Bug: Claude Desktop gateway routing fails — alias not reverse-mapped to DeepSeek model (400)

[vehiclerentropy]

Client or integration

Other

Area

Proxy and routing

Summary

Claude Desktop gateway routing fails: alias not reverse-mapped to DeepSeek model (HTTP 400). The desktop client sends the injected alias model id (claude-opus-4-8-20261210) and the proxy forwards that alias verbatim to the routed provider (deepseek) instead of reverse-mapping it to the real model (deepseek/deepseek-v4-flash).

Reproduction

1. ocx login deepseek (add deepseek provider)
2. ocx claude desktop apply — injects configLibrary json with inferenceModels: name=claude-opus-4-8-20261210, labelOverride=Deepseek V4 Flash (deepseek)
3. Open Claude Desktop (Cowork) → model picker shows "Deepseek V4 Flash (deepseek)"
4. Send a message → HTTP 400, model never reaches DeepSeek

Usage log confirms: provider=deepseek, model=claude-opus-4-8-20261210, requestedModel=claude-opus-4-8-20261210, status=400, upstreamError="The supported API model names are deepseek-v4-pro or deepseek-v4-flash, but you passed claude-opus-4-8-20261210".

Additional experiments:

• Manually changing the configLibrary name to deepseek/deepseek-v4-flash → Claude Desktop validation rejects it ("is not an Anthropic model")
• Changing to anthropic/deepseek-v4-flash → passes validation, but the proxy still forwards it verbatim and DeepSeek rejects it
• Expected: proxy should reverse-map claude-opus-4-8-20261210 → deepseek/deepseek-v4-flash via claudeCode.desktopProfile.assignments before forwarding, like it does for Claude Code CLI aliases

Version

2.8.0

Operating system

macOS 26 (Apple Silicon)

Provider and model

deepseek (openai-chat adapter), Claude Desktop 1.24012.9

[github-actions]

_{Automated translation bookkeeping — detected language: unknown.}