fix: preserve system prompts for Mistral Chat Completions

Author: SaladDay

packages/local-runtime-v2/src/service/model-system/resolution/local-model-resolver.test.ts

@@ -1384,19 +1384,22 @@ describe('LocalModelResolver custom provider compat overrides', () => {
   });
 
   // The incident was a wire-level symptom: pi chooses the system prompt role from the
-  // resolved compat, so these two cases pin the request pi would actually send.
+  // resolved compat, so these cases pin the request pi would actually send.
   const reasoningModelConfig: LocalModelConfig = {
     reasoning: true,
     thinking: { effortOptions: ['low', 'light', 'max'] },
   };
 
-  const systemPromptRoleFor = async (compat: LocalModelConfig['compat']) => {
+  const systemPromptRoleFor = async (
+    compat: LocalModelConfig['compat'],
+    baseURL = 'https://gateway.example/v1',
+  ) => {
     const resolver = new LocalModelResolver({
       byokConfigGetter: () => ({
         custom_provider: {
           gateway: {
             api: 'openai-completions',
-            options: { apiKey: 'gateway-key', baseURL: 'https://gateway.example/v1' },
+            options: { apiKey: 'gateway-key', baseURL },
             models: { 'kimi-k2-thinking': { ...reasoningModelConfig, compat } },
           },
         },
@@ -1443,4 +1446,23 @@ describe('LocalModelResolver custom provider compat overrides', () => {
   it('still sends `developer` when the gateway declares nothing, reproducing the incident', async () => {
     expect(await systemPromptRoleFor(undefined)).toBe('developer');
   });
+
+  it('keeps the system role for a thinking model on the Mistral API', async () => {
+    expect(await systemPromptRoleFor(undefined, 'https://api.mistral.ai/v1')).toBe('system');
+  });
+
+  it.each([
+    ['https://api.openai.com/v1', 'developer'],
+    ['https://api.deepseek.com/v1', 'system'],
+    ['https://api.mistral.ai.example/v1', 'developer'],
+    ['https://gateway.example/api.mistral.ai/v1', 'developer'],
+  ])('preserves the system prompt role for %s', async (baseURL, role) => {
+    expect(await systemPromptRoleFor(undefined, baseURL)).toBe(role);
+  });
+
+  it('honors an explicit developer-role override on the Mistral API', async () => {
+    expect(
+      await systemPromptRoleFor({ supportsDeveloperRole: true }, 'https://api.mistral.ai/v1'),
+    ).toBe('developer');
+  });
 });

third_party/pi-mono/MINIMAX_CHANGES.md

@@ -13,6 +13,14 @@ This directory vendors `pi-mono` as source so MiniMax can patch, validate, and s
 
 No upstream source files are changed in the baseline import.
 
+### 2026-09-19 — preserve the system role for Mistral Chat Completions
+
+- Reason: thinking-enabled custom OpenAI-compatible connections to `api.mistral.ai` emitted `developer`, which is absent from the [Mistral Chat Completions message contract](https://docs.mistral.ai/api/endpoint/chat). [OpenClaw's compatibility defaults](https://github.com/openclaw/openclaw/blob/e2bcb1614de060927121bd72de850cee3a08d308/packages/ai/src/transports/openai-completions-compat.ts#L184-L210) also disable this role for the Mistral public endpoint.
+- Affected package: `packages/ai` (`@earendil-works/pi-ai`), OpenAI Completions compatibility detection.
+- Change: default to `system` only on the exact Mistral API host. Preserve thinking, explicit compatibility overrides, all other request options, and other endpoints. Native Mistral transport is unchanged.
+- Upstream PR: not opened.
+- Validation: outgoing-payload regressions in `packages/local-runtime-v2/src/service/model-system/resolution/local-model-resolver.test.ts`; the endpoint regression fails before the fix. No live Mistral requests or full CLI acceptance were performed; the decision is based on the official contract and current upstream implementation.
+
 ### 2026-08-31 — Windows PowerShell ConstrainedLanguage compatibility
 
 - Reason: the Windows PowerShell 5.1 stdin wrapper called `Parser.ParseInput`, `ScriptBlock.Create`, and other restricted .NET APIs before user commands. Under enterprise App Control / AppLocker `ConstrainedLanguage`, the wrapper therefore failed before commands such as Python could run.

third_party/pi-mono/packages/ai/src/providers/openai-completions.ts

@@ -1090,6 +1090,7 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet
 	const isTogether =
 		provider === "together" || baseUrl.includes("api.together.ai") || baseUrl.includes("api.together.xyz");
 	const isMoonshot = provider === "moonshotai" || provider === "moonshotai-cn" || baseUrl.includes("api.moonshot.");
+	const isMistralApi = /^https?:\/\/api\.mistral\.ai(?::\d+)?(?:\/|$)/i.test(baseUrl);
 	const isOpenRouter = provider === "openrouter" || baseUrl.includes("openrouter.ai");
 	const isCloudflareWorkersAI = provider === "cloudflare-workers-ai" || baseUrl.includes("api.cloudflare.com");
 	const isCloudflareAiGateway = provider === "cloudflare-ai-gateway" || baseUrl.includes("gateway.ai.cloudflare.com");
@@ -1124,7 +1125,8 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet
 
 	return {
 		supportsStore: !isNonStandard,
-		supportsDeveloperRole: isOpenRouterDeveloperRoleModel || (!isNonStandard && !isOpenRouter),
+		supportsDeveloperRole:
+			!isMistralApi && (isOpenRouterDeveloperRoleModel || (!isNonStandard && !isOpenRouter)),
 		supportsReasoningEffort:
 			!isGrok && !isZai && !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia && !isAntLing,
 		supportsUsageInStreaming: true,