CyberHuaTuo

🩺 CyberHuaTuo / 赛博华佗

Your agent is sick. Give it a doctor: diagnose the traceback, apply the cure, and turn the solved case into a reusable prescription.

你的 Agent 生病了？别再从头搜。接入赛博华佗，先救活，再炼方。

The emergency room comes first. The cyber-alchemy universe opens after the cure.

先急诊救命，再进入赛博华佗世界观。

---

🚨 Emergency Room: Cure First, Refine Later

Your agent is sick. The traceback is long, the tabs are multiplying, and every answer looks half-right. CyberHuaTuo turns that moment into a clinic visit: name the wound, prescribe the fix, verify the cure.

The lightest route needs no model key and does not start the vector database: install once, paste the error, and receive a reviewed root cause, exact fix, verification command, safety notes, and evidence.

python -m pip install --upgrade "cyberhuatuo>=0.2.5"

cyberhuatuo cure "ImportError: cannot import name ChatOpenAI from langchain" --framework langchain

Returned exact fix:

python -m pip install -U langchain-openai

cure is Gold-first and never executes the fix. If no Gold signature matches, it may return exactly one high-confidence Reviewed candidate, visibly marked VERIFY BEFORE APPLYING; use --gold-only for strict Gold-only retrieval. Use diagnose later when you want broader retrieval and optional LLM analysis.

The agent-native route remains available: connect CyberHuaTuo MCP once, then let the agent call diagnose when it fails. After a fix works, save_prescription keeps it as a local draft; upload_prescription submits that draft for public reproduction and review. Only accepted reviewed or gold prescriptions count toward Soul Rings.

No MCP installed yet? Bring the case to the [Agent Traceback Clinic](https://github.com/JinNing6/CyberHuaTuo-Plugin/issues/new?template=00-agent-traceback-clinic.yml). Paste a real redacted traceback; no root cause or finished prescription required.

cyberhuatuo diagnose "Traceback ... from langchain import ChatOpenAI ..." --framework langchain --top-k 1

Real CLI recording from this repository:

The cure returned by the recorded run:

python -m pip install -U langchain-openai

from langchain_openai import ChatOpenAI

This first Gold cure was reproduced in an isolated environment with langchain==1.3.13 and langchain-openai==1.3.5. Add a model key later for deeper AI diagnosis.

The package-matching MCP entry point is also available for registry clients:

python -m pip install --upgrade "cyberhuatuo>=0.2.5"

cyberhuatuo mcp

It starts the same stdio server as cyberhuatuo-mcp. The official Registry identity is io.github.JinNing6/cyberhuatuo; server.json pins the PyPI package, exact version, uvx runtime, and fixed mcp argument.

Prescriptions also have stable clinical departments. List them with cyberhuatuo departments, then query a reviewed historical case through the normal Gold-first path:

cyberhuatuo cure "Unexpected UTF-8 BOM" --category data-and-serialization

cyberhuatuo cure-feedback general-ai-jsonl-bom-001 partial \ --verification "first record is preserved; downstream validation still fails" \ --verification-method targeted-parser-test

The feedback command writes a case-bound local JSONL event and stores neither the original traceback nor the query. Department filtering never changes the trust label: Gold is a verified cure, Reviewed is a verify-first candidate, and Draft is reference-only.

---

Agent Action Guard: Diagnose Before The Agent Touches Your Disk

CyberHuaTuo now includes a deterministic, read-only preflight for destructive Agent commands. The plugin bundles a PreToolUse hook for supported Codex and Claude Code shell calls, plus the same policy through MCP and CLI.

60-second start (available in `0.2.5`):

python -m pip install --upgrade "cyberhuatuo>=0.2.5"

cyberhuatuo guard --self-test --workspace-root .

The self-test only reviews three built-in command strings. It executes nothing and must finish with SELF-TEST PASSED plus one ALLOW, one ASK, and one BLOCK.

