@p-sw/brainbox 0.1.2-alpha.4 → 0.1.2-alpha.5

package/dist/index.js

@@ -430,6 +430,21 @@ function defaultReasoningEffort(effort, model, identityModel) {
     return effort;
   return model === identityModel ? "medium" : "none";
 }
+function stripThinkTags(content) {
+  return content.replace(/<think\b[^>]*>[\s\S]*?<\/think>/gi, "").trim();
+}
+function parseModelJson(content) {
+  const cleaned = stripThinkTags(content);
+  try {
+    return JSON.parse(cleaned);
+  } catch (first) {
+    const fence = cleaned.match(/```(?:json)?\s*([\s\S]*?)```/i);
+    if (fence?.[1]) {
+      return JSON.parse(fence[1].trim());
+    }
+    throw first;
+  }
+}
 function readAuthString(auth, key, envName) {
   const fromAuth = typeof auth?.[key] === "string" ? auth[key] : undefined;
   if (fromAuth)
@@ -546,7 +561,7 @@ function fromOrChoice(choice) {
   }));
   return {
     message: {
-      content: typeof msg.content === "string" ? msg.content : undefined,
+      content: typeof msg.content === "string" ? stripThinkTags(msg.content) : undefined,
       toolCalls
     }
   };
@@ -598,13 +613,14 @@ class OpenRouterExecutor extends LLMExecutor {
         stream: false
       }
     });
-    const content = result.choices[0]?.message?.content;
+    const raw = result.choices[0]?.message?.content;
+    const content = typeof raw === "string" ? stripThinkTags(raw) : raw;
     if (!content) {
       log3.debug(`call: empty content in choice 0`);
       throw new Error("Empty response from model");
     }
     log3.debug(`call: response ${content.length} chars`);
-    return jsonMode ? JSON.parse(content) : content;
+    return jsonMode ? parseModelJson(content) : content;
   }
   async chatWithTools(model, options) {
     log3.debug(`chatWithTools: model=${model} msgs=${options.messages.length} tools=${options.tools.length}`);
@@ -663,7 +679,7 @@ function fromChoice(choice) {
   }));
   return {
     message: {
-      content: typeof msg.content === "string" ? msg.content : undefined,
+      content: typeof msg.content === "string" ? stripThinkTags(msg.content) : undefined,
       toolCalls
     }
   };
@@ -778,13 +794,14 @@ class OpenAICompatibleExecutor extends LLMExecutor {
       reasoningEffort: reasoning
     });
     const data = await this.sendRequest(body, options.reasoningEffort);
-    const content = data.choices?.[0]?.message?.content;
+    const raw = data.choices?.[0]?.message?.content;
+    const content = typeof raw === "string" ? stripThinkTags(raw) : raw;
     if (!content) {
       log4.debug(`call: empty content in choice 0`);
       throw new Error("Empty response from model");
     }
     log4.debug(`call: response ${content.length} chars`);
-    return jsonMode ? JSON.parse(content) : content;
+    return jsonMode ? parseModelJson(content) : content;
   }
   async chatWithTools(model, options) {
     const reasoning = defaultReasoningEffort(options.reasoningEffort, model, this.models.identity);
@@ -1099,28 +1116,34 @@ class ZenMuxExecutor extends OpenAICompatibleExecutor {
 }
 
 // src/provider/providers/minimax.ts
-class MiniMaxExecutor extends OpenAICompatibleExecutor {
+class MiniMaxBase extends OpenAICompatibleExecutor {
+  buildBody(opts) {
+    const body = super.buildBody(opts);
+    delete body["reasoning_effort"];
+    body["reasoning_split"] = true;
+    body["thinking"] = opts.reasoningEffort && opts.reasoningEffort !== "none" ? { type: "adaptive" } : { type: "disabled" };
+    return body;
+  }
+}
+function minimaxOpts(providerName, baseURL, opts) {
+  return {
+    providerName,
+    baseURL,
+    apiKey: opts.apiKey,
+    conversationModel: opts.conversationModel,
+    identityModel: opts.identityModel
+  };
+}
+
+class MiniMaxExecutor extends MiniMaxBase {
   constructor(opts) {
-    super({
-      providerName: "minimax",
-      baseURL: "https://api.minimax.io/v1",
-      apiKey: opts.apiKey,
-      conversationModel: opts.conversationModel,
-      identityModel: opts.identityModel
-    });
