Summary

Route Claude model requests directly to Copilot's native /v1/messages endpoint instead of translating through OpenAI /chat/completions.

Motivation

The current architecture translates every request through an Anthropic ↔ OpenAI format conversion layer. However, GitHub Copilot exposes a native Anthropic /v1/messages endpoint that accepts Claude requests directly. By bypassing the translation for Claude models, we get:

• Zero translation overhead — requests/responses pass through unchanged
• Native thinking support — thinking.type: "adaptive" works out of the box
• Native cache_control — prompt caching via Anthropic semantics
• Native output_config — effort levels supported natively
• Simpler streaming — SSE pipes directly without chunk-by-chunk translation

Changes

New files

• src/lib/sanitize.ts — Copilot backend compatibility sanitizer
  • Strips unsupported fields (context_management, budget_tokens, defer_loading, tool_reference)
  • opus models: thinking.enabled → adaptive, effort → medium
  • Model name normalization (claude-sonnet-4-* → claude-sonnet-4)
• src/services/copilot/create-messages.ts — Native /v1/messages client
• tests/sanitize.test.ts — 16 test cases for the sanitizer

Modified

• src/routes/messages/handler.ts — Routes Claude models to native path, non-Claude models keep existing OpenAI translation

Routing logic

Client (Anthropic format)
  → handler.ts
    ├─ Claude models → sanitize → /v1/messages (native, zero translation)
    └─ Non-Claude   → translateToOpenAI → /chat/completions (existing path)

Test results

All 42 tests pass (26 existing + 16 new sanitizer tests).

Breaking changes

None. Non-Claude models (GPT, etc.) are completely unaffected.