Agent Skills: Critical Bug Finder

IRON LAW: NEVER REPORT A BUG WITHOUT A CONCRETE TRIGGER, A REAL FAILURE MODE, AND A SEVERE CONSEQUENCE.

Critical Bug Finder

Use $ARGUMENTS and the user's wording to define scope before scanning.

Options

• <path>: file or directory to audit
• --full: audit the full repository
• --quick: skip confirmation gates when scope and output mode are explicit
• --security-only: only report security vulnerabilities
• --concurrency-only: only report races, deadlocks, or TOCTOU bugs
• --format json|md: output format, default md; if json, load references/output-format.md
• --write-report: write a report file after presenting findings
• --max-findings N: cap reported findings, default 10

Copy this checklist and check off items as you complete them:

• [ ] Step 0: Define scope [BLOCKING]
  • [ ] Parse $ARGUMENTS
  • [ ] If <path> is provided, scope to it
  • [ ] If --full is provided, scan the full repo
  • [ ] If scope is unclear and --full is absent, ask one focused question before scanning
• [ ] Step 1: Build the risk map [BLOCKING]
  • [ ] Identify high-risk files first: input boundaries, state mutations, DB or file writes, auth checks, concurrency, error handling
  • [ ] For large repos, split one top-level directory or subsystem per subagent, starting with the riskiest areas first
• [ ] Step 2: Audit [REQUIRED]
  • [ ] Load references/checklist.md
  • [ ] Ask: can bad input crash this path, corrupt state, bypass auth, exhaust resources, or create a race
  • [ ] Ask: can two requests or two processes break assumptions here
  • [ ] Ask: does the code mutate state without atomicity, rollback, or ordering guarantees
• [ ] Step 3: Validate candidates [REQUIRED]
  • [ ] Keep only findings with a concrete trigger scenario
  • [ ] Keep only findings with a real severe consequence
  • [ ] Keep only findings with a specific code location
  • [ ] Discard theoretical, stylistic, or low-impact issues
• [ ] Step 4: Present findings [REQUIRED]
  • [ ] Report the highest-confidence findings first
  • [ ] Respect --max-findings N
  • [ ] State audited scope and anything not checked
• [ ] Step 5: Write report (conditional)
  • [ ] Only if the user requested file output or passed --write-report
  • [ ] Unless --quick was passed with explicit output mode, ask before writing files

Qualifying categories

• Crash, panic, uncaught exception, or unrecoverable runtime failure
• Data corruption, data loss, or silently wrong results
• Security vulnerability such as SQL injection, XSS, RCE, SSRF, path traversal, auth bypass, or secret exposure
• Infinite loop, deadlock, resource exhaustion, or unbounded growth that can take the system down
• Race condition or TOCTOU bug with a realistic failure path
• Core logic contradiction that breaks documented or clearly intended behavior

Exclusions

• Architecture opinions, refactors, naming, readability, or style
• Performance complaints unless they realistically cause outage, OOM, or resource exhaustion
• Missing tests, docs, or logging by themselves
• "Could be risky" claims without a concrete exploit or failure path

Output

If --format json, load references/output-format.md and follow its schema exactly.

For each finding, include:

• severity: CRITICAL or HIGH
• location: file:line or exact symbol
• category: one qualifying category
• description: one sentence
• trigger: concrete scenario
• why_fatal: why the impact is severe
• suggested_fix: minimal fix direction

If no qualifying bugs are found, say so clearly and list the audited scope.

Report file

If writing a file, save it in the project root as bugs_<4-char>.md.

Typical downstream: use bugfix-dispatch when the user wants to split confirmed findings into PR-sized worktrees.

Do not

• Do not report generic code review issues
• Do not report speculative vulnerabilities without a trigger path
• Do not inflate severity for medium or low issues
• Do not write a report file unless the user asked for it or passed --write-report
• Do not claim a full-repo audit if the scan was partial