+    super(minimaxOpts("minimax", "https://api.minimax.io/v1", opts));
   }
 }
 
-// src/provider/providers/minimax_cn.ts
-class MiniMaxCnExecutor extends OpenAICompatibleExecutor {
+class MiniMaxCnExecutor extends MiniMaxBase {
   constructor(opts) {
-    super({
-      providerName: "minimax-cn",
-      baseURL: "https://api.minimaxi.com/v1",
-      apiKey: opts.apiKey,
-      conversationModel: opts.conversationModel,
-      identityModel: opts.identityModel
-    });
+    super(minimaxOpts("minimax-cn", "https://api.minimaxi.com/v1", opts));
   }
 }
 
@@ -1273,7 +1296,17 @@ class CloudflareGatewayExecutor extends OpenAICompatibleExecutor {
 // src/provider/providers/cloudflare_workers.ts
 function toCloudflareMessage(m) {
   if (m.role === "assistant") {
-    return { role: "assistant", content: m.content ?? "" };
+    return {
+      role: "assistant",
+      content: m.content ?? "",
+      tool_calls: m.toolCalls?.map((c) => {
+        let args = c.function.arguments;
+        try {
+          args = JSON.parse(c.function.arguments);
+        } catch {}
+        return { id: c.id, name: c.function.name, arguments: args };
+      })
+    };
   }
   if (m.role === "tool") {
     return { role: "tool", content: m.content, tool_call_id: m.toolCallId };
@@ -1335,11 +1368,11 @@ class CloudflareWorkersExecutor extends LLMExecutor {
     if (data.errors && data.errors.length > 0) {
       throw new Error(`cloudflare-workers API error: ${data.errors.map((e) => e.message).join("; ")}`);
     }
-    const content = data.result?.response ?? "";
+    const content = stripThinkTags(data.result?.response ?? "");
     if (!content) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? JSON.parse(content) : content;
+    return jsonMode ? parseModelJson(content) : content;
   }
   async chatWithTools(model, options) {
     const reasoning = defaultReasoningEffort(options.reasoningEffort, model, this.models.identity);
@@ -1364,7 +1397,7 @@ class CloudflareWorkersExecutor extends LLMExecutor {
     if (data.errors && data.errors.length > 0) {
       throw new Error(`cloudflare-workers API error: ${data.errors.map((e) => e.message).join("; ")}`);
     }
-    const content = data.result?.response ?? "";
+    const content = stripThinkTags(data.result?.response ?? "");
     const toolCalls = data.result?.tool_calls?.map((c, idx) => ({
       id: `call_${idx}`,
       function: {
@@ -1402,7 +1435,8 @@ class AzureOpenAIExecutor extends OpenAICompatibleExecutor {
     super({
       providerName: "azure-openai",
       baseURL: `https://${resource || "__resource__"}.openai.azure.com/openai/deployments`,
-      apiKey: opts.apiKey,
+      apiKey: "",
+      defaultHeaders: { "api-key": opts.apiKey },
       conversationModel: opts.conversationModel,
       identityModel: opts.identityModel
     });
@@ -1422,7 +1456,8 @@ class AzureCognitiveExecutor extends OpenAICompatibleExecutor {
     super({
       providerName: "azure-cognitive",
       baseURL: `https://${resource || "__resource__"}.cognitiveservices.azure.com/openai/deployments`,
-      apiKey: opts.apiKey,
+      apiKey: "",
+      defaultHeaders: { "api-key": opts.apiKey },
       conversationModel: opts.conversationModel,
       identityModel: opts.identityModel
     });
@@ -1546,44 +1581,50 @@ class AnthropicExecutor extends LLMExecutor {
     const reasoning = defaultReasoningEffort(options.reasoningEffort, model, this.models.identity);
     const log5 = logger.child("llm:anthropic");
     log5.debug(`call: model=${model} jsonSchema=${jsonMode ? options.jsonSchemaName : "-"} msgLen=${options.message.length}`);
+    const outputCap = 4096;
     const body = {
       model,
-      max_tokens: 4096,
+      max_tokens: outputCap,
       system: options.instruction,
       messages: [{ role: "user", content: options.message }]
     };
     if (reasoning !== "none") {
+      const budget = REASONING_BUDGET[reasoning];
+      body["max_tokens"] = budget + outputCap;
       body["thinking"] = {
         type: "enabled",
