Your codebase is on trial. Eight AI experts are the judges.
In this court, your project is guilty until proven shippable.
One command convenes a full review tribunal — Project Manager, Product Manager, Architect, Backend Engineer, Frontend/UI Engineer, Security Engineer, QA Engineer, DevOps Engineer. Each judge inspects your project through their own professional lens, independently, without seeing each other's notes. Then the court hands down:
- ⚖️ A quantified verdict — per-role scores, plus an overall score under sentencing guidelines: any Critical finding caps the project at 7/10; three Criticals cap it at 5. A beautiful average can't save you.
- 📋 A rap sheet with evidence — every finding cites
path:line, severity (🔴 Critical / 🟡 Warning / 🟢 Suggestion), and why it matters. Dynamic findings are verified by actually running your app locally (实测= tested, not guessed). No "the code could be better" hand-waving. - 🔧 A court-ordered fix plan (
MASTER_FIX_PLAN.md) — every task carries the exact file, line number, current code snippet, expected change, and a copy-pasteable verification command. Hand it to Claude Code / Cursor / your coding agent and walk away.
Field-tested, not vibes. On a real flu-prediction platform, a manual 5-role sequential pass found 16 issues. The same project before this tribunal: 31 issues — including unauthenticated delete endpoints, path traversal, and a CVE in the auth library that nobody had noticed. Parallel judges with independent context windows surface what a single-pass review misses. All 44 finding patterns in this repo's precedent library were found in the wild, not invented.
| Judge | Asks | Catches |
|---|---|---|
| 🎯 Project Manager | "Is this shippable? What's blocking?" | Mock data in deliverables, missing deployment artifacts, unpinned deps |
| 📋 Product Manager | "Does this solve the user's real problem?" | Dead-end user flows, missing feedback loops, placeholder values in user-facing output |
| 🏗️ Architect | "Will this structure hold up?" | God files, wrong dependency direction, circular coupling |
| ⚙️ Backend Engineer | "What breaks at 3am?" | Zero-auth endpoints, silent exception swallowing, sync I/O inside async def |
| 🎨 Frontend/UI Engineer | "Would I be proud to show this?" | No loading/error/empty states, 100+ hardcoded hex colors, memory-leaking resize handlers |
| 🛡️ Security Engineer | "Where is the trust boundary that wasn't checked?" | Weak JWT secrets, SQL string concatenation, CSRF not initialized, secrets committed to git |
| 🧪 QA Engineer | "Is there evidence this works, or just developer optimism?" | Zero test suites, stale tests, "tests" that can't even run in CI |
| 🚀 DevOps Engineer | "How does this reach production? How would I know it failed?" | Root-user Dockerfiles, non-existent base-image versions, placeholder systemd paths |
The tribunal adapts to the project: data pipelines get a Data Engineer instead of Frontend; ESP32 firmware gets a Firmware Engineer + Companion App Engineer (blocking delay() in loop(), WiFi passwords baked into public binaries, BLE MTU traps...).
- Bailiff work — stack detection: manifest, framework, lockfile, package manager, deployment artifacts. Never assume the stack.
- Jury selection — roles chosen by stack and project maturity (prototype → 3 judges; production platform → full 8).
- Independent deliberation — each judge reads the code through their own lens (ideally as parallel sub-agents with separate context windows) and returns a signed Output Contract: score + findings + health assessment.
- Evidence testing — Critical security claims are verified by actually running the service locally and probing: no-credential request → 200? confirmed. 401? finding refuted as false positive.
