luulam/9router
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(claude-to-openai): handle OpenAI-format responses in non-streaming path

Browse Source 
Some providers (e.g. xiaomi-tokenplan -claude models) return OpenAI-format
responses even when request was translated to Claude. Early-return now detects
choices[]. Also strip reasoning_content only when content is non-empty so
thinking models keep their only output.

Closes #1836

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
fjia
2026-06-17 09:43:58 +07:00
committed by decolua
parent 740093d852
commit 411a589781
1 changed files with 12 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
open-sse/handlers/chatCore/nonStreamingHandler.js
View File

@@ -76,7 +76,12 @@ export function translateNonStreamingResponse(responseBody, targetFormat, source
// missing/null (e.g. M3 with max_tokens:1 spends the budget on thinking
// and returns `content: null`). Returning the raw body would leave the
// OpenAI client without a `choices` array and surface as a UI test error.
if (responseBody.content && !Array.isArray(responseBody.content)) return responseBody;
// Early return if the response is already in OpenAI format (has choices array)
// or if it has content as a non-array value (likely a different non-Claude format).
// Some providers (e.g. xiaomi-tokenplan) return OpenAI-format responses even when
// the request was translated to Claude format — the targetFormat is Claude but the
// actual response is OpenAI-native and needs no further translation.
if (responseBody.choices || (responseBody.content && !Array.isArray(responseBody.content))) return responseBody;
let textContent = "", thinkingContent = "";
const toolCalls = [];
@@ -193,11 +198,14 @@ export async function handleNonStreamingResponse({ providerResponse, provider, m
translatedResponse.usage = filterUsageForFormat(addBufferToUsage(translatedResponse.usage), sourceFormat);
}
// Strip reasoning_content — some clients (e.g. Firecrawl AI SDK) have JSON parsers that
// break on this non-standard field, even though OpenAI allows it in extensions.
// Strip reasoning_content only when content is non-empty.
// When content is empty (e.g. thinking models that used all tokens for reasoning),
// reasoning_content is the only useful output and must be preserved.
if (translatedResponse?.choices) {
for (const choice of translatedResponse.choices) {
if (choice?.message) delete choice.message.reasoning_content;
if (choice?.message?.reasoning_content && choice.message.content) {
delete choice.message.reasoning_content;
}
}
}