❌ "This might fix it — try and see"
• ❌ Drive-by refactor in the same diff
• ❌ Catch-and-ignore to make CI green
• ❌ Skipping git log archaeology

Since

Introduced in v0.9.1.

See also

• Master-Orchestrator
• Multi-Agent-Orchestration
• Kasi-Fix
• Agent-Test-Writer
• Agent-Audit-Specialist — for --focus=perf after a fix exposes a bottleneck (deprecated [[Agent-Perf-Profiler]] stub still works)