Support model names containing / without URL encoding

[Copilot]

parsePathAndModel() splits the URL pathname on / to extract the model name as the last segment, making model names like z-ai/glm4.7 impossible to express — the slash is consumed as a path separator.

Changes

• source/proxy.ts — Added a /~/ tilde-separator format that explicitly marks where the base URL ends and the model name begins, allowing any characters (including /) in the model name without encoding. The legacy last-segment format is preserved for backward compatibility, with decodeURIComponent() so %2F-encoded slashes also continue to work.
• test/dynamic-routing.test.ts — Added tests for both the new /~/ separator format and the legacy %2F percent-encoding format, including assertions that the upstream API is called with the correct base URL and decoded model name.
• README.md — Documented both URL formats with examples in the provider table and quick-start section.

URL Formats

Two formats are supported:

New tilde-separator format (recommended — no encoding required):

# Model "z-ai/glm4.7" — use /~/ to separate the base URL path from the model name
export ANTHROPIC_BASE_URL=https://claude-proxy.yinxulai.com/https/api.example.com/v1/~/z-ai/glm4.7

Legacy format (backward-compatible — %2F encoding still works):

# Model "z-ai/glm4.7" — encode the slash as %2F in the path
curl -X POST "https://claude-proxy.yinxulai.com/https/api.example.com/v1/z-ai%2Fglm4.7/v1/messages" \
  -H "x-api-key: your-key" \
  -d '{"model": "z-ai/glm4.7", "messages": [...], "max_tokens": 100}'

In the tilde-separator format, everything between /~/ and /v1/messages is treated as the model name, so slashes in model names require no special treatment.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>模型名能否支持包含/，如"model": "z-ai/glm4.7",</issue_title>
<issue_description>openai请求如下：
curl -X POST "https://api-domain/path"
-H "Content-Type: application/json"
-H "Authorization: Bearer api-key"
-d '{
"model": "z-ai/glm4.7",
"messages": [
{
"role": "user",
"content": "Hello, world!"
}
]
}'
目前看好像不支持z-ai/glm4.7，这种包含/的模型名</issue_description>

<agent_instructions>看看目前是否支持，如果不支持可以提交一个 PR 处理下</agent_instructions>

Comments on the Issue (you are @copilot in this section)

@yinxulai 暂不支持，考虑提个 PR 吗

• Fixes 模型名能否支持包含/，如"model": "z-ai/glm4.7", #1

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.

[yinxulai]

@copilot
不要求用户进行编码的方式下可以支持吗

[Copilot]

@copilot
不要求用户进行编码的方式下可以支持吗

支持了。新增了 /~/ 分隔符格式，在路径中用 /~/ 将 base URL 和模型名分开，/~/ 之后的所有内容（直到 /v1/messages）都作为模型名，无需任何编码：

# z-ai/glm4.7 直接放在 /~/ 后面，无需编码
export ANTHROPIC_BASE_URL=https://claude-proxy.yinxulai.com/https/api.example.com/v1/~/z-ai/glm4.7

代理收到 https://claude-proxy.yinxulai.com/https/api.example.com/v1/~/z-ai/glm4.7/v1/messages 后解析为：

• base URL：https://api.example.com/v1
• 模型名：z-ai/glm4.7

原有格式（模型名作为最后一段路径）以及 %2F 编码均保持向后兼容。相关变更在 b6eed75。

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• api.unavailable.com
  • Triggering command: /opt/hostedtoolcache/node/24.14.0/x64/bin/node /opt/hostedtoolcache/node/24.14.0/x64/bin/node --conditions node --conditions development /home/REDACTED/work/claude-proxy/claude-proxy/node_modules/tinypool/dist/entry/process.js (dns block)
• invalid
  • Triggering command: /opt/hostedtoolcache/node/24.14.0/x64/bin/node /opt/hostedtoolcache/node/24.14.0/x64/bin/node --conditions node --conditions development /home/REDACTED/work/claude-proxy/claude-proxy/node_modules/tinypool/dist/entry/process.js (dns block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)

[yinxulai]

@copilot 给我提出几个实现方案，我希望的方案是对用户影响最小的那种，最好是可以保持当前用法完全不影响用户那种

[Copilot]

@copilot 给我提出几个实现方案，我希望的方案是对用户影响最小的那种，最好是可以保持当前用法完全不影响用户那种

以下是几种可行方案，从对现有用户影响最大到最小排列：

---

方案 A：%2F 百分号编码（PR 原始版本）

ANTHROPIC_BASE_URL=.../https/api.example.com/v1/z-ai%2Fglm4.7

• 需要用户对 / 进行 URL 编码
• 对现有用户无影响，但新用户需要手动编码

---

方案 B：/~/ 分隔符（当前 PR 版本）

ANTHROPIC_BASE_URL=.../https/api.example.com/v1/~/z-ai/glm4.7

• 明确分隔 base URL 和模型名，无需编码
• 对现有用户无影响，但 URL 格式稍有变化

---

方案 C：x-model-name 请求头（对用户影响最小）

• URL 只包含 base URL，模型名通过自定义请求头传递：

  ANTHROPIC_BASE_URL=.../https/api.example.com/v1
  Header: x-model-name: z-ai/glm4.7
  

• 完全不影响 URL 格式，现有用户零改动
• 缺点：客户端需要支持发送自定义请求头（Claude Code 通过 ANTHROPIC_HEADERS 环境变量可以做到）

---

方案 D：/vN/ 版本号启发式识别（URL 格式完全不变）

检测路径中最后一个 /v{数字}/ 作为 base URL 的结束点，其后所有内容为模型名：

/https/api.example.com/v1/z-ai/glm4.7  →  baseUrl=.../v1, model=z-ai/glm4.7
/https/api.openai.com/v1/gpt-4o        →  baseUrl=.../v1, model=gpt-4o

• URL 格式完全不变，现有用户零改动
• 缺点：依赖路径中必须有 /v{N}/；对于 api.some-provider.com/api/v2/openai/custom-model 这类 base URL 含有非版本额外路径段的情况会解析错误（已有测试用例覆盖这一场景，会 break）

---

我的推荐

如果目标是"对现有用户影响最小"：

• 方案 C（自定义请求头） 是唯一一个不需要修改任何已有 URL 格式的方案，现有用户完全无感知。可以同时保留 URL 路径方式作为兜底，两种方式共存。

如果接受方案 C，我可以实现：当请求头 x-model-name 存在时，优先用它作为模型名，URL 路径只作为 base URL；否则回退到当前的路径解析逻辑（兼容所有现有用法）。请告知选择哪种方案。