Verify whether your Claude Code endpoint is really serving GLM-5.2 — using tokenizer fingerprinting, reasoning-token checks, config inspection, and an optional 1M context probe.
verify-glm is not another LLM eval framework. It verifies whether the model behind your Claude Code / LLM gateway is actually the model you think it is.
verify-glm 不是又一个模型评测框架，它验证的是：你接入的 Claude Code / LLM 网关背后，到底是不是真的那个模型。
verify-glm is a Claude Code skill + standalone Python CLI for developers who route Claude Code through a GLM-compatible / OpenAI-compatible gateway and want evidence that the backend model is actually GLM-5.2, not a silently substituted or downgraded model.
If you configure Claude Code with ANTHROPIC_MODEL=glm-5.2, you still might not be talking to GLM-5.2.
A gateway can:
- echo
"model": "glm-5.2"while routing to a cheaper model - list
glm-5.2in/v1/modelswhile serving GLM-4 or another backend - hide routing changes behind an OpenAI-compatible or Claude-compatible API
- pass a simple "who are you?" prompt because the model follows the system prompt
You need out-of-band evidence. verify-glm gives you that evidence in a reproducible CLI report.
- Claude Code users routing through a GLM-5.2 endpoint
- Developers using OpenAI-compatible or Claude-compatible LLM gateways
- Teams evaluating GLM-5.2 for AI coding workflows
- Engineers who need to verify tokenizer behavior, reasoning-token metadata, or long-context support
- Anyone asking: "Is this endpoint really GLM-5.2?"
| # | Check | Why it matters |
|---|---|---|
| 1 | Config inspection | Verifies local Claude Code config / env actually points at GLM |
| 2 | API echo | Confirms the endpoint accepts requests and returns the expected model name |
| 3 | Tokenizer fingerprint | Fixed strings produce model-family-specific token counts. ChatGLM-style tokenization is visibly different from tiktoken / Qwen / Llama families. |
| 4 | reasoning_tokens field |
GLM-4.5+ exposes completion_tokens_details.reasoning_tokens in usage metadata. |
| 5 optional | 1M context probe | Sends a near-1M-token prompt to test whether long-context support is real. |
Passing checks 1-4 means the endpoint is likely GLM-family. Passing the optional long-context probe gives stronger evidence for GLM-5.2-class context support.
verify-glm — checking active endpoint
base_url: https://<your-glm-gateway>/v1
model: glm-5.2
[1/4] Config check
✅ Configured for GLM
[2/4] API echo
✅ API echoed model=glm-5.2
[3/4] Tokenizer fingerprint
✅ CN char-level (ratio=0.92), EN subword (7 tokens) — matches ChatGLM family
[4/4] reasoning_tokens field
✅ reasoning_tokens=0 (GLM-4.5+ feature)
✅ Verified GLM family (4/4). Model claims to be: glm-5.2
With --full, the tool also runs the long-context probe:
[5/5] 1M context probe
✅ accepted ~900K tokens
Verdict: likely GLM-5.2-class endpoint
After installation, ask Claude Code:
verify I'm running GLM-5.2
Claude Code will discover the verify-glm skill and run the verifier.
git clone https://github.com/xiaoyanfei-tech/verify-glm.git
cd verify-glm
# Linux / macOS / Git Bash
./install.sh
# Windows PowerShell
.\install.ps1Or copy skills/verify-glm/ into ~/.claude/skills/ manually.
# Default 4 checks: fast, low token usage
python ~/.claude/skills/verify-glm/verify_glm.py
# Add 1M context probe: expensive, use only when needed
python ~/.claude/skills/verify-glm/verify_glm.py --full
# Machine-readable output
python ~/.claude/skills/verify-glm/verify_glm.py --jsonIf this tool helps you verify your setup, you can add a badge to your project README:
[](https://github.com/xiaoyanfei-tech/verify-glm)Rendered badge:
| Failure | What it likely means |
|---|---|
| Config ❌ | ANTHROPIC_BASE_URL or ANTHROPIC_MODEL is missing / misconfigured |
| Echo ❌ HTTP 401/403 | API token is wrong, expired, or blocked |
| Echo ❌ HTTP 400 | Model name not recognized by the gateway |
| Tokenizer ❌ looks like tiktoken | Gateway may be routing to Claude / GPT-style backend while echoing GLM |
| Tokenizer ❌ CN ratio too low | Tokenizer likely belongs to another model family such as Qwen / Llama / custom |
reasoning_tokens ❌ |
Could be pre-4.5 GLM, non-GLM, or gateway stripping usage details |
--full ❌ context error |
Context window is likely below GLM-5.2-class long-context capacity |
The script reads credentials and endpoint config from:
- Environment variables:
ANTHROPIC_BASE_URL,ANTHROPIC_MODEL,ANTHROPIC_AUTH_TOKENorANTHROPIC_API_KEY ~/.claude/settings.json→env./.claude/settings.json→env
If you're setting up Claude Code with GLM-5.2 in a corporate, proxy-heavy, or custom-gateway environment, this tool can help diagnose common routing and model-verification issues.
For setup help, endpoint verification reports, or enterprise AI coding environment consulting, open a GitHub issue with your use case.
- This is evidence-based verification, not a cryptographic proof of model identity.
- Tokenizer fingerprinting identifies the model family; the optional long-context probe gives stronger GLM-5.2-class evidence.
- Network errors, quota limits, and gateway filtering can look like verification failures. Re-run when the environment is healthy.
- A determined adversarial gateway could theoretically proxy selected checks to a real GLM endpoint.
None. Pure Python 3.8+ standard library.
