titan-agent 5.4.0 → 5.4.2

package/dist/providers/google.js

@@ -7,7 +7,89 @@ import logger from "../utils/logger.js";
 import { fetchWithRetry } from "../utils/helpers.js";
 import { resolveApiKey } from "./authResolver.js";
 import { v4 as uuid } from "uuid";
+import { clampMaxTokens } from "./modelCapabilities.js";
+import { mkdirSync, writeFileSync } from "fs";
+import { join } from "path";
+import { homedir } from "os";
 const COMPONENT = "Google";
+function shouldDumpRequestBody() {
+  if (process.env.GOOGLE_DUMP_REQUEST_BODY === "1" || process.env.GOOGLE_DUMP_REQUEST_BODY === "true") {
+    return true;
+  }
+  try {
+    const cfg = loadConfig();
+    const p = cfg.providers?.google;
+    return Boolean(p?.dumpRequestBody);
+  } catch {
+    return false;
+  }
+}
+const GEMINI_DEBUG_DIR = join(homedir(), ".titan", "debug", "gemini-requests");
+function dumpRequestBody(reason, body, extra) {
+  if (!shouldDumpRequestBody()) return;
+  try {
+    mkdirSync(GEMINI_DEBUG_DIR, { recursive: true });
+    const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
+    const path = join(GEMINI_DEBUG_DIR, `${stamp}-${reason}.json`);
+    writeFileSync(path, JSON.stringify({ reason, body, ...extra ?? {} }, null, 2));
+    logger.info(COMPONENT, `Dumped Gemini request body \u2192 ${path}`);
+  } catch (err) {
+    logger.warn(COMPONENT, `Failed to dump Gemini request body: ${err.message}`);
+  }
+}
+function buildContents(messages) {
+  const toolCallNameById = /* @__PURE__ */ new Map();
+  for (const m of messages) {
+    if (m.role === "assistant" && Array.isArray(m.toolCalls)) {
+      for (const tc of m.toolCalls) {
+        if (tc.id && tc.function?.name) {
+          toolCallNameById.set(tc.id, tc.function.name);
+        }
+      }
+    }
+  }
+  let corrections = 0;
+  const contents = [];
+  for (const m of messages.filter((x) => x.role !== "system")) {
+    if (m.role === "tool") {
+      const callId = m.toolCallId || "";
+      const recordedName = callId ? toolCallNameById.get(callId) : void 0;
+      const claimedName = (m.name || "").trim();
+      if (!recordedName) {
+        logger.warn(
+          COMPONENT,
+          `Malformed tool message: tool_call_id="${callId}" has no matching prior tool_call. name="${claimedName}". Dropping to prevent Gemini 400.`
+        );
+        corrections++;
+        continue;
+      }
+      const finalName = claimedName || recordedName;
+      if (!claimedName) {
+        logger.warn(
+          COMPONENT,
+          `Tool message missing name for tool_call_id="${callId}"; inferred "${finalName}" from assistant history.`
+        );
+        corrections++;
+      } else if (claimedName !== recordedName) {
+        logger.warn(
+          COMPONENT,
+          `Tool message name mismatch for tool_call_id="${callId}": claimed "${claimedName}" but tool_call recorded "${recordedName}". Using recorded.`
+        );
+        corrections++;
+      }
+      contents.push({
+        role: "function",
+        parts: [{ functionResponse: { name: recordedName, response: { result: m.content } } }]
+      });
+      continue;
+    }
+    contents.push({
+      role: m.role === "assistant" ? "model" : "user",
+      parts: [{ text: m.content }]
+    });
+  }
+  return { contents, corrections };
+}
 class GoogleProvider extends LLMProvider {
   name = "google";
   displayName = "Google (Gemini)";
@@ -22,22 +104,14 @@ class GoogleProvider extends LLMProvider {
     if (!apiKey) throw new Error("Google API key not configured");
     logger.debug(COMPONENT, `Chat request: model=${model}, messages=${options.messages.length}`);
     const systemInstruction = options.messages.find((m) => m.role === "system")?.content;
-    const contents = options.messages.filter((m) => m.role !== "system").map((m) => {
-      if (m.role === "tool") {
-        return {
-          role: "function",
-          parts: [{ functionResponse: { name: m.name || "tool", response: { result: m.content } } }]
-        };
-      }
-      return {
-        role: m.role === "assistant" ? "model" : "user",
-        parts: [{ text: m.content }]
-      };
-    });
+    const { contents, corrections } = buildContents(options.messages);
+    if (corrections > 0) {
+      logger.warn(COMPONENT, `Applied ${corrections} tool-message correction(s) before sending to Gemini.`);
+    }
     const body = {
       contents,
       generationConfig: {
-        maxOutputTokens: options.maxTokens || 8192,
+        maxOutputTokens: clampMaxTokens(options.model || "google/gemini-2.0-flash", options.maxTokens),
         temperature: options.temperature ?? 0.7
       }
     };
@@ -64,6 +138,7 @@ class GoogleProvider extends LLMProvider {
     });
     if (!response.ok) {
       const errorText = await response.text();
+      dumpRequestBody(`http-${response.status}`, body, { errorText, model });
       const { createProviderError } = await import("./errorTaxonomy.js");
       throw createProviderError("Google API", response, errorText, { provider: "google", model });
     }
@@ -112,15 +187,13 @@ class GoogleProvider extends LLMProvider {
       return;
     }