-        budget_tokens: REASONING_BUDGET[reasoning]
+        budget_tokens: budget
       };
     }
     const data = await this.send(body);
     if (data.error) {
       throw new Error(`anthropic API error: ${data.error.message ?? "unknown"}`);
     }
-    const text = (data.content ?? []).filter((b) => b.type === "text").map((b) => b.text).join("");
+    const text = stripThinkTags((data.content ?? []).filter((b) => b.type === "text").map((b) => b.text).join(""));
     if (!text) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? JSON.parse(text) : text;
+    return jsonMode ? parseModelJson(text) : text;
   }
   async chatWithTools(model, options) {
     const reasoning = defaultReasoningEffort(options.reasoningEffort, model, this.models.identity);
     const log5 = logger.child("llm:anthropic");
     log5.debug(`chatWithTools: model=${model} msgs=${options.messages.length} tools=${options.tools.length}`);
     const { system, msgs } = toAnthropicMessages(options.messages);
+    const outputCap = 4096;
     const body = {
       model,
-      max_tokens: 4096,
+      max_tokens: outputCap,
       system: system ?? options.instruction,
       messages: msgs,
       tools: options.tools.map(toAnthropicTool)
     };
     if (reasoning !== "none") {
+      const budget = REASONING_BUDGET[reasoning];
+      body["max_tokens"] = budget + outputCap;
       body["thinking"] = {
         type: "enabled",
-        budget_tokens: REASONING_BUDGET[reasoning]
+        budget_tokens: budget
       };
     }
     const data = await this.send(body);
@@ -1591,7 +1632,7 @@ class AnthropicExecutor extends LLMExecutor {
       throw new Error(`anthropic API error: ${data.error.message ?? "unknown"}`);
     }
     const blocks = data.content ?? [];
-    const text = blocks.filter((b) => b.type === "text").map((b) => b.text).join("");
+    const text = stripThinkTags(blocks.filter((b) => b.type === "text").map((b) => b.text).join(""));
     const toolCalls = blocks.filter((b) => b.type === "tool_use").map((b) => ({
       id: b.id,
       function: {
@@ -1717,7 +1758,7 @@ function toAnthropicTool2(t) {
 }
 function extractAnthropicContent(data) {
   const content = data["content"] ?? [];
-  const text = content.filter((b) => b["type"] === "text").map((b) => b["text"]).join("");
+  const text = stripThinkTags(content.filter((b) => b["type"] === "text").map((b) => b["text"]).join(""));
   const toolCalls = content.filter((b) => b["type"] === "tool_use").map((b) => ({
     id: b["id"],
     function: {
@@ -1792,7 +1833,7 @@ class BedrockExecutor extends LLMExecutor {
     if (!text) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? JSON.parse(text) : text;
+    return jsonMode ? parseModelJson(text) : text;
   }
   async chatWithTools(model, options) {
     const log5 = logger.child("llm:bedrock");
@@ -1822,6 +1863,7 @@ var VertexAuthSchema = z5.object({
 }).loose();
 function toGeminiContents(messages) {
   const out = [];
+  const nameByCallId = new Map;
   for (const m of messages) {
     if (m.role === "system")
       continue;
@@ -1835,6 +1877,7 @@ function toGeminiContents(messages) {
         parts.push({ text: m.content });
       if (m.toolCalls) {
         for (const c of m.toolCalls) {
+          nameByCallId.set(c.id, c.function.name);
           let args = {};
           try {
             args = c.function.arguments ? JSON.parse(c.function.arguments) : {};
@@ -1856,21 +1899,26 @@ function toGeminiContents(messages) {
       }
       out.push({
         role: "user",
-        parts: [{ functionResponse: { name: "", response } }]
+        parts: [
+          {
+            functionResponse: {
+              name: nameByCallId.get(m.toolCallId) ?? "",
+              response
+            }
+          }
+        ]
       });
     }
   }
   return out;
 }
-function toGeminiTool(t) {
+function toGeminiTools(tools) {
   return {
-    functionDeclarations: [
-      {
-        name: t.name,
-        description: t.description,
