@p-sw/brainbox 0.1.2-alpha.11 → 0.1.2-alpha.12

package/dist/index.js

@@ -489,15 +489,100 @@ function stripThinkTags(content) {
 }
 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());
+  const candidates = [cleaned];
+  const fence = cleaned.match(/```(?:json)?\s*([\s\S]*?)```/i);
+  if (fence?.[1])
+    candidates.push(fence[1].trim());
+  const extracted = extractJsonSlice(cleaned);
+  if (extracted)
+    candidates.push(extracted);
+  let lastErr;
+  for (const candidate of candidates) {
+    try {
+      return decodeJsonValue(candidate);
+    } catch (err) {
+      lastErr = err;
+    }
+  }
+  throw lastErr instanceof Error ? lastErr : new Error("Failed to parse model JSON");
+}
+function decodeJsonValue(text) {
+  let value = JSON.parse(text);
+  for (let i = 0;i < 2 && typeof value === "string"; i++) {
+    const inner = value.trim();
+    if (!(inner.startsWith("{") && inner.endsWith("}") || inner.startsWith("[") && inner.endsWith("]") || inner.startsWith('"') && inner.endsWith('"'))) {
+      break;
     }
-    throw first;
+    value = JSON.parse(inner);
   }
+  return value;
+}
+function extractJsonSlice(text) {
+  const start = text.search(/[\{\[]/);
+  if (start < 0)
+    return;
+  const open = text[start];
+  const close = open === "{" ? "}" : "]";
+  let depth = 0;
+  let inString = false;
+  let escape = false;
+  for (let i = start;i < text.length; i++) {
+    const ch = text[i];
+    if (inString) {
+      if (escape) {
+        escape = false;
+      } else if (ch === "\\") {
+        escape = true;
+      } else if (ch === '"') {
+        inString = false;
+      }
+      continue;
+    }
+    if (ch === '"') {
+      inString = true;
+      continue;
+    }
+    if (ch === open)
+      depth++;
+    else if (ch === close) {
+      depth--;
+      if (depth === 0)
+        return text.slice(start, i + 1);
+    }
+  }
+  return;
+}
+function schemaToolName(name) {
+  const cleaned = name.replace(/[^a-zA-Z0-9_]/g, "_").replace(/^(\d)/, "_$1");
+  return cleaned.length > 0 ? cleaned : "submit_result";
+}
+function buildStructuredJsonRequest(options) {
+  const toolName = schemaToolName(options.jsonSchemaName);
+  return {
+    toolName,
+    tool: {
+      name: toolName,
+      description: `Submit the structured ${options.jsonSchemaName} result.`,
+      parameters: options.jsonSchema ?? {
+        type: "object",
+        additionalProperties: true
+      }
+    },
+    instruction: `${options.instruction}
+
+You MUST call the \`${toolName}\` tool exactly once with the complete answer.
+Do not write the JSON as plain text or inside a markdown code fence.`
+  };
+}
+function parseStructuredJsonResult(choice, toolName) {
+  const call = choice.message.toolCalls?.find((c) => c.function.name === toolName) ?? choice.message.toolCalls?.[0];
+  if (call?.function.arguments) {
+    return parseModelJson(call.function.arguments);
+  }
+  if (choice.message.content) {
+    return parseModelJson(choice.message.content);
+  }
+  throw new Error("Empty response from model");
 }
 function readAuthString(auth, key, envName) {
   const fromAuth = typeof auth?.[key] === "string" ? auth[key] : undefined;
@@ -781,6 +866,7 @@ class OpenAICompatibleExecutor extends LLMExecutor {
   chatPath;
   noBearerPrefix;
   reasoningEffortInQuery;
+  supportsResponseFormat;
   constructor(opts) {
     super();
     this.providerName = opts.providerName;
@@ -790,6 +876,7 @@ class OpenAICompatibleExecutor extends LLMExecutor {
     this.chatPath = opts.chatPath ?? "/chat/completions";
     this.noBearerPrefix = opts.noBearerPrefix ?? false;
     this.reasoningEffortInQuery = opts.reasoningEffortInQuery ?? false;
+    this.supportsResponseFormat = opts.supportsResponseFormat ?? true;
     this.models = {
       conversation: opts.conversationModel,
       identity: opts.identityModel
@@ -861,6 +948,9 @@ class OpenAICompatibleExecutor extends LLMExecutor {
     return data;
   }
   async call(model, options) {
+    if ("jsonSchemaName" in options && !this.supportsResponseFormat) {
+      return this.callJsonViaTool(model, options);
+    }
     const jsonMode = "jsonSchemaName" in options;
     const reasoning = defaultReasoningEffort(options.reasoningEffort, model, this.models.identity);
     log4.debug(`call: provider=${this.providerName} model=${model} jsonSchema=${jsonMode ? options.jsonSchemaName : "-"} msgLen=${options.message.length}`);