     const systemInstruction = options.messages.find((m) => m.role === "system")?.content;
-    const contents = options.messages.filter((m) => m.role !== "system").map((m) => {
-      if (m.role === "tool") {
-        return { role: "function", parts: [{ functionResponse: { name: m.name || "tool", response: { result: m.content } } }] };
-      }
-      return { role: m.role === "assistant" ? "model" : "user", parts: [{ text: m.content }] };
-    });
+    const { contents, corrections } = buildContents(options.messages);
+    if (corrections > 0) {
+      logger.warn(COMPONENT, `Applied ${corrections} tool-message correction(s) before streaming to Gemini.`);
+    }
     const body = {
       contents,
-      generationConfig: { maxOutputTokens: options.maxTokens || 8192, temperature: options.temperature ?? 0.7 }
+      generationConfig: { maxOutputTokens: clampMaxTokens(options.model || "google/gemini-2.0-flash", options.maxTokens), temperature: options.temperature ?? 0.7 }
     };
     if (systemInstruction) body.systemInstruction = { parts: [{ text: systemInstruction }] };
     if (options.tools && options.tools.length > 0) {
@@ -135,6 +208,7 @@ class GoogleProvider extends LLMProvider {
       });
       if (!response.ok || !response.body) {
         const errorText = await response.text();
+        dumpRequestBody(`stream-http-${response.status}`, body, { errorText, model });
         yield { type: "error", error: `Google API error (${response.status}): ${errorText}` };
         return;
       }

package/dist/providers/google.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/providers/google.ts"],"sourcesContent":["/**\n * TITAN — Google Gemini Provider\n */\nimport {\n    LLMProvider,\n    type ChatOptions,\n    type ChatResponse,\n    type ChatStreamChunk,\n    type ToolCall,\n} from './base.js';\nimport { loadConfig } from '../config/config.js';\nimport logger from '../utils/logger.js';\nimport { fetchWithRetry } from '../utils/helpers.js';\nimport { resolveApiKey } from './authResolver.js';\nimport { v4 as uuid } from 'uuid';\n\nconst COMPONENT = 'Google';\n\nexport class GoogleProvider extends LLMProvider {\n    readonly name = 'google';\n    readonly displayName = 'Google (Gemini)';\n\n    private get apiKey(): string {\n        const config = loadConfig();\n        const p = config.providers.google;\n        return resolveApiKey('google', p.authProfiles || [], p.apiKey || '', 'GOOGLE_API_KEY', p.rotationStrategy, p.credentialCooldownMs);\n    }\n\n    async chat(options: ChatOptions): Promise<ChatResponse> {\n        const model = (options.model || 'gemini-2.0-flash').replace('google/', '');\n        const apiKey = this.apiKey;\n        if (!apiKey) throw new Error('Google API key not configured');\n\n        logger.debug(COMPONENT, `Chat request: model=${model}, messages=${options.messages.length}`);\n\n        const systemInstruction = options.messages.find((m) => m.role === 'system')?.content;\n        const contents = options.messages\n            .filter((m) => m.role !== 'system')\n            .map((m) => {\n                if (m.role === 'tool') {\n                    return {\n                        role: 'function' as const,\n                        parts: [{ functionResponse: { name: m.name || 'tool', response: { result: m.content } } }],\n                    };\n                }\n                return {\n                    role: (m.role === 'assistant' ? 'model' : 'user') as string,\n                    parts: [{ text: m.content }],\n                };\n            });\n\n        const body: Record<string, unknown> = {\n            contents,\n            generationConfig: {\n                maxOutputTokens: options.maxTokens || 8192,\n                temperature: options.temperature ?? 0.7,\n            },\n        };\n\n        if (systemInstruction) {\n            body.systemInstruction = { parts: [{ text: systemInstruction }] };\n        }\n\n        if (options.tools && options.tools.length > 0) {\n            body.tools = [{\n                functionDeclarations: options.tools.map((t) => ({\n                    name: t.function.name,\n                    description: t.function.description,\n                    parameters: t.function.parameters,\n                })),\n            }];\n        }\n\n        const url = `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent`;\n        const response = await fetchWithRetry(url, {\n            method: 'POST',\n            headers: {\n                'Content-Type': 'application/json',\n                'x-goog-api-key': apiKey,\n            },\n            body: JSON.stringify(body),\n        });\n\n        if (!response.ok) {\n            const errorText = await response.text();\n            // Hunt Finding #37: attach status + Retry-After so the router can respect backoff\n            const { createProviderError } = await import('./errorTaxonomy.js');\n            throw createProviderError('Google API', response, errorText, { provider: 'google', model });\n        }\n\n        const data = await response.json() as Record<string, unknown>;\n        const candidates = data.candidates as Array<Record<string, unknown>>;\n\n        let textContent = '';\n        const toolCalls: ToolCall[] = [];\n\n        if (candidates && candidates.length > 0) {\n            const parts = (candidates[0].content as Record<string, unknown>)?.parts as Array<Record<string, unknown>> || [];\n            for (const part of parts) {\n                if (part.text) {\n                    textContent += part.text as string;\n                }\n                if (part.functionCall) {\n                    const fc = part.functionCall as Record<string, unknown>;\n                    toolCalls.push({\n                        id: uuid(),\n                        type: 'function',\n                        function: {\n                            name: fc.name as string,\n                            arguments: JSON.stringify(fc.args),\n                        },\n                    });\n                }\n            }\n        }\n\n        const usageMeta = data.usageMetadata as { promptTokenCount?: number; candidatesTokenCount?: number; totalTokenCount?: number } | undefined;\n\n        return {\n            id: uuid(),\n            content: textContent,\n            toolCalls: toolCalls.length > 0 ? toolCalls : undefined,\n            usage: usageMeta\n                ? {\n                    promptTokens: usageMeta.promptTokenCount || 0,\n                    completionTokens: usageMeta.candidatesTokenCount || 0,\n                    totalTokens: usageMeta.totalTokenCount || 0,\n                }\n                : undefined,\n            finishReason: toolCalls.length > 0 ? 'tool_calls' : 'stop',\n            model: `google/${model}`,\n        };\n    }\n\n    async *chatStream(options: ChatOptions): AsyncGenerator<ChatStreamChunk> {\n        const model = (options.model || 'gemini-2.0-flash').replace('google/', '');\n        const apiKey = this.apiKey;\n        if (!apiKey) { yield { type: 'error', error: 'Google API key not configured' }; return; }\n\n        const systemInstruction = options.messages.find((m) => m.role === 'system')?.content;\n        const contents = options.messages.filter((m) => m.role !== 'system').map((m) => {\n            if (m.role === 'tool') {\n                return { role: 'function' as const, parts: [{ functionResponse: { name: m.name || 'tool', response: { result: m.content } } }] };\n            }\n            return { role: (m.role === 'assistant' ? 'model' : 'user') as string, parts: [{ text: m.content }] };\n        });\n\n        const body: Record<string, unknown> = {\n            contents,\n            generationConfig: { maxOutputTokens: options.maxTokens || 8192, temperature: options.temperature ?? 0.7 },\n        };\n        if (systemInstruction) body.systemInstruction = { parts: [{ text: systemInstruction }] };\n        if (options.tools && options.tools.length > 0) {\n            body.tools = [{ functionDeclarations: options.tools.map((t) => ({ name: t.function.name, description: t.function.description, parameters: t.function.parameters })) }];\n        }\n\n        try {\n            const url = `https://generativelanguage.googleapis.com/v1beta/models/${model}:streamGenerateContent?alt=sse`;\n            const response = await fetch(url, {\n                method: 'POST',\n                headers: { 'Content-Type': 'application/json', 'x-goog-api-key': apiKey },\n                body: JSON.stringify(body),\n            });\n\n            if (!response.ok || !response.body) {\n                const errorText = await response.text();\n                yield { type: 'error', error: `Google API error (${response.status}): ${errorText}` };\n                return;\n            }\n\n            const reader = response.body.getReader();\n            const decoder = new TextDecoder();\n            let buffer = '';\n\n            while (true) {\n                const { done, value } = await reader.read();\n                if (done) break;\n                buffer += decoder.decode(value, { stream: true });\n\n                const lines = buffer.split('\\n');\n                buffer = lines.pop() || '';\n\n                for (const line of lines) {\n                    if (!line.startsWith('data: ')) continue;\n                    const json = line.slice(6).trim();\n                    if (!json) continue;\n\n                    try {\n                        const chunk = JSON.parse(json);\n                        const candidates = chunk.candidates as Array<Record<string, unknown>> | undefined;\n                        if (candidates && candidates.length > 0) {\n                            const parts = (candidates[0].content as Record<string, unknown>)?.parts as Array<Record<string, unknown>> || [];\n                            for (const part of parts) {\n                                if (part.text) yield { type: 'text', content: part.text as string };\n                                if (part.functionCall) {\n                                    const fc = part.functionCall as Record<string, unknown>;\n                                    yield {\n                                        type: 'tool_call',\n                                        toolCall: { id: uuid(), type: 'function', function: { name: fc.name as string, arguments: JSON.stringify(fc.args) } },\n                                    };\n                                }\n                            }\n                        }\n                    } catch { /* skip malformed SSE lines */ }\n                }\n            }\n            yield { type: 'done' };\n        } catch (error) {\n            yield { type: 'error', error: (error as Error).message };\n        }\n    }\n\n    async listModels(): Promise<string[]> {\n        return ['gemini-2.5-pro', 'gemini-2.5-flash', 'gemini-2.0-flash', 'gemini-1.5-pro'];\n    }\n\n    async healthCheck(): Promise<boolean> {\n        try {\n            if (!this.apiKey) return false;\n            const url = `https://generativelanguage.googleapis.com/v1beta/models`;\n            const response = await fetch(url, {\n                headers: { 'x-goog-api-key': this.apiKey },\n            });\n            return response.ok;\n        } catch {\n            return false;\n        }\n    }\n}\n"],"mappings":";AAGA;AAAA,EACI;AAAA,OAKG;AACP,SAAS,kBAAkB;AAC3B,OAAO,YAAY;AACnB,SAAS,sBAAsB;AAC/B,SAAS,qBAAqB;AAC9B,SAAS,MAAM,YAAY;AAE3B,MAAM,YAAY;AAEX,MAAM,uBAAuB,YAAY;AAAA,EACnC,OAAO;AAAA,EACP,cAAc;AAAA,EAEvB,IAAY,SAAiB;AACzB,UAAM,SAAS,WAAW;AAC1B,UAAM,IAAI,OAAO,UAAU;AAC3B,WAAO,cAAc,UAAU,EAAE,gBAAgB,CAAC,GAAG,EAAE,UAAU,IAAI,kBAAkB,EAAE,kBAAkB,EAAE,oBAAoB;AAAA,EACrI;AAAA,EAEA,MAAM,KAAK,SAA6C;AACpD,UAAM,SAAS,QAAQ,SAAS,oBAAoB,QAAQ,WAAW,EAAE;AACzE,UAAM,SAAS,KAAK;AACpB,QAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,+BAA+B;AAE5D,WAAO,MAAM,WAAW,uBAAuB,KAAK,cAAc,QAAQ,SAAS,MAAM,EAAE;AAE3F,UAAM,oBAAoB,QAAQ,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AAC7E,UAAM,WAAW,QAAQ,SACpB,OAAO,CAAC,MAAM,EAAE,SAAS,QAAQ,EACjC,IAAI,CAAC,MAAM;AACR,UAAI,EAAE,SAAS,QAAQ;AACnB,eAAO;AAAA,UACH,MAAM;AAAA,UACN,OAAO,CAAC,EAAE,kBAAkB,EAAE,MAAM,EAAE,QAAQ,QAAQ,UAAU,EAAE,QAAQ,EAAE,QAAQ,EAAE,EAAE,CAAC;AAAA,QAC7F;AAAA,MACJ;AACA,aAAO;AAAA,QACH,MAAO,EAAE,SAAS,cAAc,UAAU;AAAA,QAC1C,OAAO,CAAC,EAAE,MAAM,EAAE,QAAQ,CAAC;AAAA,MAC/B;AAAA,IACJ,CAAC;AAEL,UAAM,OAAgC;AAAA,MAClC;AAAA,MACA,kBAAkB;AAAA,QACd,iBAAiB,QAAQ,aAAa;AAAA,QACtC,aAAa,QAAQ,eAAe;AAAA,MACxC;AAAA,IACJ;AAEA,QAAI,mBAAmB;AACnB,WAAK,oBAAoB,EAAE,OAAO,CAAC,EAAE,MAAM,kBAAkB,CAAC,EAAE;AAAA,IACpE;AAEA,QAAI,QAAQ,SAAS,QAAQ,MAAM,SAAS,GAAG;AAC3C,WAAK,QAAQ,CAAC;AAAA,QACV,sBAAsB,QAAQ,MAAM,IAAI,CAAC,OAAO;AAAA,UAC5C,MAAM,EAAE,SAAS;AAAA,UACjB,aAAa,EAAE,SAAS;AAAA,UACxB,YAAY,EAAE,SAAS;AAAA,QAC3B,EAAE;AAAA,MACN,CAAC;AAAA,IACL;AAEA,UAAM,MAAM,2DAA2D,KAAK;AAC5E,UAAM,WAAW,MAAM,eAAe,KAAK;AAAA,MACvC,QAAQ;AAAA,MACR,SAAS;AAAA,QACL,gBAAgB;AAAA,QAChB,kBAAkB;AAAA,MACtB;AAAA,MACA,MAAM,KAAK,UAAU,IAAI;AAAA,IAC7B,CAAC;AAED,QAAI,CAAC,SAAS,IAAI;AACd,YAAM,YAAY,MAAM,SAAS,KAAK;AAEtC,YAAM,EAAE,oBAAoB,IAAI,MAAM,OAAO,oBAAoB;AACjE,YAAM,oBAAoB,cAAc,UAAU,WAAW,EAAE,UAAU,UAAU,MAAM,CAAC;AAAA,IAC9F;AAEA,UAAM,OAAO,MAAM,SAAS,KAAK;AACjC,UAAM,aAAa,KAAK;AAExB,QAAI,cAAc;AAClB,UAAM,YAAwB,CAAC;AAE/B,QAAI,cAAc,WAAW,SAAS,GAAG;AACrC,YAAM,QAAS,WAAW,CAAC,EAAE,SAAqC,SAA2C,CAAC;AAC9G,iBAAW,QAAQ,OAAO;AACtB,YAAI,KAAK,MAAM;AACX,yBAAe,KAAK;AAAA,QACxB;AACA,YAAI,KAAK,cAAc;AACnB,gBAAM,KAAK,KAAK;AAChB,oBAAU,KAAK;AAAA,YACX,IAAI,KAAK;AAAA,YACT,MAAM;AAAA,YACN,UAAU;AAAA,cACN,MAAM,GAAG;AAAA,cACT,WAAW,KAAK,UAAU,GAAG,IAAI;AAAA,YACrC;AAAA,UACJ,CAAC;AAAA,QACL;AAAA,MACJ;AAAA,IACJ;AAEA,UAAM,YAAY,KAAK;AAEvB,WAAO;AAAA,MACH,IAAI,KAAK;AAAA,MACT,SAAS;AAAA,MACT,WAAW,UAAU,SAAS,IAAI,YAAY;AAAA,MAC9C,OAAO,YACD;AAAA,QACE,cAAc,UAAU,oBAAoB;AAAA,QAC5C,kBAAkB,UAAU,wBAAwB;AAAA,QACpD,aAAa,UAAU,mBAAmB;AAAA,MAC9C,IACE;AAAA,MACN,cAAc,UAAU,SAAS,IAAI,eAAe;AAAA,MACpD,OAAO,UAAU,KAAK;AAAA,IAC1B;AAAA,EACJ;AAAA,EAEA,OAAO,WAAW,SAAuD;AACrE,UAAM,SAAS,QAAQ,SAAS,oBAAoB,QAAQ,WAAW,EAAE;AACzE,UAAM,SAAS,KAAK;AACpB,QAAI,CAAC,QAAQ;AAAE,YAAM,EAAE,MAAM,SAAS,OAAO,gCAAgC;AAAG;AAAA,IAAQ;AAExF,UAAM,oBAAoB,QAAQ,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AAC7E,UAAM,WAAW,QAAQ,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,QAAQ,EAAE,IAAI,CAAC,MAAM;AAC5E,UAAI,EAAE,SAAS,QAAQ;AACnB,eAAO,EAAE,MAAM,YAAqB,OAAO,CAAC,EAAE,kBAAkB,EAAE,MAAM,EAAE,QAAQ,QAAQ,UAAU,EAAE,QAAQ,EAAE,QAAQ,EAAE,EAAE,CAAC,EAAE;AAAA,MACnI;AACA,aAAO,EAAE,MAAO,EAAE,SAAS,cAAc,UAAU,QAAmB,OAAO,CAAC,EAAE,MAAM,EAAE,QAAQ,CAAC,EAAE;AAAA,IACvG,CAAC;AAED,UAAM,OAAgC;AAAA,MAClC;AAAA,MACA,kBAAkB,EAAE,iBAAiB,QAAQ,aAAa,MAAM,aAAa,QAAQ,eAAe,IAAI;AAAA,IAC5G;AACA,QAAI,kBAAmB,MAAK,oBAAoB,EAAE,OAAO,CAAC,EAAE,MAAM,kBAAkB,CAAC,EAAE;AACvF,QAAI,QAAQ,SAAS,QAAQ,MAAM,SAAS,GAAG;AAC3C,WAAK,QAAQ,CAAC,EAAE,sBAAsB,QAAQ,MAAM,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,SAAS,MAAM,aAAa,EAAE,SAAS,aAAa,YAAY,EAAE,SAAS,WAAW,EAAE,EAAE,CAAC;AAAA,IACzK;AAEA,QAAI;AACA,YAAM,MAAM,2DAA2D,KAAK;AAC5E,YAAM,WAAW,MAAM,MAAM,KAAK;AAAA,QAC9B,QAAQ;AAAA,QACR,SAAS,EAAE,gBAAgB,oBAAoB,kBAAkB,OAAO;AAAA,QACxE,MAAM,KAAK,UAAU,IAAI;AAAA,MAC7B,CAAC;AAED,UAAI,CAAC,SAAS,MAAM,CAAC,SAAS,MAAM;AAChC,cAAM,YAAY,MAAM,SAAS,KAAK;AACtC,cAAM,EAAE,MAAM,SAAS,OAAO,qBAAqB,SAAS,MAAM,MAAM,SAAS,GAAG;AACpF;AAAA,MACJ;AAEA,YAAM,SAAS,SAAS,KAAK,UAAU;AACvC,YAAM,UAAU,IAAI,YAAY;AAChC,UAAI,SAAS;AAEb,aAAO,MAAM;AACT,cAAM,EAAE,MAAM,MAAM,IAAI,MAAM,OAAO,KAAK;AAC1C,YAAI,KAAM;AACV,kBAAU,QAAQ,OAAO,OAAO,EAAE,QAAQ,KAAK,CAAC;AAEhD,cAAM,QAAQ,OAAO,MAAM,IAAI;AAC/B,iBAAS,MAAM,IAAI,KAAK;AAExB,mBAAW,QAAQ,OAAO;AACtB,cAAI,CAAC,KAAK,WAAW,QAAQ,EAAG;AAChC,gBAAM,OAAO,KAAK,MAAM,CAAC,EAAE,KAAK;AAChC,cAAI,CAAC,KAAM;AAEX,cAAI;AACA,kBAAM,QAAQ,KAAK,MAAM,IAAI;AAC7B,kBAAM,aAAa,MAAM;AACzB,gBAAI,cAAc,WAAW,SAAS,GAAG;AACrC,oBAAM,QAAS,WAAW,CAAC,EAAE,SAAqC,SAA2C,CAAC;AAC9G,yBAAW,QAAQ,OAAO;AACtB,oBAAI,KAAK,KAAM,OAAM,EAAE,MAAM,QAAQ,SAAS,KAAK,KAAe;AAClE,oBAAI,KAAK,cAAc;AACnB,wBAAM,KAAK,KAAK;AAChB,wBAAM;AAAA,oBACF,MAAM;AAAA,oBACN,UAAU,EAAE,IAAI,KAAK,GAAG,MAAM,YAAY,UAAU,EAAE,MAAM,GAAG,MAAgB,WAAW,KAAK,UAAU,GAAG,IAAI,EAAE,EAAE;AAAA,kBACxH;AAAA,gBACJ;AAAA,cACJ;AAAA,YACJ;AAAA,UACJ,QAAQ;AAAA,UAAiC;AAAA,QAC7C;AAAA,MACJ;AACA,YAAM,EAAE,MAAM,OAAO;AAAA,IACzB,SAAS,OAAO;AACZ,YAAM,EAAE,MAAM,SAAS,OAAQ,MAAgB,QAAQ;AAAA,IAC3D;AAAA,EACJ;AAAA,EAEA,MAAM,aAAgC;AAClC,WAAO,CAAC,kBAAkB,oBAAoB,oBAAoB,gBAAgB;AAAA,EACtF;AAAA,EAEA,MAAM,cAAgC;AAClC,QAAI;AACA,UAAI,CAAC,KAAK,OAAQ,QAAO;AACzB,YAAM,MAAM;AACZ,YAAM,WAAW,MAAM,MAAM,KAAK;AAAA,QAC9B,SAAS,EAAE,kBAAkB,KAAK,OAAO;AAAA,MAC7C,CAAC;AACD,aAAO,SAAS;AAAA,IACpB,QAAQ;AACJ,aAAO;AAAA,IACX;AAAA,EACJ;AACJ;","names":[]}
+{"version":3,"sources":["../../src/providers/google.ts"],"sourcesContent":["/**\n * TITAN — Google Gemini Provider\n */\nimport {\n    LLMProvider,\n    type ChatOptions,\n    type ChatMessage,\n    type ChatResponse,\n    type ChatStreamChunk,\n    type ToolCall,\n} from './base.js';\nimport { loadConfig } from '../config/config.js';\nimport logger from '../utils/logger.js';\nimport { fetchWithRetry } from '../utils/helpers.js';\nimport { resolveApiKey } from './authResolver.js';\nimport { v4 as uuid } from 'uuid';\nimport { clampMaxTokens } from './modelCapabilities.js';\nimport { mkdirSync, writeFileSync } from 'fs';\nimport { join } from 'path';\nimport { homedir } from 'os';\n\nconst COMPONENT = 'Google';\n\n/**\n * When true, every Gemini request body that fails serialization-validation OR\n * gets a non-2xx response is dumped to ~/.titan/debug/gemini-requests/ for\n * post-mortem. Toggled via `GOOGLE_DUMP_REQUEST_BODY=1` env var or the\n * provider's `dumpRequestBody` config flag — keeps it off by default since\n * each dump is a JSON file with full prompt content.\n */\nfunction shouldDumpRequestBody(): boolean {\n    if (process.env.GOOGLE_DUMP_REQUEST_BODY === '1' || process.env.GOOGLE_DUMP_REQUEST_BODY === 'true') {\n        return true;\n    }\n    try {\n        const cfg = loadConfig();\n        const p = (cfg.providers as Record<string, unknown> | undefined)?.google as\n            | { dumpRequestBody?: boolean }\n            | undefined;\n        return Boolean(p?.dumpRequestBody);\n    } catch {\n        return false;\n    }\n}\n\nconst GEMINI_DEBUG_DIR = join(homedir(), '.titan', 'debug', 'gemini-requests');\n\nfunction dumpRequestBody(reason: string, body: unknown, extra?: Record<string, unknown>): void {\n    if (!shouldDumpRequestBody()) return;\n    try {\n        mkdirSync(GEMINI_DEBUG_DIR, { recursive: true });\n        const stamp = new Date().toISOString().replace(/[:.]/g, '-');\n        const path = join(GEMINI_DEBUG_DIR, `${stamp}-${reason}.json`);\n        writeFileSync(path, JSON.stringify({ reason, body, ...(extra ?? {}) }, null, 2));\n        logger.info(COMPONENT, `Dumped Gemini request body → ${path}`);\n    } catch (err) {\n        logger.warn(COMPONENT, `Failed to dump Gemini request body: ${(err as Error).message}`);\n    }\n}\n\n/**\n * Build the Gemini `contents[]` array from TITAN ChatMessages, with strict\n * pre-serialization validation of `tool` messages.\n *\n * Why this matters:\n *   Gemini's `functionResponse` requires a non-empty `name` field paired with\n *   a valid `tool_call_id` from a prior assistant turn. If the agent loop\n *   ever emits a tool result whose corresponding tool call cannot be located\n *   in conversation history, Gemini rejects the whole request with a 400 and\n *   the error message is opaque (\"function_response without function_call\").\n *\n *   Rather than push the malformed message and let Gemini blow up, we:\n *     1. Build a map of every tool_call.id → name from prior assistant messages.\n *     2. For each `tool` message, ensure (a) the name is non-empty (use the\n *        toolCallId map as a backstop) and (b) the toolCallId references a\n *        known prior call.\n *     3. Drop or relabel messages that fail validation, with a warning that\n *        names the offending message so it shows up in logs.\n *     4. If `dumpRequestBody` is enabled, write the full pre-validation body\n *        to disk for inspection before any silent corrections.\n */\nfunction buildContents(messages: ChatMessage[]): { contents: Array<Record<string, unknown>>; corrections: number } {\n    // Pass 1: build a lookup of valid tool_call_id → function name from\n    // every prior assistant turn that emitted toolCalls.\n    const toolCallNameById = new Map<string, string>();\n    for (const m of messages) {\n        if (m.role === 'assistant' && Array.isArray(m.toolCalls)) {\n            for (const tc of m.toolCalls) {\n                if (tc.id && tc.function?.name) {\n                    toolCallNameById.set(tc.id, tc.function.name);\n                }\n            }\n        }\n    }\n\n    let corrections = 0;\n    const contents: Array<Record<string, unknown>> = [];\n\n    for (const m of messages.filter((x) => x.role !