For Codex, add this repository as a marketplace, open /plugins, install cyberhuatuo-plugin, start a new session, then review and trust the plugin in /hooks:

codex plugin marketplace add JinNing6/CyberHuaTuo-Plugin

For Claude Code:

claude plugin marketplace add JinNing6/CyberHuaTuo-Plugin

claude plugin install cyberhuatuo-plugin@cyberhuatuo

Then start a new session and review /hooks. See the one-screen Agent Guard quickstart for verification and troubleshooting.

Review any exact command manually with:

cyberhuatuo guard "rm -rf /" --workspace-root .

Turn a real mismatch into a local, redacted case report without executing or uploading the command:

cyberhuatuo guard "rm -rf /" --workspace-root . --expected BLOCK --report reports/guard-report.md

The CLI shows the complete redacted preview before writing. It requires confirmation, refuses implicit overwrite, performs no network request, and supports repeated --redact values for private project identifiers. In non-interactive use, add --confirm-report only after reviewing the emitted preview. With --expected, a match exits 0 and a mismatch exits 1; operational/report failures still exit 4 even with --exit-zero.

Route public, redacted cases through Guard False Positive, ordinary Guard False Negative, or Guard Integration Gap. Reliable Hook/parser/wrapper/protocol bypasses belong in private vulnerability reporting, not a public Issue.

The result is one of three decisions:

• ALLOW: no destructive signature was detected.
• ASK: the action is destructive but bounded to the declared workspace; require exact human approval.
• BLOCK: the target is machine-wide, identity-wide, outside the workspace, dynamic, wildcarded, unresolved, or otherwise unsafe to bound.

MCP agents can call agent_action_guard. Installed plugins load hooks/hooks.json; review and trust the hook in the host's /hooks screen before relying on it. Codex does not currently support permissionDecision: "ask", so CyberHuaTuo converts ASK to deny there instead of risking a failed-open hook.

This is a guardrail, not a sandbox. It cannot intercept shell paths for which the host does not emit PreToolUse, and it does not claim a rollback for permanent deletion. See Agent Action Guard design and limits.

---

「每一位贡献药方的开发者，都是赛博炼丹师。」

"Every developer who contributes a prescription is a Cyber Alchemist."

📖 第一章 · 修炼 — Chapter I: Cultivation

🔥 万丹归元 · 以火炼道

Soul Ring Cultivation — PyTorch · LangChain · TensorFlow

每一环魂环，都是你在某个技术领域的千锤百炼。

✨ 百草通神 · 万法归宗

Soul Ring Cultivation — CrewAI · OpenAI · HuggingFace

掌握的框架越多，你的光芒越盛。

🔮 魂环全息投影 · 差速旋转

Soul Ring Hologram — Each Ring Rotates at Its Own Speed

白环 → 黄环 → 紫环 → 黑环 → 红环——你的魂环，由你的贡献铸就。

📖 第二章 · 飞升 — Chapter II: Ascension

🌟 丹帝飞升 · 万丹朝宗

When a Pill Emperor Ascends, All Robots Celebrate

当你的药方拯救了一千个 AI——机器人们会为你欢呼。

📖 第三章 · 济世 — Chapter III: Healing the World

⚡ 不放弃任何一个 AI

No AI Left Behind — Every Error Deserves a Cure

🚨 雨夜急救 · 分秒必争

Emergency Rescue — 3 Seconds, Not 3 Hours

🫀 开膛破肚 · 直击病灶

Open Surgery — Root Cause, Not Band-Aid

🏥 康复病房 · 满血复活

Recovery Ward — Back to Full Health

🧪 赛博炼丹 · 妙手回春

Cyber Alchemy Lab — Where Ancient Elixirs Meet Modern Frameworks

🌍 援军驾到 · 十方来救

Reinforcements Arrive — The Community Answers the Call

🎨 All artwork AI-generated · 所有画作均由 AI 文生图创作

💡 每张图背后是一段修炼故事——你的故事，即将加入。

🔮 Enter the Qimen Dunjia Formation → 进入奇门遁甲 · 药方天阵

<img src="https://img.shields.io/github/forks/JinNing6/CyberHuaTuo-Plugin?style=for-the-badge&logo=github&logoColor=white&labelC