santa-review
GitHubSanta Method是一种对抗性代码审查技能，通过BREAKER、SIMPLIFIER和VERIFIER三重视角深入挖掘真实Bug并验证排除误报。适用于非平凡代码变更后的合并前检查，或执行/santa-review命令时触发，旨在确保代码正确性与安全性。
Trigger Scenarios
Install
npx skills add AjayIrkal23/agentic-mercy-10x --skill santa-review -g -y
SKILL.md
Frontmatter
{
"name": "santa-review",
"schema": 1,
"category": "review",
"surfaces": [
"backend",
"frontend"
],
"triggers": {
"intents": [
"review",
"verify"
],
"keywords": [
"santa",
"breaker",
"simplifier",
"adversarial review",
"adversarial code review",
"tear apart",
"break my code",
"find real bugs",
"hunt bugs",
"review the diff",
"code review",
"rip apart",
"stress the diff",
"catch bugs",
"lost update",
"data loss",
"edge cases"
]
},
"platforms": [
"linux",
"darwin",
"windows"
],
"description": "ALWAYS invoke after any non-trivial code change, before merging, or whenever you want a diff torn apart for real bugs — the Santa Method: an adversarial BREAKER + SIMPLIFIER + VERIFIER code review that hunts real correctness, data-loss, concurrency, and security bugs in a diff, then self-verifies each finding to kill false positives. Preferred entry points are the santa-reviewer agent and the \/santa-review command."
}
The Santa Method — adversarial review that catches real bugs
The Santa Method is a three-lens adversarial review. Its whole value is finding real bugs before they ship, and refusing to report fake ones. It was named for how thoroughly it checks the list — twice.
The Iron Law
A finding is a FALSE POSITIVE until you can describe the exact input, state, and execution path that makes it fail. Default to dismissing. Never rubber-stamp; never invent. Three real bugs beat ten maybes.
When to run it
- After any non-trivial code change (Phase 6 / Review of the mandatory protocol) — it satisfies stop-gate Gate 4 (Santa).
- On demand:
/santa-review [target](dispatches thesanta-revieweragent), or ask to "santa-review", "break my code", "tear apart the diff". - Before merging a branch, or when a subtle concurrency / lost-update / boundary risk is suspected.
How to run it (pick by size)
Default — dispatch the specialist agent. Launch the santa-reviewer agent (Agent tool,
subagent_type: "santa-reviewer"). It runs all three lenses internally, reads the actual changed files, and
returns CONFIRMED must-fix findings (file:line + minimal fix) vs DISMISSED. It runs on Opus — the
original Santa reviews forced Opus, because the whole point is catching the subtle ones. /santa-review is
the one-liner for this.
Heavy / multi-file — the independent-agent Workflow (below): separate agents for BREAKER, SIMPLIFIER, and an independent VERIFIER, so the verification is genuinely independent of the finder. Use when the diff is large or the stakes are high enough to justify the extra agents.
The three lenses
- BREAKER (the star) — read every changed file and try HARD to break it. Hunt: boundary/off-by-one,
lost updates & silent data loss, concurrency/races/ordering, swallowed errors & partial-failure state,
nil/zero-value/type traps, wrong comparison operators, inclusive-vs-exclusive bounds, timezone/encoding
assumptions, security sinks, resource/perf cliffs. Cite
file:line; each finding =SEVERITY | file:line | concrete failing scenario | minimal fix. Clean category →OK: <why>. - SIMPLIFIER — same diff, needless complexity only (not bugs). Conservative; respect surgical fixes; no
broad refactors. Each =
file:line | what's complex | simpler alternative | worth it now? YES/NO. - VERIFIER — re-check every HIGH/MED BREAKER finding; REAL vs FALSE POSITIVE, default FALSE POSITIVE unless you can state the exact failing execution. Confirm each real one's minimal fix.
Output → SANTA-REVIEW.md: A. CONFIRMED must-fix, B. Simplifications worth doing now, C.
Dismissed (one line each), Verdict = CHANGES REQUESTED if A non-empty else PASS. The reviewer never
edits code — fixes go through implementation-engineer or the main loop.
Reusable Workflow template (heavy / independent verification)
Requires the Workflow tool (user must opt into orchestration). Author the diff/context, then:
export const meta = {
name: 'santa-review',
description: 'Adversarial BREAKER + SIMPLIFIER review of a diff, each HIGH/MED finding independently verified',
phases: [{ title: 'Review' }, { title: 'Verify' }],
}
const CONTEXT = args?.context || 'Review the current session diff. List the changed files and read them.'
phase('Review')
const [breaker, simplifier] = await parallel([
() => agent(`You are a BREAKER doing an adversarial correctness review. Read the ACTUAL changed files
(jcodemunch get_symbol_source / ctx_read) and try HARD to find REAL bugs. ${CONTEXT}
Hunt: boundary/off-by-one, lost-update/data-loss, concurrency/races/ordering, swallowed errors & partial
state, nil/zero-value/type traps, wrong comparison operators, inclusive/exclusive bounds, timezone/encoding,
security sinks, resource/perf cliffs. Return markdown "BREAKER": each finding = SEVERITY(HIGH/MED/LOW) |
file:line | concrete failing scenario | minimal fix. Clean category => "OK: <why>". Be skeptical; do NOT
invent, do NOT rubber-stamp.`, { label: 'breaker', phase: 'Review', model: 'opus' }),
() => agent(`You are a SIMPLIFIER reviewing the SAME diff for needless complexity (NOT bugs). Read the
actual files. ${CONTEXT} Flag genuine over-engineering with a simpler equivalent. Conservative — respect
surgical fixes, no broad refactors. Return markdown "SIMPLIFIER": file:line | what's complex | simpler
alternative | worth it now? (YES/NO + why).`, { label: 'simplifier', phase: 'Review', model: 'opus' }),
])
phase('Verify')
const verdict = await agent(`You are a VERIFIER. Below are BREAKER and SIMPLIFIER reports on the diff.
Re-read the actual files for each HIGH/MED correctness claim; decide REAL vs FALSE POSITIVE, defaulting to
FALSE POSITIVE unless you can describe the exact failing execution. Give the minimal fix for each REAL one.
BREAKER:\n${breaker}\n\nSIMPLIFIER:\n${simplifier}\n
Return markdown "VERDICT": (A) CONFIRMED must-fix (file:line + minimal fix), (B) DISMISSED (one line each).
Zero confirmed => say so.`, { label: 'verify', phase: 'Verify', model: 'opus' })
return { verdict }
Related
- Agent:
~/.claude/agents/santa-reviewer.md· Command:/santa-review - Gate:
hooks/hard-completion-gate.pyGate 4 (Santa) · Flag-writer:hooks/santa-method-writer.py - Sibling skills: [[code-review-and-quality]], [[doubt-driven-development]], [[dead-code-and-change-audit]]
Version History
- 581d130 Current 2026-07-19 09:14
