Claude models behind OpenAI-compatible proxies get 400 prefill errors

[bussnet]

Bug

Claude models routed through non-native SDK paths (@ai-sdk/openai-compatible, GitHub Copilot, Azure AI / LiteLLM) fail with:

[400]: This model does not support assistant message prefill.
The conversation must end with a user message.

This is likely the shared root cause behind #19517 and #13768.

Root Cause

ProviderTransform in transform.ts uses three different conditions to detect Claude models, and they disagree:

Transform	Condition	Lines
Empty content stripping	api.npm === "@ai-sdk/anthropic"	L56
Tool call ID scrubbing	api.id.includes("claude")	L76
Prompt caching	providerID || api.id || model.id || api.npm	L282-287

When a Claude model uses any SDK other than @ai-sdk/anthropic — proxies, Copilot, Azure AI — caching activates but empty-content stripping does not. Empty assistant messages survive, get cache metadata, and Anthropic rejects them as prefill.

Reproduction

1. Configure a Claude model through an OpenAI-compatible proxy with a model name containing "claude"
2. Use it in a multi-agent conversation with tool calls
3. After several turns, the request fails with the prefill error

Fix

PR #20414 — extracts isClaudeModel() helper used at all three detection sites.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Azure AI Claude rejects requests after tool abort: assistant message prefill not supported #19517: Azure AI Claude rejects requests after tool abort: assistant message prefill not supported — same error message and similar root cause analysis around trailing assistant messages with OpenAI-compatible proxies/LiteLLM
• This model does not support assistant message prefill / Github Copilot with Opus 4.6 #13768: This model does not support assistant message prefill / Github Copilot with Opus 4.6 — same 400 prefill error via an OpenAI-compatible proxy layer

[github-actions]

To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.