@@ -886,6 +976,19 @@ class OpenAICompatibleExecutor extends LLMExecutor {
     log4.debug(`call: response ${content.length} chars`);
     return jsonMode ? parseModelJson(content) : content;
   }
+  async callJsonViaTool(model, options) {
+    const { toolName, tool, instruction } = buildStructuredJsonRequest(options);
+    log4.debug(`callJsonViaTool: provider=${this.providerName} model=${model} tool=${toolName}`);
+    const choice = await this.chatWithTools(model, {
+      caller: options.caller ?? options.jsonSchemaName,
+      instruction,
+      messages: [{ role: "user", content: options.message }],
+      tools: [tool],
+      reasoningEffort: "none",
+      parallelToolCalls: false
+    });
+    return parseStructuredJsonResult(choice, toolName);
+  }
   async chatWithTools(model, options) {
     const reasoning = defaultReasoningEffort(options.reasoningEffort, model, this.models.identity);
     log4.debug(`chatWithTools: provider=${this.providerName} model=${model} msgs=${options.messages.length} tools=${options.tools.length}`);
@@ -1219,7 +1322,8 @@ function minimaxOpts(providerName, baseURL, opts) {
     baseURL,
     apiKey: opts.apiKey,
     conversationModel: opts.conversationModel,
-    identityModel: opts.identityModel
+    identityModel: opts.identityModel,
+    supportsResponseFormat: false
   };
 }
 
@@ -1295,7 +1399,8 @@ class LmStudioExecutor extends OpenAICompatibleExecutor {
       baseURL: "http://127.0.0.1:1234/v1",
       apiKey: opts.apiKey || "lm-studio",
       conversationModel: opts.conversationModel,
-      identityModel: opts.identityModel
+      identityModel: opts.identityModel,
+      supportsResponseFormat: false
     });
   }
 }
@@ -1308,7 +1413,8 @@ class OllamaExecutor extends OpenAICompatibleExecutor {
       baseURL: "http://127.0.0.1:11434/v1",
       apiKey: opts.apiKey || "ollama",
       conversationModel: opts.conversationModel,
-      identityModel: opts.identityModel
+      identityModel: opts.identityModel,
+      supportsResponseFormat: false
     });
   }
 }
@@ -1334,7 +1440,8 @@ class LlamaCppExecutor extends OpenAICompatibleExecutor {
       baseURL: "http://127.0.0.1:8080/v1",
       apiKey: opts.apiKey || "llama.cpp",
       conversationModel: opts.conversationModel,
-      identityModel: opts.identityModel
+      identityModel: opts.identityModel,
+      supportsResponseFormat: false
     });
   }
 }
@@ -1573,17 +1680,17 @@ var AnthropicAuthSchema = z4.object({
   apiVersion: z4.string().optional()
 }).loose();
 function toAnthropicMessages(messages) {
-  let system;
-  const msgs = [];
+  let system2;
+  const msgs2 = [];
   for (const m of messages) {
     if (m.role === "system") {
-      system = (system ? system + `
+      system2 = (system2 ? system2 + `
 
 ` : "") + m.content;
       continue;
     }
     if (m.role === "user") {
-      msgs.push({ role: "user", content: m.content });
+      msgs2.push({ role: "user", content: m.content });
       continue;
     }
     if (m.role === "assistant") {
@@ -1606,11 +1713,11 @@ function toAnthropicMessages(messages) {
           });
         }
       }
-      msgs.push({ role: "assistant", content: blocks });
+      msgs2.push({ role: "assistant", content: blocks });
       continue;
     }
     if (m.role === "tool") {
-      msgs.push({
+      msgs2.push({
         role: "user",
         content: [
           {
@@ -1622,7 +1729,7 @@ function toAnthropicMessages(messages) {
       });
     }
   }
-  return { system, msgs };
+  return { system: system2, msgs: msgs2 };
 }
 function toAnthropicTool(t) {
   return {
@@ -1669,10 +1776,25 @@ class AnthropicExecutor extends LLMExecutor {
     return JSON.parse(responseRaw);
   }
   async call(model, options) {
-    const jsonMode = "jsonSchemaName" in options;
+    if ("jsonSchemaName" in options) {
+      const { toolName, tool, instruction } = buildStructuredJsonRequest({
+        instruction: options.instruction,
+        jsonSchemaName: options.jsonSchemaName,
+        jsonSchema: options.jsonSchema
+      });
+      const choice = await this.chatWithTools(model, {
+        caller: options.caller ?? options.jsonSchemaName,
+        instruction,
+        messages: [{ role: "user", content: options.message }],
+        tools: [tool],
+        reasoningEffort: "none",
+        toolChoice: { type: "tool", name: toolName }
+      });
+      return parseStructuredJsonResult(choice, toolName);
+    }