- Synthesis — findings de-duplicated across judges (
[roles: backend, security]= higher confidence), conflicts surfaced openly with a tie-break, gaps noted. - The verdict — overall score computed mechanically, caps applied, arithmetic shown:
总分 5.0 = min(均值 6.1, 3+ Critical 封顶 5) - Sentencing —
REVIEW_REPORT.md(the audit trail) +MASTER_FIX_PLAN.md(the executable sentence), cross-mapped so every finding traces to a task.
| Situation | Score cap |
|---|---|
| Any Critical finding stands | ≤ 7 |
| 3+ Critical findings | ≤ 5 |
| Any single judge scores ≤ 3 | ≤ 6 |
| Core business logic absent (scaffold only) | ≤ 4 |
Why caps exist — a real case: a production platform averaged 5.9 across roles, but carried 60 completely unauthenticated endpoints. The first review shipped the 5.9 anyway. The re-review recomputed with caps: 5.0. One severely deficient area is a systemic risk; the average must not launder it.
| Mode | Command vibe | What happens |
|---|---|---|
| Full Trial (default) | "review this project" / "审查这个项目" | Whole-project audit, all judges |
| Spot Check | "review this PR/diff" | Same lenses, scoped to changes |
| Appeal (delta re-review) | "re-review the project" | The parole hearing 🎫 — every previous finding re-verified line-by-line (✅ 已修 / |
references/findings-patterns.md — 44 patterns distilled from 8 real project reviews, each with the grep command that finds it and the fix pattern that resolves it. A taste:
- #1 Zero-auth business endpoints — appeared in every reviewed platform.
grep -rn "@router.post\|@router.delete" | grep -v get_current_user - #28 Non-existent dependency versions —
pandas==3.0.2doesn't exist; Docker build fails 100%. (Hallucinated versions from AI coding assistants are now a standard check.) - #33 Written-but-never-wired modules — a fully implemented evaluator that nothing imports, while the README claims the feature. Dead code disguised as features.
- #36 NaN silently treated as zero — pandas
.sum(skipna=True)on surveillance data where NaN means "not tested", not "zero". Corrupts every downstream number without an error message. - #43 Generated artifacts landing in a public static dir — each half looks fine in isolation; only the overlap is the vulnerability.
Works with any agent that reads SKILL.md (Claude Code, ZCode, and compatible runners). Best experience with sub-agent support (parallel judges); falls back to sequential personas automatically.
# Claude Code (user-level)
git clone https://github.com/MUST-panxiao/codebase-tribunal.git ~/.claude/skills/codebase-tribunal
# or generic agents convention
git clone https://github.com/MUST-panxiao/codebase-tribunal.git ~/.agents/skills/codebase-tribunal
# project-level (any agent): clone into .claude/skills/ or .github/skills/ in your repoThen open your project and say "review this project" or “审查这个项目”. That's it — the tribunal self-organizes.
👉 examples/sample-verdict.md — a condensed, anonymized real-format verdict (score table, cap computation, findings with evidence, fix-plan excerpt).
代码审判庭：一句话把你的项目送上法庭。5–8 位 AI 专家（项目经理、产品、架构、后端、前端、安全、测试、运维）各自独立审查同一个项目，互不通气，避免视角趋同；随后交叉验证、去重合议，产出：
- 量化评分（含封顶规则：存在任一 Critical 直接封顶 7 分）
- 每条发现都带
文件:行号证据和影响说明，安全类发现会本地起服务实测 MASTER_FIX_PLAN.md修复计划——精确到行号、附当前代码和验收命令，可直接交给 Claude Code 执行
支持全量审查、PR 增量审查、以及复审模式（旧问题逐项核对：已修/部分/未修/失效，并列出本轮新增）。内置 44 条真实项目沉淀的发现模式（零鉴权接口、路径穿越、NaN 当 0、幻觉依赖版本、写好但从未接线的模块……）。审查可用中文或英文进行。
Distilled from 8 real project reviews — delivery-stage platforms, a 20+-module production system, a copy-paste product family, server-rendered Flask apps, an admin-template prototype, a WHO-data forecasting pipeline, and an ESP32 smart-watch firmware. Project names anonymized (Projects A–H); every scar, cap rule, and pattern was earned in the field. Forged in real Hermes Agent and ZCode review sessions, now public for any agent to wield.
MIT © 2026 panxiao