MIT — see LICENSE.
verify-glm 是一个 Claude Code skill + Python 命令行工具，用来验证你当前接入的 Claude Code endpoint 是否真的在跑 GLM-5.2。
它不是简单问模型"你是谁"，而是用：
- 配置检查
- API 回显
- Tokenizer 指纹
reasoning_tokens元数据- 可选 1M 上下文探测
来判断后端是否像真正的 GLM-5.2。
你把 Claude Code 配成 ANTHROPIC_MODEL=glm-5.2，不代表后端一定是真的 GLM-5.2。
网关可能：
- 响应里写
"model": "glm-5.2"，实际后端跑更便宜的模型 /v1/models列表里有glm-5.2，实际路由到 GLM-4 或别的模型- 在 OpenAI-compatible / Claude-compatible API 后面悄悄换模型
- 让模型按 system prompt 自称 GLM，骗过普通问答
所以你需要带外证据。verify-glm 就是用可复现的检测项给你证据。
- 正在用 Claude Code 接 GLM-5.2 的开发者
- 使用 OpenAI-compatible / Claude-compatible 网关的人
- 正在评估 GLM-5.2 做 AI Coding 的团队
- 想确认上下文窗口、tokenizer、reasoning metadata 是否真实的人
- 担心模型被替换、降级、代理转发的人
| # | 检查项 | 作用 |
|---|---|---|
| 1 | 配置检查 | 确认本机 Claude Code 配置确实指向 GLM |
| 2 | API 回显 | 确认 endpoint 可用，并返回期望 model 名 |
| 3 | Tokenizer 指纹 | 固定字符串 token 数能反映模型家族；ChatGLM-style 与 tiktoken / Qwen / Llama 差异明显 |
| 4 | reasoning_tokens 字段 |
GLM-4.5+ 会在 usage metadata 里返回该字段 |
| 5 可选 | 1M 上下文探测 | 发送接近 1M token 的 prompt，验证长上下文是否真实 |
1-4 全过，说明大概率是 GLM 家族。再跑过 --full，则更接近 GLM-5.2 级别长上下文能力。
verify-glm — checking active endpoint
base_url: https://<your-glm-gateway>/v1
model: glm-5.2
[1/4] Config check
✅ Configured for GLM
[2/4] API echo
✅ API echoed model=glm-5.2
[3/4] Tokenizer fingerprint
✅ CN char-level (ratio=0.92), EN subword (7 tokens) — matches ChatGLM family
[4/4] reasoning_tokens field
✅ reasoning_tokens=0 (GLM-4.5+ feature)
✅ Verified GLM family (4/4). Model claims to be: glm-5.2
加 --full 后会额外跑 1M 上下文探测：
[5/5] 1M context probe
✅ accepted ~900K tokens
Verdict: likely GLM-5.2-class endpoint
安装后，直接对 Claude Code 说：
验证一下我是不是真的在跑 GLM-5.2
Claude Code 会发现 verify-glm skill 并运行检测。
git clone https://github.com/xiaoyanfei-tech/verify-glm.git
cd verify-glm
# Linux / macOS / Git Bash
./install.sh
# Windows PowerShell
.\install.ps1也可以直接把 skills/verify-glm/ 拷贝到 ~/.claude/skills/。
# 默认 4 项检查：快，token 消耗低
python ~/.claude/skills/verify-glm/verify_glm.py
# 加 1M 上下文探测：消耗大，只在需要时跑
python ~/.claude/skills/verify-glm/verify_glm.py --full
# JSON 输出
python ~/.claude/skills/verify-glm/verify_glm.py --json如果这个工具帮你验证了环境，可以把 badge 放到你的 README：
[](https://github.com/xiaoyanfei-tech/verify-glm)效果：
| 失败 | 可能原因 |
|---|---|
| 配置 ❌ | ANTHROPIC_BASE_URL / ANTHROPIC_MODEL 没设置或设置错 |
| 回显 ❌ HTTP 401/403 | API token 错误、过期或被拦截 |
| 回显 ❌ HTTP 400 | 网关不认识这个模型名 |
| Tokenizer ❌ 像 tiktoken | 可能被路由到 Claude / GPT 风格后端，但 model 字段仍写 GLM |
| Tokenizer ❌ 中文 ratio 太低 | 可能是 Qwen / Llama / 自研 tokenizer，不是 GLM 家族 |
reasoning_tokens ❌ |
可能是 4.5 之前的 GLM，也可能不是 GLM，或网关删掉了 usage 细节 |
--full ❌ 上下文错误 |
上下文窗口很可能低于 GLM-5.2 级别 |
脚本按以下顺序读取配置：
- 环境变量：
ANTHROPIC_BASE_URL、ANTHROPIC_MODEL、ANTHROPIC_AUTH_TOKEN或ANTHROPIC_API_KEY ~/.claude/settings.json的env./.claude/settings.json的env
如果你正在企业网络、代理环境、自建网关里配置 Claude Code + GLM-5.2，这个工具可以帮助诊断常见的模型路由和配置问题。
如需 endpoint 验证报告、Claude Code 接入配置、企业 AI Coding 环境排障，可以开 GitHub issue 描述你的场景。
- 这是基于证据的验证，不是密码学级别的模型身份证明。
- Tokenizer 指纹主要识别模型家族；
--full长上下文探测能提供更强的 GLM-5.2 级别证据。 - 网络错误、配额耗尽、网关过滤都可能表现为验证失败，环境恢复后需要重跑。
- 极端对抗场景下，恶意网关理论上可以把部分检测转发给真实 GLM endpoint。
零依赖。Python 3.8+ 标准库即可。
MIT — 见 LICENSE。