-        parameters: t.parameters ?? { type: "object", properties: {} }
-      }
-    ]
+    functionDeclarations: tools.map((t) => ({
+      name: t.name,
+      description: t.description,
+      parameters: t.parameters ?? { type: "object", properties: {} }
+    }))
   };
 }
 function extractFromGemini(data) {
@@ -1891,7 +1939,10 @@ function extractFromGemini(data) {
       });
     }
   }
-  return { text, toolCalls: toolCalls.length > 0 ? toolCalls : undefined };
+  return {
+    text: stripThinkTags(text),
+    toolCalls: toolCalls.length > 0 ? toolCalls : undefined
+  };
 }
 
 class VertexExecutor extends LLMExecutor {
@@ -1957,7 +2008,7 @@ class VertexExecutor extends LLMExecutor {
     if (!text) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? JSON.parse(text) : text;
+    return jsonMode ? parseModelJson(text) : text;
   }
   async chatWithTools(model, options) {
     const log5 = logger.child("llm:vertex");
@@ -1966,7 +2017,7 @@ class VertexExecutor extends LLMExecutor {
     const body = {
       contents,
       systemInstruction: { parts: [{ text: options.instruction }] },
-      tools: options.tools.length > 0 ? [toGeminiTool(options.tools[0])] : undefined
+      tools: options.tools.length > 0 ? [toGeminiTools(options.tools)] : undefined
     };
     const data = await this.generate(model, body);
     if (data.error) {
@@ -2074,11 +2125,11 @@ class GitLabDuoExecutor extends LLMExecutor {
     if (data.error) {
       throw new Error(`gitlab-duo API error: ${data.error.message ?? "unknown"}`);
     }
-    const content = data.choices?.[0]?.message?.content ?? "";
+    const content = stripThinkTags(data.choices?.[0]?.message?.content ?? "");
     if (!content) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? JSON.parse(content) : content;
+    return jsonMode ? parseModelJson(content) : content;
   }
   async chatWithTools(model, options) {
     const log5 = logger.child("llm:gitlab-duo");
@@ -2102,7 +2153,7 @@ class GitLabDuoExecutor extends LLMExecutor {
     }));
     return {
       message: {
-        content: choice?.message?.content ?? undefined,
+        content: choice?.message?.content ? stripThinkTags(choice.message.content) : undefined,
         toolCalls
       }
     };
@@ -2193,11 +2244,11 @@ class SnowflakeCortexExecutor extends LLMExecutor {
     if (data.error) {
       throw new Error(`snowflake-cortex API error: ${data.error.message ?? "unknown"}`);
     }
-    const content = data.choices?.[0]?.message?.content ?? data.message ?? "";
+    const content = stripThinkTags(data.choices?.[0]?.message?.content ?? data.message ?? "");
     if (!content) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? JSON.parse(content) : content;
+    return jsonMode ? parseModelJson(content) : content;
   }
   async chatWithTools(model, options) {
     const log5 = logger.child("llm:snowflake-cortex");
@@ -2215,7 +2266,7 @@ class SnowflakeCortexExecutor extends LLMExecutor {
       throw new Error(`snowflake-cortex API error: ${data.error.message ?? "unknown"}`);
     }
     const choice = data.choices?.[0];
-    const content = choice?.message?.content ?? data.message ?? "";
+    const content = stripThinkTags(choice?.message?.content ?? data.message ?? "");
     const toolCalls = choice?.message?.tool_calls?.map((c) => ({
       id: c.id,
       function: {
@@ -5282,7 +5333,7 @@ function BrainApp({
       }, undefined, false, undefined, this),
       busy ? /* @__PURE__ */ jsxDEV5(Text5, {
         dimColor: true,
-        children: "Creating brain…"
+        children: "Creating brain… (This can take few minutes depending on the model's response speed)"
       }, undefined, false, undefined, this) : /* @__PURE__ */ jsxDEV5(TextInput, {
         prompt: "seed> ",
         onSubmit: (raw) => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@p-sw/brainbox",
-  "version": "0.1.2-alpha.4",
+  "version": "0.1.2-alpha.5",
   "module": "dist/index.js",
   "main": "dist/index.js",
   "type": "module",