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): strip assistant reasoning_content for Groq and Cerebras (#1093)

suantea (asuan_suan)2b1fc7b8269b
diff --git a/server/src/__tests__/providers/openai-compat.test.ts b/server/src/__tests__/providers/openai-compat.test.ts
index 80df5915..d7ab31f6 100644
--- a/server/src/__tests__/providers/openai-compat.test.ts
+++ b/server/src/__tests__/providers/openai-compat.test.ts
@@ -413,6 +413,68 @@ describe('OpenAICompatProvider', () => {
});
});
+ it('strips assistant reasoning_content before sending messages to Groq (#1070)', async () => {
+ let body: any = null;
+ vi.spyOn(global, 'fetch').mockImplementation(async (_url, init) => {
+ body = JSON.parse((init as any).body);
+ return {
+ ok: true,
+ json: () => Promise.resolve({
+ id: 'id', object: 'chat.completion', created: 1, model: 'm',
+ choices: [{ index: 0, message: { role: 'assistant', content: 'ok' }, finish_reason: 'stop' }],
+ usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+ }),
+ } as any;
+ });
+
+ const groq = new OpenAICompatProvider({ platform: 'groq', name: 'Groq', baseUrl: 'https://api.groq.com/openai/v1' });
+ await groq.chatCompletion(
+ 'k',
+ [{
+ role: 'assistant',
+ content: 'replayed turn',
+ reasoning_content: 'private chain from a prior thinking turn',
+ }],
+ 'llama-3.3-70b-versatile',
+ );
+
+ // Groq is strict about unknown fields: reasoning_content must NOT reach the wire.
+ expect(body.messages[0]).not.toHaveProperty('reasoning_content');
+ expect(body.messages[0]).toEqual({ role: 'assistant', content: 'replayed turn' });
+ });
+
+ it('strips assistant reasoning_content before sending messages to Cerebras (#1070)', async () => {
+ let body: any = null;
+ vi.spyOn(global, 'fetch').mockImplementation(async (_url, init) => {
+ body = JSON.parse((init as any).body);
+ return {
+ ok: true,
+ json: () => Promise.resolve({
+ id: 'id', object: 'chat.completion', created: 1, model: 'm',
+ choices: [{ index: 0, message: { role: 'assistant', content: 'ok' }, finish_reason: 'stop' }],
+ usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+ }),
+ } as any;
+ });
+
+ const cerebras = new OpenAICompatProvider({ platform: 'cerebras', name: 'Cerebras', baseUrl: 'https://api.cerebras.ai/v1' });
+ await cerebras.chatCompletion(
+ 'k',
+ [{
+ role: 'assistant',
+ content: 'replayed turn',
+ reasoning_content: 'private chain from a prior thinking turn',
+ }],
+ 'zai-glm-4.7',
+ );
+
+ // Cerebras is strict about unknown fields: reasoning_content must NOT reach
+ // the wire (verified: 400 wrong_api_format 'property ... reasoning_content
+ // is unsupported', see vercel/ai#15042 / opencode#26762).
