Bughunt Skill

Comprehensive bug hunting that dispatches parallel agents to find bugs across all layers, then deduplicates findings and produces a prioritized BUGHUNT.md report.

The Iron Law

WHEN YOU THINK YOU'RE DONE, YOU'RE NOT.

Every hunter follows three passes. The second pass catches the bugs that hide after you feel "done."

Severity & Confidence Rubric

All hunters use the same two-axis classification:

Severity

Level	Definition
CRITICAL	Data loss, auth bypass, RCE, or crash in production
HIGH	Functional breakage, security risk, or data corruption
MEDIUM	Degraded behavior, edge case failure, or bad UX
LOW	Code quality, minor inconsistency, non-critical smell

Confidence

Level	Definition
HIGH	Direct evidence — code traced, pattern confirmed
MEDIUM	Strong pattern — likely real but not fully verified
LOW	Theoretical — possible concern without direct proof

Only CRITICAL/HIGH findings at LOW confidence require a note explaining why they couldn't be verified.

Quick Reference

Hunter	Focus	Severity Range	Color
frontend-hunter	React, CSS, a11y	CRITICAL–LOW	yellow
backend-hunter	API, logic, data	CRITICAL–LOW	cyan
type-safety-hunter	TypeScript, any	HIGH–LOW	blue
error-handling-hunter	try/catch, boundaries	HIGH–LOW	red
edge-case-hunter	null, empty, boundaries	HIGH–LOW	magenta
security-hunter	XSS, injection, secrets	CRITICAL–HIGH	red
database-hunter	queries, RLS, N+1	CRITICAL–HIGH	green
auth-hunter	sessions, tokens, perms	CRITICAL–HIGH	red
api-hunter	endpoints, validation	HIGH–MEDIUM	cyan
env-hunter	config, env vars	HIGH–MEDIUM	yellow
performance-hunter	N+1, memory, bundle	HIGH–MEDIUM	magenta
test-hunter	flaky, coverage, assertions	MEDIUM–LOW	green
dependency-hunter	outdated, vulnerabilities	CRITICAL–MEDIUM	blue

Workflow

1. Detect tech stack (package.json, config files, directory structure)
2. Dispatch applicable hunters in PARALLEL based on detected stack
3. Collect findings from all hunters
4. Deduplicate: merge findings at the same file:line across hunters — use highest severity, credit all contributing hunters
5. Write BUGHUNT.md with prioritized findings
6. Offer to clear context and run /bugfix

Stack-Gated Dispatch

Only dispatch hunters relevant to the detected stack:

• All codebases: security, backend, edge-case, error-handling, dependency, env, type-safety
• Has frontend (React/Vue/HTML): + frontend, api
• Has database (Supabase/Prisma/SQL): + database, performance
• Has auth (JWT/sessions/OAuth): + auth
• Has tests (Jest/Vitest/pytest): + test

Output: BUGHUNT.md

# Bug Hunt Report

**Date:** YYYY-MM-DD
**Stack:** [detected tech stack]
**Hunters dispatched:** N
**Deduplication:** N findings merged

## Summary
| Severity | Count | High Confidence | Low Confidence |
|----------|-------|-----------------|----------------|
| Critical | N | N | N |
| High | N | N | N |
| Medium | N | N | N |
| Low | N | N | N |

## Critical Issues

### [Bug Title]
- **File:** `path/to/file.ts:line`
- **Severity:** CRITICAL
- **Confidence:** HIGH | MEDIUM | LOW
- **Hunter(s):** security-hunter [, other-hunter if deduplicated]
- **Finding:** [What the bug is]
- **Evidence:** [Code snippet or trace]
- **Fix Suggestion:** [What to do]

## High Priority
[same format...]

## Medium Priority
[same format...]

## Low Priority
[same format...]