From 2e385bd40188f4e09d98977319c35bbc197d3290 Mon Sep 17 00:00:00 2001
From: digua <i@digua.me>
Date: Tue, 22 Sep 2026 10:39:10 +0800
Subject: [PATCH] fix(ai): preserve system role for Qwen chat completions

---
 .../src/ai/__tests__/llm-builder.test.ts      |  6 +--
 packages/node-runtime/src/ai/llm-builder.ts   |  6 ++-
 .../src/services/topics/model-client.test.ts  | 54 +++++++++++++++++++
 3 files changed, 62 insertions(+), 4 deletions(-)

diff --git a/packages/node-runtime/src/ai/__tests__/llm-builder.test.ts b/packages/node-runtime/src/ai/__tests__/llm-builder.test.ts
index 8300615b..dc8dfffa 100644
--- a/packages/node-runtime/src/ai/__tests__/llm-builder.test.ts
+++ b/packages/node-runtime/src/ai/__tests__/llm-builder.test.ts
@@ -121,7 +121,7 @@ describe('buildPiModel', () => {
       baseUrl: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
     })
     assert.equal(model.reasoning, true)
-    assert.deepEqual(model.compat, { thinkingFormat: 'qwen' })
+    assert.deepEqual(model.compat, { thinkingFormat: 'qwen', supportsDeveloperRole: false })
   })
 
   it('auto-infers reasoning=true and compat.thinkingFormat=qwen for custom self-hosted qwen3 model', () => {
@@ -131,7 +131,7 @@ describe('buildPiModel', () => {
       { findModelFn: () => null }
     )
     assert.equal(model.reasoning, true)
-    assert.deepEqual(model.compat, { thinkingFormat: 'qwen' })
+    assert.deepEqual(model.compat, { thinkingFormat: 'qwen', supportsDeveloperRole: false })
   })
 
   it('auto-infers reasoning=true for custom model saved with only capabilities:["chat"]', () => {
@@ -155,7 +155,7 @@ describe('buildPiModel', () => {
       }
     )
     assert.equal(model.reasoning, true)
-    assert.deepEqual(model.compat, { thinkingFormat: 'qwen' })
+    assert.deepEqual(model.compat, { thinkingFormat: 'qwen', supportsDeveloperRole: false })
   })
 
   it('auto-infers reasoning=false and compat=undefined for non-reasoning Anthropic model', () => {
diff --git a/packages/node-runtime/src/ai/llm-builder.ts b/packages/node-runtime/src/ai/llm-builder.ts
index 89bf4d7f..1e86aac8 100644
--- a/packages/node-runtime/src/ai/llm-builder.ts
+++ b/packages/node-runtime/src/ai/llm-builder.ts
@@ -156,6 +156,10 @@ export function buildPiModel(config: PiModelConfig, options?: BuildPiModelOption
       : baseUrl
 
   const { reasoning, compat, thinkingLevelMap } = inferReasoning(config.provider, modelId, modelDef)
+  // Qwen chat-completions endpoints reject developer messages, even for reasoning models.
+  // Override Pi's default for official and custom-hosted Qwen models without disabling thinking.
+  const requiresSystemRole =
+    apiFormat === 'openai-completions' && (config.provider.toLowerCase() === 'qwen' || /\bqwen|qwq/i.test(modelId))
 
   return {
     id: modelId,
@@ -171,6 +175,6 @@ export function buildPiModel(config: PiModelConfig, options?: BuildPiModelOption
     contextWindow,
     // Reasoning and visible text share the output budget on compatible APIs.
     maxTokens: config.maxTokens ?? (reasoning ? 16_384 : 4096),
-    compat,
+    compat: requiresSystemRole ? { ...compat, supportsDeveloperRole: false } : compat,
   }
 }
diff --git a/packages/node-runtime/src/services/topics/model-client.test.ts b/packages/node-runtime/src/services/topics/model-client.test.ts
index 9e7bc81f..e71a4821 100644
--- a/packages/node-runtime/src/services/topics/model-client.test.ts
+++ b/packages/node-runtime/src/services/topics/model-client.test.ts
@@ -1,8 +1,62 @@
 import assert from 'node:assert/strict'
 import test from 'node:test'
 import type { AIServiceConfig } from '../../ai'
+import { completeSimple } from '../../ai/pi-runtime'
 import { createChatTopicModelClient } from './model-client'
 
+// Prevent topic requests from failing with HTTP 400 on Qwen endpoints that reject the developer role.
+for (const { provider, model, expectedRole } of [
+  { provider: 'qwen', model: 'qwen3.8', expectedRole: 'system' },
+  { provider: 'openai-compatible', model: 'qwen3.8', expectedRole: 'system' },
+  { provider: 'openai-compatible', model: 'Qwen/QwQ-32B', expectedRole: 'system' },
+  { provider: 'openai', model: 'o3', expectedRole: 'developer' },
+]) {
+  test(`topic requests preserve the supported prompt role for ${provider}/${model}`, async () => {
+    let messages: Array<{ role: string; content: unknown }> | undefined
+    let reasoning: boolean | undefined
+    const client = createChatTopicModelClient(
+      {
+        id: 'synthetic',
+        name: 'Synthetic',
+        provider,
+        model,
+        baseUrl: 'https://example.invalid/v1',
+        apiKey: 'test-key',
+        createdAt: 1,
+        updatedAt: 1,
+      },
+      {
+        completeSimple: (model, context, options) => {
+          reasoning = model.reasoning
+          return completeSimple(model, context, {
+            ...options,
+            onPayload: async (payload, model) => {
+              const normalized = (await options?.onPayload?.(payload, model)) ?? payload
+              messages = (normalized as { messages: typeof messages }).messages
+              // Capture the real SDK serialization, then stop before any network request.
+              throw new Error('Captured topic request before network')
+            },
+          })
+        },
+      }
+    )
+    await assert.rejects(
+      () =>
+        client.complete(
+          { systemPrompt: 'Synthetic system prompt', userPrompt: 'Synthetic user prompt' },
+          { signal: new AbortController().signal, sessionId: 'synthetic' }
+        ),
+      /Captured topic request before network/
+    )
+    assert.equal(reasoning, true)
+    assert.deepEqual(
+      messages?.map((message) => message.role),
+      [expectedRole, 'user']
+    )
+    assert.equal(messages?.[0].content, 'Synthetic system prompt')
+  })
+}
+
 test('topic model calls disable reasoning and constrain the DeepSeek payload', async () => {
   let capturedOptions: Parameters<typeof import('../../ai').completeSimple>[2]
   let normalizedPayload: unknown