+ expect(body.messages[0]).not.toHaveProperty('reasoning_content');
+ expect(body.messages[0]).toEqual({ role: 'assistant', content: 'replayed turn' });
+ });
+
it('folds reasoning into content when content is empty (Ollama style — bare `reasoning` field)', async () => {
vi.spyOn(global, 'fetch').mockResolvedValueOnce({
ok: true,
diff --git a/server/src/__tests__/routes/proxy-tools.test.ts b/server/src/__tests__/routes/proxy-tools.test.ts
index c42d26ad..5f5743c1 100644
--- a/server/src/__tests__/routes/proxy-tools.test.ts
+++ b/server/src/__tests__/routes/proxy-tools.test.ts
@@ -3,6 +3,7 @@ import type { Express } from 'express';
import { createApp } from '../../app.js';
import { initDb, getDb, getUnifiedApiKey } from '../../db/index.js';
import { mintDashboardToken, isGatedApiPath } from '../helpers/auth.js';
+import { clearReasoningMemory } from '../../routes/proxy.js';
let dashToken = '';
@@ -46,6 +47,8 @@ describe('Proxy tool-calling support', () => {
db.prepare('DELETE FROM api_keys').run();
db.prepare('DELETE FROM requests').run();
+ clearReasoningMemory();
+
const addKey = await request(app, 'POST', '/api/keys', {
platform: 'groq',
key: 'gsk_proxy_tool_test',
@@ -219,7 +222,7 @@ describe('Proxy tool-calling support', () => {
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...');
+ expect(providerBody.messages[1].reasoning_content).toBeUndefined();
});
it('restores session reasoning_content the client dropped on replay (#797)', async () => {
@@ -274,7 +277,7 @@ describe('Proxy tool-calling support', () => {
}, sessHeaders);
expect(second.status).toBe(200);
expect(providerBody.messages[1].role).toBe('assistant');
- expect(providerBody.messages[1].reasoning_content).toBe('session trace from turn one');
+ expect(providerBody.messages[1].reasoning_content).toBeUndefined();
});
it('sends the client\'s messages untouched when the session remembers no reasoning (#797)', async () => {
@@ -364,7 +367,7 @@ describe('Proxy tool-calling support', () => {
expect(calls).toHaveLength(3);
// Attempt 1 goes to the model that produced the trace — restored.
- expect(calls[1].messages[1].reasoning_content).toBe('trace bound to the first model');
+ expect(calls[1].messages[1].reasoning_content).toBeUndefined();
// Attempt 2 is a different model. The restore lived on the outbound copy
// only, so this hop carries exactly what the client sent — no leaked
// reasoning_content from the mutated request body.
diff --git a/server/src/providers/openai-compat.ts b/server/src/providers/openai-compat.ts
index 7747096f..52de8bfa 100644
--- a/server/src/providers/openai-compat.ts
+++ b/server/src/providers/openai-compat.ts
@@ -138,12 +138,20 @@ export class OpenAICompatProvider extends BaseProvider {
return { temperature: options?.temperature, topP: options?.top_p };
}
- /** Mistral's OpenAI-compatible endpoint is strict about unknown nested fields
- * and returns 422 for provider-private replay fields that other gateways
- * ignore. Keep the OpenAI wire shape, but strip our internal reasoning /
- * thought-signature extensions before sending to Mistral. */
+ /**
+ * OpenAI-compatible endpoints that are strict about unknown nested fields:
+ * Mistral returns 422 for provider-private replay fields, Groq rejects
+ * assistant `reasoning_content` with 400 (verified in production, #1070),
+ * and Cerebras rejects it too — `property 'messages.N.assistant.
+ * reasoning_content' is unsupported` (confirmed via vercel/ai#15042 and
+ * opencode#26762, and by Cerebras' own docs, which use a `reasoning` field
+ * instead). Other gateways ignore the fields, so keep the OpenAI wire
+ * shape but strip our internal reasoning / thought-signature extensions
+ * before sending to these platforms.
+ */
+ private static readonly STRICT_PLATFORMS = new Set(['mistral', 'groq', 'cerebras']);
private messagesForPlatform(messages: ChatMessage[]): ChatMessage[] {
- if (this.platform !== 'mistral') return messages;
+ if (!OpenAICompatProvider.STRICT_PLATFORMS.has(this.platform)) return messages;
return messages.map((m) => {
if (m.role === 'assistant') {
diff --git a/server/src/routes/proxy.ts b/server/src/routes/proxy.ts
index ae4d3a41..c4eeb981 100644
--- a/server/src/routes/proxy.ts
+++ b/server/src/routes/proxy.ts
@@ -193,6 +193,11 @@ function rememberReasoning(sessionKey: string | undefined, modelKey: string, rea
// The remembered trace for this session, or undefined when there is none, it
// expired, or it came from a different model than the one about to be called.
// An expired entry is dropped on read rather than left for the size sweep.
+export function clearReasoningMemory() {
+ reasoningMemory.clear();
+ stickySessionMap.clear();
+}
+
function rememberedReasoningFor(sessionKey: string, modelKey: string): string | undefined {
if (!sessionKey) return undefined;
const entry = reasoningMemory.get(sessionKey);