fix(ollama): make agent + tool calls work against Ollama Cloud

patches/@mariozechner+pi-ai+0.60.0.patch

@@ -1,8 +1,25 @@
 diff --git a/node_modules/@mariozechner/pi-ai/dist/providers/openai-completions.js b/node_modules/@mariozechner/pi-ai/dist/providers/openai-completions.js
-index 29ae8cf..6d9ea61 100644
+index 29ae8cf..7204589 100644
 --- a/node_modules/@mariozechner/pi-ai/dist/providers/openai-completions.js
 +++ b/node_modules/@mariozechner/pi-ai/dist/providers/openai-completions.js
-@@ -474,13 +474,32 @@ export function convertMessages(model, context, compat) {
+@@ -446,9 +446,15 @@ export function convertMessages(model, context, compat) {
+                     : content;
+                 if (filteredContent.length === 0)
+                     continue;
++                // Flatten text-only content to a plain string for maximum endpoint
++                // compatibility. Ollama and some OpenAI-compatible servers reject the
++                // content-block array form for plain text with "invalid message format".
++                const isTextOnly = filteredContent.every((c) => c.type === "text");
+                 params.push({
+                     role: "user",
+-                    content: filteredContent,
++                    content: isTextOnly
++                        ? filteredContent.map((c) => c.text).join("")
++                        : filteredContent,
+                 });
+             }
+         }
+@@ -474,13 +480,32 @@ export function convertMessages(model, context, compat) {
              // Filter out empty thinking blocks to avoid API validation errors
              const nonEmptyThinkingBlocks = thinkingBlocks.filter((b) => b.thinking && b.thinking.trim().length > 0);
              if (nonEmptyThinkingBlocks.length > 0) {
@@ -37,23 +54,23 @@ index 29ae8cf..6d9ea61 100644
                      else {
                          assistantMsg.content = [{ type: "text", text: thinkingText }];
                      }
-@@ -658,6 +677,7 @@ function mapStopReason(reason) {
+@@ -658,6 +683,7 @@ function mapStopReason(reason) {
  function detectCompat(model) {
      const provider = model.provider;
      const baseUrl = model.baseUrl;
 +    const isDeepSeekV4 = /(?:^|[/_-])deepseek[-_.]?v4(?:$|[-:_.])/i.test(model.id);
      const isZai = provider === "zai" || baseUrl.includes("api.z.ai");
      const isNonStandard = provider === "cerebras" ||
          baseUrl.includes("cerebras.ai") ||
-@@ -690,6 +710,7 @@ function detectCompat(model) {
+@@ -690,6 +716,7 @@ function detectCompat(model) {
          requiresToolResultName: false,
          requiresAssistantAfterToolResult: false,
          requiresThinkingAsText: false,
 +        requiresThinkingInContent: isDeepSeekV4,
          thinkingFormat: isZai ? "zai" : "openai",
          openRouterRouting: {},
          vercelGatewayRouting: {},
-@@ -714,6 +735,7 @@ function getCompat(model) {
+@@ -714,6 +741,7 @@ function getCompat(model) {
          requiresToolResultName: model.compat.requiresToolResultName ?? detected.requiresToolResultName,
          requiresAssistantAfterToolResult: model.compat.requiresAssistantAfterToolResult ?? detected.requiresAssistantAfterToolResult,
          requiresThinkingAsText: model.compat.requiresThinkingAsText ?? detected.requiresThinkingAsText,

src/main/claude/pi-model-resolution.ts

@@ -328,15 +328,20 @@ export function applyPiModelRuntimeOverrides(
     } as typeof nextModel;
   }
 
-  // DeepSeek V4 models on custom/relay endpoints need thinking blocks in content[] array.
+  // DeepSeek V4 models on some custom/relay endpoints (and local Ollama) need thinking
+  // blocks in the content[] array. Ollama CLOUD (ollama.com), however, rejects that
+  // non-standard {type:"thinking"} content block with "invalid message format". Local
+  // Ollama (loopback :11434) is a different service and keeps the existing behavior.
   if (nextModel.api === 'openai-completions' && DEEPSEEK_V4_MODEL_PATTERN.test(nextModel.id)) {
     const currentCompat = (nextModel.compat || {}) as Record<string, unknown>;
-    if (!currentCompat.requiresThinkingInContent) {
+    const endpoint = `${options.customBaseUrl || ''} ${nextModel.baseUrl || ''}`.toLowerCase();
+    const isOllamaCloud = endpoint.includes('ollama.com');
+    if (currentCompat.requiresThinkingInContent === undefined) {
       nextModel = {
         ...nextModel,
         compat: {
           ...currentCompat,
-          requiresThinkingInContent: true,
+          requiresThinkingInContent: !isOllamaCloud,
         },
       } as typeof nextModel;
     }

src/tests/claude/pi-model-resolution.test.ts

@@ -18,6 +18,19 @@ const openAIResponsesModel = {
   maxTokens: 16384,
 } as Model<Api>;
 
+const deepseekV4CompletionsModel = {
+  id: 'deepseek-v4-pro',
+  name: 'deepseek-v4-pro',
+  api: 'openai-completions',
+  provider: 'ollama',
+  baseUrl: 'https://ollama.com/v1',
+  reasoning: true,
+  input: ['text'],
+  cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+  contextWindow: 128000,
+  maxTokens: 16384,
+} as Model<Api>;
+
 describe('pi model runtime overrides', () => {
   it('keeps OpenAI Responses for custom OpenAI configs that target official OpenAI', () => {
     const model = resolvePiRegistryModel('openai/gpt-5.4', {
@@ -46,4 +59,34 @@ describe('pi model runtime overrides', () => {
       supportsStore: false,
     });
   });
+
+  it('disables thinking-in-content for DeepSeek V4 on Ollama Cloud (ollama.com)', () => {
+    // Ollama Cloud rejects the non-standard {type:"thinking"} content block with
+    // "invalid message format", so thinking must not be serialized into content[].
+    const model = applyPiModelRuntimeOverrides(deepseekV4CompletionsModel, {
+      configProvider: 'ollama',
+      rawProvider: 'ollama',
+      customBaseUrl: 'https://ollama.com/v1',
+    });
+
+    expect(model.compat).toMatchObject({ requiresThinkingInContent: false });
+  });
+
+  it('keeps thinking-in-content for DeepSeek V4 on other custom/relay endpoints', () => {
+    const model = applyPiModelRuntimeOverrides(
+      {
+        ...deepseekV4CompletionsModel,
+        provider: 'custom',
+        baseUrl: 'https://relay.example.test/v1',
+      } as Model<Api>,
+      {
+        configProvider: 'custom',
+        rawProvider: 'custom',
+        customProtocol: 'openai',
+        customBaseUrl: 'https://relay.example.test/v1',
+      }
+    );
+
+    expect(model.compat).toMatchObject({ requiresThinkingInContent: true });
+  });
 });