← ClaudeAtlas

bug-hunterlisted

Audit an existing codebase for verified functional, logic, runtime, concurrency, data-integrity, contract, and security bugs without fixing them. Use only when the user explicitly invokes $bug-hunter for adversarial bug finding in a repository, path, staged change, or branch diff. Not for ordinary code review, style issues, missing tests, speculative hardening, implementation, auto-fix, or remediation work.
martonpaulo/skills · ★ 0 · AI & Automation · score 72
Install: claude install-skill martonpaulo/skills
# Bug Hunter Find behavioral bugs that survive an adversarial evidence pass. Report them in the conversation; never edit production code, tests, configuration, or repository documentation. ## Resolve the target - With a path argument, audit that file or directory. With no argument, use the current repository. - `--staged` audits the full current contents of staged source files, using the staged diff only to select scope. - `--branch <base>` audits source files changed from the merge base with `<base>`, while reading unchanged callers, callees, tests, and configuration when required to verify behavior. - Apply repository guidance and exclude generated output, vendored dependencies, minified assets, and irrelevant documentation. Read tests as intent and coverage evidence; do not report test-only defects unless the test code ships in the product or the user explicitly includes test infrastructure. If the target is too large for credible full coverage, audit complete risk-bounded slices in this order: external/trust boundaries, authorization and sensitive state, persistence and transactions, concurrency and retries, error boundaries, then remaining business logic. State the exact queue and unscanned scope; never claim full coverage from a sample. ## Workflow 1. Inspect repository guidance, worktree status, architecture, tests, recent relevant history, and the resolved target without changing anything. 2. Map entry points, trust boundaries, state transitions, data ownership,