Cursor agent mode: Anthropic tool blocks rejected on OpenAI endpoint (upstream fix available)

[HatriGt]

Summary

When using Cursor in agent / tool mode pointed at VibeProxy's OpenAI-compatible endpoint with a Claude (OAuth) model, tool calling fails. Normal chat works, but any tool use breaks with errors like:

messages.N: user messages must have non-empty content

or silently fails.

Root cause

Cursor sends Anthropic-native content blocks to the OpenAI /v1/chat/completions endpoint:

• user.content with { "type": "tool_result", "tool_use_id": ..., "content": ... }
• assistant.content with { "type": "tool_use", "id": ..., "name": ..., "input": {...} }
• bare tool definitions { "name", "description", "input_schema" }

The OpenAI -> Claude translator expects standard OpenAI shapes (tool_calls, role:"tool", tools[].type=="function"), so these payloads are dropped. This is also tracked upstream in router-for-me/CLIProxyAPI#1165 and the Cursor forum report "Wrong tools handling on OpenAI compatible endpoint".

Today the only workaround is running an external request-rewriting proxy/shim in front of the proxy.

Upstream fix

I opened a PR against the core proxy that normalizes these Anthropic-native blocks into standard OpenAI shapes before translation (standard OpenAI payloads pass through untouched):

• fix(translator): normalize Cursor Anthropic tool blocks on OpenAI endpoint router-for-me/CLIProxyAPI#3804

Once that lands in a release, bundling the updated cli-proxy-api binary in VibeProxy would let Cursor agent mode work out of the box with no shim.

Minor: stale reference

The README links to https://github.com/router-for-me/CLIProxyAPIPlus, which now 404s (it appears to have been consolidated into router-for-me/CLIProxyAPI). Worth updating the README and the bundled binary source accordingly.

Environment

• VibeProxy (latest), macOS
• Client: Cursor (agent mode)
• Model: Claude via OAuth