Skip to content

7.4 billion tokens per month. 34 free LLM providers. 635 free model endpoints. All behind one /v1 endpoint, plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypted keys. Personal experimentation only.

TypeScript

fix(proxy): round-trip reasoning_content for DeepSeek thinking models (#255)

Tashfeen2f860c34abc8
diff --git a/server/src/__tests__/routes/proxy-tools.test.ts b/server/src/__tests__/routes/proxy-tools.test.ts
index 4493c43e..85209aef 100644
--- a/server/src/__tests__/routes/proxy-tools.test.ts
+++ b/server/src/__tests__/routes/proxy-tools.test.ts
@@ -181,4 +181,43 @@ describe('Proxy tool-calling support', () => {
expect(providerBody.messages[2].tool_call_id).toBe('call_weather_1');
expect(body.choices[0].message.content).toContain('30C');
});
+
+ it('round-trips assistant reasoning_content on follow-up turns (DeepSeek thinking — #255)', async () => {
+ const origFetch = global.fetch;
+ let providerBody: any = null;
+
+ vi.spyOn(global, 'fetch').mockImplementation(async (url, init) => {
+ const urlStr = typeof url === 'string' ? url : url.toString();
+ if (urlStr.includes('api.groq.com/openai/v1/chat/completions')) {
+ providerBody = JSON.parse((init as any).body);
+ return {
+ ok: true,
+ json: () => Promise.resolve({
+ id: 'chatcmpl-r', object: 'chat.completion', created: 1, model: 'm',
+ choices: [{ index: 0, message: { role: 'assistant', content: 'done' }, finish_reason: 'stop' }],
+ usage: { prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 },
+ }),
+ } as any;
+ }
+ return origFetch(url, init);
+ });
+
+ const { status } = await request(app, 'POST', '/v1/chat/completions', {
+ messages: [
+ { role: 'user', content: 'think then answer' },
+ {
+ role: 'assistant',
+ content: 'partial',
+ // What a DeepSeek thinking model returned last turn and the client
+ // replayed. Stripping it makes OpenCode Zen 400 on this request.
+ reasoning_content: 'Let me reason about this step by step...',
+ },
+ { role: 'user', content: 'continue' },
+ ],
+ }, authHeaders());
+
+ expect(status).toBe(200);
+ expect(providerBody.messages[1].role).toBe('assistant');
+ expect(providerBody.messages[1].reasoning_content).toBe('Let me reason about this step by step...');
+ });
});
diff --git a/server/src/routes/proxy.ts b/server/src/routes/proxy.ts
index 1fb5e283..9f11fd32 100644
--- a/server/src/routes/proxy.ts
+++ b/server/src/routes/proxy.ts
@@ -238,6 +238,11 @@ const assistantMessageSchema = z.object({
// no-tool assistant turns — aionrs (AionUI's engine) writes it into every
// session-resumed assistant echo. Treated as absent. (#200)
tool_calls: z.array(toolCallSchema).nullable().optional(),
+ // Thinking trace echoed back by a client. DeepSeek thinking models on
+ // OpenCode Zen 400 ("reasoning_content in thinking mode must be passed back")
+ // unless the prior turn's reasoning_content is replayed, so keep it through
+ // validation instead of stripping it. See issue #255.
+ reasoning_content: z.string().nullable().optional(),
});
// Tool results may arrive with null/missing content (a tool that returned
@@ -507,6 +512,13 @@ proxyRouter.post('/chat/completions', async (req: Request, res: Response) => {
role: 'assistant',
content: assistantContent,
...(m.name ? { name: m.name } : {}),
+ // Replay the thinking trace verbatim. DeepSeek thinking models on
+ // OpenCode Zen reject a follow-up turn that drops it; other providers
+ // ignore the unknown field. Same round-trip rationale as
+ // thought_signature below. (#255)
+ ...(typeof m.reasoning_content === 'string' && m.reasoning_content.length > 0
+ ? { reasoning_content: m.reasoning_content }
+ : {}),
// hasToolCalls (not a bare truthiness check) so null AND empty-array
// tool_calls are dropped rather than forwarded — strict upstreams
// reject both shapes. (#200)
diff --git a/shared/types.ts b/shared/types.ts
index 777fc4ce..d5b8dc4e 100644
--- a/shared/types.ts
+++ b/shared/types.ts
@@ -139,6 +139,10 @@ export interface ChatMessage {
name?: string;
tool_call_id?: string;
tool_calls?: ChatToolCall[];
+ // The model's thinking trace on an assistant turn. Some thinking models
+ // (DeepSeek on OpenCode Zen) require it to be replayed verbatim on the next
+ // turn or they 400; the proxy preserves and forwards it. See issue #255.
+ reasoning_content?: string;
}
export interface ChatCompletionRequest {