     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=${options.message.length}`);
+    log5.debug(`call: model=${model} jsonSchema=- msgLen=${options.message.length}`);
     const outputCap = 4096;
     const body = {
       model,
@@ -1696,21 +1818,27 @@ class AnthropicExecutor extends LLMExecutor {
     if (!text) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? parseModelJson(text) : text;
+    return text;
   }
   async chatWithTools(model, options) {
     const reasoning = defaultReasoningEffort(options.reasoningEffort, model, this.models.identity);
     const log5 = logger.child("llm:anthropic");
     log5.debug(`chatWithTools: model=${model} msgs=${options.messages.length} tools=${options.tools.length}`);
-    const { system, msgs } = toAnthropicMessages(options.messages);
+    const { system: system2, msgs: msgs2 } = toAnthropicMessages(options.messages);
     const outputCap = 4096;
     const body = {
       model,
       max_tokens: outputCap,
-      system: system ?? options.instruction,
-      messages: msgs,
+      system: system2 ?? options.instruction,
+      messages: msgs2,
       tools: options.tools.map(toAnthropicTool)
     };
+    if (options.toolChoice) {
+      body["tool_choice"] = {
+        type: "tool",
+        name: options.toolChoice.name
+      };
+    }
     if (reasoning !== "none") {
       const budget = REASONING_BUDGET[reasoning];
       body["max_tokens"] = budget + outputCap;
@@ -1789,58 +1917,6 @@ function signRequest(opts) {
     headers["X-Amz-Security-Token"] = opts.sessionToken;
   return headers;
 }
-function toAnthropicMessages2(messages) {
-  let system;
-  const msgs = [];
-  for (const m of messages) {
-    if (m.role === "system") {
-      system = (system ? system + `
-
-` : "") + m.content;
-      continue;
-    }
-    if (m.role === "user") {
-      msgs.push({ role: "user", content: [{ type: "text", text: m.content }] });
-      continue;
-    }
-    if (m.role === "assistant") {
-      const blocks = [];
-      if (m.content)
-        blocks.push({ type: "text", text: m.content });
-      if (m.toolCalls) {
-        for (const c of m.toolCalls) {
-          let input = {};
-          try {
-            input = c.function.arguments ? JSON.parse(c.function.arguments) : {};
-          } catch {
-            input = {};
-          }
-          blocks.push({
-            type: "tool_use",
-            id: c.id,
-            name: c.function.name,
-            input
-          });
-        }
-      }
-      msgs.push({ role: "assistant", content: blocks });
-      continue;
-    }
-    if (m.role === "tool") {
-      msgs.push({
-        role: "user",
-        content: [
-          {
-            type: "tool_result",
-            tool_use_id: m.toolCallId,
-            content: [{ type: "text", text: m.content }]
-          }
-        ]
-      });
-    }
-  }
-  return { system, msgs };
-}
 function toAnthropicTool2(t) {
   return {
     name: t.name,
@@ -1909,9 +1985,24 @@ class BedrockExecutor extends LLMExecutor {
     return JSON.parse(responseRaw);
   }
   async call(model, options) {
-    const jsonMode = "jsonSchemaName" in options;
     const log5 = logger.child("llm:bedrock");
-    log5.debug(`call: model=${model} jsonSchema=${jsonMode ? options.jsonSchemaName : "-"} msgLen=${options.message.length}`);
+    if ("jsonSchemaName" in options) {
+      log5.debug(`call: model=${model} jsonSchema=${options.jsonSchemaName} via tool`);
+      const { toolName, tool, instruction } = buildStructuredJsonRequest({
+        instruction: options.instruction,
+        jsonSchemaName: options.jsonSchemaName,
+        jsonSchema: options.jsonSchema
+      });
+      const choice = await this.chatWithTools(model, {
+        caller: options.caller ?? options.jsonSchemaName,
+        instruction,
+        messages: [{ role: "user", content: options.message }],
+        tools: [tool],
+        toolChoice: { type: "tool", name: toolName }
+      });
+      return parseStructuredJsonResult(choice, toolName);
+    }
+    log5.debug(`call: model=${model} jsonSchema=- msgLen=${options.message.length}`);
     if (!model.startsWith("anthropic.") && !model.startsWith("us.anthropic.")) {
       throw new Error(`bedrock provider currently only supports Anthropic models (got ${model})`);
     }
@@ -1926,7 +2017,7 @@ class BedrockExecutor extends LLMExecutor {
     if (!text) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? parseModelJson(text) : text;
+    return text;
   }
   async chatWithTools(model, options) {
     const log5 = logger.child("llm:bedrock");
@@ -1934,7 +2025,6 @@ class BedrockExecutor extends LLMExecutor {
     if (!model.startsWith("anthropic.") && !model.startsWith("us.anthropic.")) {
       throw new Error(`bedrock provider currently only supports Anthropic models (got ${model})`);
     }
-    const { system, msgs } = toAnthropicMessages2(options.messages);
     const body = {
       anthropic_version: "bedrock-2023-05-31",
       max_tokens: 4096,
@@ -1942,6 +2032,12 @@ class BedrockExecutor extends LLMExecutor {
       messages: msgs,
       tools: options.tools.map(toAnthropicTool2)
     };
+    if (options.toolChoice) {
+      body.tool_choice = {
+        type: "tool",
+        name: options.toolChoice.name
+      };
+    }
     const data = await this.invoke(model, body, resolveLlmCaller(options));
     const { text, toolCalls } = extractAnthropicContent(data);
     return { message: { content: text || undefined, toolCalls } };
@@ -2208,9 +2304,23 @@ class GitLabDuoExecutor extends LLMExecutor {
     return JSON.parse(responseRaw);
   }
   async call(model, options) {
-    const jsonMode = "jsonSchemaName" in options;
     const log5 = logger.child("llm:gitlab-duo");
-    log5.debug(`call: model=${model} jsonSchema=${jsonMode ? options.jsonSchemaName : "-"} msgLen=${options.message.length}`);
+    if ("jsonSchemaName" in options) {
+      log5.debug(`call: model=${model} jsonSchema=${options.jsonSchemaName} via tool`);
+      const { toolName, tool, instruction } = buildStructuredJsonRequest({
+        instruction: options.instruction,
+        jsonSchemaName: options.jsonSchemaName,
+        jsonSchema: options.jsonSchema
+      });
+      const choice = await this.chatWithTools(model, {
+        caller: options.caller ?? options.jsonSchemaName,
+        instruction,
+        messages: [{ role: "user", content: options.message }],
+        tools: [tool]
+      });
+      return parseStructuredJsonResult(choice, toolName);
+    }
+    log5.debug(`call: model=${model} jsonSchema=- msgLen=${options.message.length}`);
     const body = {
       model,
       messages: [
@@ -2226,7 +2336,7 @@ class GitLabDuoExecutor extends LLMExecutor {
     if (!content) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? parseModelJson(content) : content;
+    return content;
   }
   async chatWithTools(model, options) {
     const log5 = logger.child("llm:gitlab-duo");
@@ -2325,10 +2435,25 @@ class SnowflakeCortexExecutor extends LLMExecutor {
     return JSON.parse(responseRaw);
   }
   async call(model, options) {
-    const jsonMode = "jsonSchemaName" in options;
-    const reasoning = defaultReasoningEffort(options.reasoningEffort, model, this.models.identity);
     const log5 = logger.child("llm:snowflake-cortex");
-    log5.debug(`call: model=${model} jsonSchema=${jsonMode ? options.jsonSchemaName : "-"} msgLen=${options.message.length}`);
+    if ("jsonSchemaName" in options) {
+      log5.debug(`call: model=${model} jsonSchema=${options.jsonSchemaName} via tool`);
+      const { toolName, tool, instruction } = buildStructuredJsonRequest({
+        instruction: options.instruction,
+        jsonSchemaName: options.jsonSchemaName,
+        jsonSchema: options.jsonSchema
+      });
+      const choice = await this.chatWithTools(model, {
+        caller: options.caller ?? options.jsonSchemaName,
+        instruction,
+        messages: [{ role: "user", content: options.message }],
+        tools: [tool],
+        reasoningEffort: "none"
+      });
+      return parseStructuredJsonResult(choice, toolName);
+    }
+    const reasoning = defaultReasoningEffort(options.reasoningEffort, model, this.models.identity);
+    log5.debug(`call: model=${model} jsonSchema=- msgLen=${options.message.length}`);
     const body = {
       model,
       messages: [
@@ -2347,7 +2472,7 @@ class SnowflakeCortexExecutor extends LLMExecutor {
     if (!content) {
       throw new Error("Empty response from model");
     }
-    return jsonMode ? parseModelJson(content) : content;
+    return content;
   }
   async chatWithTools(model, options) {
     const log5 = logger.child("llm:snowflake-cortex");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@p-sw/brainbox",
-  "version": "0.1.2-alpha.11",
+  "version": "0.1.2-alpha.12",
   "module": "dist/index.js",
   "main": "dist/index.js",
   "type": "module",