== 'system')) {\n        if (m.role === 'tool') {\n            // Validation: name must be non-empty AND toolCallId must reference\n            // a known prior call. Either failure → log + best-effort repair.\n            const callId = m.toolCallId || '';\n            const recordedName = callId ? toolCallNameById.get(callId) : undefined;\n            const claimedName = (m.name || '').trim();\n\n            if (!recordedName) {\n                logger.warn(\n                    COMPONENT,\n                    `Malformed tool message: tool_call_id=\"${callId}\" has no matching prior tool_call. ` +\n                    `name=\"${claimedName}\". Dropping to prevent Gemini 400.`,\n                );\n                corrections++;\n                continue;\n            }\n\n            const finalName = claimedName || recordedName;\n            if (!claimedName) {\n                logger.warn(\n                    COMPONENT,\n                    `Tool message missing name for tool_call_id=\"${callId}\"; ` +\n                    `inferred \"${finalName}\" from assistant history.`,\n                );\n                corrections++;\n            } else if (claimedName !== recordedName) {\n                logger.warn(\n                    COMPONENT,\n                    `Tool message name mismatch for tool_call_id=\"${callId}\": ` +\n                    `claimed \"${claimedName}\" but tool_call recorded \"${recordedName}\". Using recorded.`,\n                );\n                corrections++;\n            }\n\n            contents.push({\n                role: 'function' as const,\n                parts: [{ functionResponse: { name: recordedName, response: { result: m.content } } }],\n            });\n            continue;\n        }\n\n        contents.push({\n            role: (m.role === 'assistant' ? 'model' : 'user') as string,\n            parts: [{ text: m.content }],\n        });\n    }\n\n    return { contents, corrections };\n}\n\nexport class GoogleProvider extends LLMProvider {\n    readonly name = 'google';\n    readonly displayName = 'Google (Gemini)';\n\n    private get apiKey(): string {\n        const config = loadConfig();\n        const p = config.providers.google;\n        return resolveApiKey('google', p.authProfiles || [], p.apiKey || '', 'GOOGLE_API_KEY', p.rotationStrategy, p.credentialCooldownMs);\n    }\n\n    async chat(options: ChatOptions): Promise<ChatResponse> {\n        const model = (options.model || 'gemini-2.0-flash').replace('google/', '');\n        const apiKey = this.apiKey;\n        if (!apiKey) throw new Error('Google API key not configured');\n\n        logger.debug(COMPONENT, `Chat request: model=${model}, messages=${options.messages.length}`);\n\n        const systemInstruction = options.messages.find((m) => m.role === 'system')?.content;\n        const { contents, corrections } = buildContents(options.messages);\n        if (corrections > 0) {\n            logger.warn(COMPONENT, `Applied ${corrections} tool-message correction(s) before sending to Gemini.`);\n        }\n\n        const body: Record<string, unknown> = {\n            contents,\n            generationConfig: {\n                maxOutputTokens: clampMaxTokens(options.model || 'google/gemini-2.0-flash', options.maxTokens),\n                temperature: options.temperature ?? 0.7,\n            },\n        };\n\n        if (systemInstruction) {\n            body.systemInstruction = { parts: [{ text: systemInstruction }] };\n        }\n\n        if (options.tools && options.tools.length > 0) {\n            body.tools = [{\n                functionDeclarations: options.tools.map((t) => ({\n                    name: t.function.name,\n                    description: t.function.description,\n                    parameters: t.function.parameters,\n                })),\n            }];\n        }\n\n        const url = `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent`;\n        const response = await fetchWithRetry(url, {\n            method: 'POST',\n            headers: {\n                'Content-Type': 'application/json',\n                'x-goog-api-key': apiKey,\n            },\n            body: JSON.stringify(body),\n        });\n\n        if (!response.ok) {\n            const errorText = await response.text();\n            // Dump body when the API rejected it so post-mortem has full context\n            dumpRequestBody(`http-${response.status}`, body, { errorText, model });\n            // Hunt Finding #37: attach status + Retry-After so the router can respect backoff\n            const { createProviderError } = await import('./errorTaxonomy.js');\n            throw createProviderError('Google API', response, errorText, { provider: 'google', model });\n        }\n\n        const data = await response.json() as Record<string, unknown>;\n        const candidates = data.candidates as Array<Record<string, unknown>>;\n\n        let textContent = '';\n        const toolCalls: ToolCall[] = [];\n\n        if (candidates && candidates.length > 0) {\n            const parts = (candidates[0].content as Record<string, unknown>)?.parts as Array<Record<string, unknown>> || [];\n            for (const part of parts) {\n                if (part.text) {\n                    textContent += part.text as string;\n                }\n                if (part.functionCall) {\n                    const fc = part.functionCall as Record<string, unknown>;\n                    toolCalls.push({\n                        id: uuid(),\n                        type: 'function',\n                        function: {\n                            name: fc.name as string,\n                            arguments: JSON.stringify(fc.args),\n                        },\n                    });\n                }\n            }\n        }\n\n        const usageMeta = data.usageMetadata as { promptTokenCount?: number; candidatesTokenCount?: number; totalTokenCount?: number } | undefined;\n\n        return {\n            id: uuid(),\n            content: textContent,\n            toolCalls: toolCalls.length > 0 ? toolCalls : undefined,\n            usage: usageMeta\n                ? {\n                    promptTokens: usageMeta.promptTokenCount || 0,\n                    completionTokens: usageMeta.candidatesTokenCount || 0,\n                    totalTokens: usageMeta.totalTokenCount || 0,\n                }\n                : undefined,\n            finishReason: toolCalls.length > 0 ? 'tool_calls' : 'stop',\n            model: `google/${model}`,\n        };\n    }\n\n    async *chatStream(options: ChatOptions): AsyncGenerator<ChatStreamChunk> {\n        const model = (options.model || 'gemini-2.0-flash').replace('google/', '');\n        const apiKey = this.apiKey;\n        if (!apiKey) { yield { type: 'error', error: 'Google API key not configured' }; return; }\n\n        const systemInstruction = options.messages.find((m) => m.role === 'system')?.content;\n        const { contents, corrections } = buildContents(options.messages);\n        if (corrections > 0) {\n            logger.warn(COMPONENT, `Applied ${corrections} tool-message correction(s) before streaming to Gemini.`);\n        }\n\n        const body: Record<string, unknown> = {\n            contents,\n            generationConfig: { maxOutputTokens: clampMaxTokens(options.model || 'google/gemini-2.0-flash', options.maxTokens), temperature: options.temperature ?? 0.7 },\n        };\n        if (systemInstruction) body.systemInstruction = { parts: [{ text: systemInstruction }] };\n        if (options.tools && options.tools.length > 0) {\n            body.tools = [{ functionDeclarations: options.tools.map((t) => ({ name: t.function.name, description: t.function.description, parameters: t.function.parameters })) }];\n        }\n\n        try {\n            const url = `https://generativelanguage.googleapis.com/v1beta/models/${model}:streamGenerateContent?alt=sse`;\n            const response = await fetch(url, {\n                method: 'POST',\n                headers: { 'Content-Type': 'application/json', 'x-goog-api-key': apiKey },\n                body: JSON.stringify(body),\n            });\n\n            if (!response.ok || !response.body) {\n                const errorText = await response.text();\n                dumpRequestBody(`stream-http-${response.status}`, body, { errorText, model });\n                yield { type: 'error', error: `Google API error (${response.status}): ${errorText}` };\n                return;\n            }\n\n            const reader = response.body.getReader();\n            const decoder = new TextDecoder();\n            let buffer = '';\n\n            while (true) {\n                const { done, value } = await reader.read();\n                if (done) break;\n                buffer += decoder.decode(value, { stream: true });\n\n                const lines = buffer.split('\\n');\n                buffer = lines.pop() || '';\n\n                for (const line of lines) {\n                    if (!line.startsWith('data: ')) continue;\n                    const json = line.slice(6).trim();\n                    if (!json) continue;\n\n                    try {\n                        const chunk = JSON.parse(json);\n                        const candidates = chunk.candidates as Array<Record<string, unknown>> | undefined;\n                        if (candidates && candidates.length > 0) {\n                            const parts = (candidates[0].content as Record<string, unknown>)?.parts as Array<Record<string, unknown>> || [];\n                            for (const part of parts) {\n                                if (part.text) yield { type: 'text', content: part.text as string };\n                                if (part.functionCall) {\n                                    const fc = part.functionCall as Record<string, unknown>;\n                                    yield {\n                                        type: 'tool_call',\n                                        toolCall: { id: uuid(), type: 'function', function: { name: fc.name as string, arguments: JSON.stringify(fc.args) } },\n                                    };\n                                }\n                            }\n                        }\n                    } catch { /* skip malformed SSE lines */ }\n                }\n            }\n            yield { type: 'done' };\n        } catch (error) {\n            yield { type: 'error', error: (error as Error).message };\n        }\n    }\n\n    async listModels(): Promise<string[]> {\n        return ['gemini-2.5-pro', 'gemini-2.5-flash', 'gemini-2.0-flash', 'gemini-1.5-pro'];\n    }\n\n    async healthCheck(): Promise<boolean> {\n        try {\n            if (!this.apiKey) return false;\n            const url = `https://generativelanguage.googleapis.com/v1beta/models`;\n            const response = await fetch(url, {\n                headers: { 'x-goog-api-key': this.apiKey },\n            });\n            return response.ok;\n        } catch {\n            return false;\n        }\n    }\n}\n"],"mappings":";AAGA;AAAA,EACI;AAAA,OAMG;AACP,SAAS,kBAAkB;AAC3B,OAAO,YAAY;AACnB,SAAS,sBAAsB;AAC/B,SAAS,qBAAqB;AAC9B,SAAS,MAAM,YAAY;AAC3B,SAAS,sBAAsB;AAC/B,SAAS,WAAW,qBAAqB;AACzC,SAAS,YAAY;AACrB,SAAS,eAAe;AAExB,MAAM,YAAY;AASlB,SAAS,wBAAiC;AACtC,MAAI,QAAQ,IAAI,6BAA6B,OAAO,QAAQ,IAAI,6BAA6B,QAAQ;AACjG,WAAO;AAAA,EACX;AACA,MAAI;AACA,UAAM,MAAM,WAAW;AACvB,UAAM,IAAK,IAAI,WAAmD;AAGlE,WAAO,QAAQ,GAAG,eAAe;AAAA,EACrC,QAAQ;AACJ,WAAO;AAAA,EACX;AACJ;AAEA,MAAM,mBAAmB,KAAK,QAAQ,GAAG,UAAU,SAAS,iBAAiB;AAE7E,SAAS,gBAAgB,QAAgB,MAAe,OAAuC;AAC3F,MAAI,CAAC,sBAAsB,EAAG;AAC9B,MAAI;AACA,cAAU,kBAAkB,EAAE,WAAW,KAAK,CAAC;AAC/C,UAAM,SAAQ,oBAAI,KAAK,GAAE,YAAY,EAAE,QAAQ,SAAS,GAAG;AAC3D,UAAM,OAAO,KAAK,kBAAkB,GAAG,KAAK,IAAI,MAAM,OAAO;AAC7D,kBAAc,MAAM,KAAK,UAAU,EAAE,QAAQ,MAAM,GAAI,SAAS,CAAC,EAAG,GAAG,MAAM,CAAC,CAAC;AAC/E,WAAO,KAAK,WAAW,qCAAgC,IAAI,EAAE;AAAA,EACjE,SAAS,KAAK;AACV,WAAO,KAAK,WAAW,uCAAwC,IAAc,OAAO,EAAE;AAAA,EAC1F;AACJ;AAuBA,SAAS,cAAc,UAA4F;AAG/G,QAAM,mBAAmB,oBAAI,IAAoB;AACjD,aAAW,KAAK,UAAU;AACtB,QAAI,EAAE,SAAS,eAAe,MAAM,QAAQ,EAAE,SAAS,GAAG;AACtD,iBAAW,MAAM,EAAE,WAAW;AAC1B,YAAI,GAAG,MAAM,GAAG,UAAU,MAAM;AAC5B,2BAAiB,IAAI,GAAG,IAAI,GAAG,SAAS,IAAI;AAAA,QAChD;AAAA,MACJ;AAAA,IACJ;AAAA,EACJ;AAEA,MAAI,cAAc;AAClB,QAAM,WAA2C,CAAC;AAElD,aAAW,KAAK,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACzD,QAAI,EAAE,SAAS,QAAQ;AAGnB,YAAM,SAAS,EAAE,cAAc;AAC/B,YAAM,eAAe,SAAS,iBAAiB,IAAI,MAAM,IAAI;AAC7D,YAAM,eAAe,EAAE,QAAQ,IAAI,KAAK;AAExC,UAAI,CAAC,cAAc;AACf,eAAO;AAAA,UACH;AAAA,UACA,yCAAyC,MAAM,4CACtC,WAAW;AAAA,QACxB;AACA;AACA;AAAA,MACJ;AAEA,YAAM,YAAY,eAAe;AACjC,UAAI,CAAC,aAAa;AACd,eAAO;AAAA,UACH;AAAA,UACA,+CAA+C,MAAM,gBACxC,SAAS;AAAA,QAC1B;AACA;AAAA,MACJ,WAAW,gBAAgB,cAAc;AACrC,eAAO;AAAA,UACH;AAAA,UACA,gDAAgD,MAAM,eAC1C,WAAW,6BAA6B,YAAY;AAAA,QACpE;AACA;AAAA,MACJ;AAEA,eAAS,KAAK;AAAA,QACV,MAAM;AAAA,QACN,OAAO,CAAC,EAAE,kBAAkB,EAAE,MAAM,cAAc,UAAU,EAAE,QAAQ,EAAE,QAAQ,EAAE,EAAE,CAAC;AAAA,MACzF,CAAC;AACD;AAAA,IACJ;AAEA,aAAS,KAAK;AAAA,MACV,MAAO,EAAE,SAAS,cAAc,UAAU;AAAA,MAC1C,OAAO,CAAC,EAAE,MAAM,EAAE,QAAQ,CAAC;AAAA,IAC/B,CAAC;AAAA,EACL;AAEA,SAAO,EAAE,UAAU,YAAY;AACnC;AAEO,MAAM,uBAAuB,YAAY;AAAA,EACnC,OAAO;AAAA,EACP,cAAc;AAAA,EAEvB,IAAY,SAAiB;AACzB,UAAM,SAAS,WAAW;AAC1B,UAAM,IAAI,OAAO,UAAU;AAC3B,WAAO,cAAc,UAAU,EAAE,gBAAgB,CAAC,GAAG,EAAE,UAAU,IAAI,kBAAkB,EAAE,kBAAkB,EAAE,oBAAoB;AAAA,EACrI;AAAA,EAEA,MAAM,KAAK,SAA6C;AACpD,UAAM,SAAS,QAAQ,SAAS,oBAAoB,QAAQ,WAAW,EAAE;AACzE,UAAM,SAAS,KAAK;AACpB,QAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,+BAA+B;AAE5D,WAAO,MAAM,WAAW,uBAAuB,KAAK,cAAc,QAAQ,SAAS,MAAM,EAAE;AAE3F,UAAM,oBAAoB,QAAQ,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AAC7E,UAAM,EAAE,UAAU,YAAY,IAAI,cAAc,QAAQ,QAAQ;AAChE,QAAI,cAAc,GAAG;AACjB,aAAO,KAAK,WAAW,WAAW,WAAW,uDAAuD;AAAA,IACxG;AAEA,UAAM,OAAgC;AAAA,MAClC;AAAA,MACA,kBAAkB;AAAA,QACd,iBAAiB,eAAe,QAAQ,SAAS,2BAA2B,QAAQ,SAAS;AAAA,QAC7F,aAAa,QAAQ,eAAe;AAAA,MACxC;AAAA,IACJ;AAEA,QAAI,mBAAmB;AACnB,WAAK,oBAAoB,EAAE,OAAO,CAAC,EAAE,MAAM,kBAAkB,CAAC,EAAE;AAAA,IACpE;AAEA,QAAI,QAAQ,SAAS,QAAQ,MAAM,SAAS,GAAG;AAC3C,WAAK,QAAQ,CAAC;AAAA,QACV,sBAAsB,QAAQ,MAAM,IAAI,CAAC,OAAO;AAAA,UAC5C,MAAM,EAAE,SAAS;AAAA,UACjB,aAAa,EAAE,SAAS;AAAA,UACxB,YAAY,EAAE,SAAS;AAAA,QAC3B,EAAE;AAAA,MACN,CAAC;AAAA,IACL;AAEA,UAAM,MAAM,2DAA2D,KAAK;AAC5E,UAAM,WAAW,MAAM,eAAe,KAAK;AAAA,MACvC,QAAQ;AAAA,MACR,SAAS;AAAA,QACL,gBAAgB;AAAA,QAChB,kBAAkB;AAAA,MACtB;AAAA,MACA,MAAM,KAAK,UAAU,IAAI;AAAA,IAC7B,CAAC;AAED,QAAI,CAAC,SAAS,IAAI;AACd,YAAM,YAAY,MAAM,SAAS,KAAK;AAEtC,sBAAgB,QAAQ,SAAS,MAAM,IAAI,MAAM,EAAE,WAAW,MAAM,CAAC;AAErE,YAAM,EAAE,oBAAoB,IAAI,MAAM,OAAO,oBAAoB;AACjE,YAAM,oBAAoB,cAAc,UAAU,WAAW,EAAE,UAAU,UAAU,MAAM,CAAC;AAAA,IAC9F;AAEA,UAAM,OAAO,MAAM,SAAS,KAAK;AACjC,UAAM,aAAa,KAAK;AAExB,QAAI,cAAc;AAClB,UAAM,YAAwB,CAAC;AAE/B,QAAI,cAAc,WAAW,SAAS,GAAG;AACrC,YAAM,QAAS,WAAW,CAAC,EAAE,SAAqC,SAA2C,CAAC;AAC9G,iBAAW,QAAQ,OAAO;AACtB,YAAI,KAAK,MAAM;AACX,yBAAe,KAAK;AAAA,QACxB;AACA,YAAI,KAAK,cAAc;AACnB,gBAAM,KAAK,KAAK;AAChB,oBAAU,KAAK;AAAA,YACX,IAAI,KAAK;AAAA,YACT,MAAM;AAAA,YACN,UAAU;AAAA,cACN,MAAM,GAAG;AAAA,cACT,WAAW,KAAK,UAAU,GAAG,IAAI;AAAA,YACrC;AAAA,UACJ,CAAC;AAAA,QACL;AAAA,MACJ;AAAA,IACJ;AAEA,UAAM,YAAY,KAAK;AAEvB,WAAO;AAAA,MACH,IAAI,KAAK;AAAA,MACT,SAAS;AAAA,MACT,WAAW,UAAU,SAAS,IAAI,YAAY;AAAA,MAC9C,OAAO,YACD;AAAA,QACE,cAAc,UAAU,oBAAoB;AAAA,QAC5C,kBAAkB,UAAU,wBAAwB;AAAA,QACpD,aAAa,UAAU,mBAAmB;AAAA,MAC9C,IACE;AAAA,MACN,cAAc,UAAU,SAAS,IAAI,eAAe;AAAA,MACpD,OAAO,UAAU,KAAK;AAAA,IAC1B;AAAA,EACJ;AAAA,EAEA,OAAO,WAAW,SAAuD;AACrE,UAAM,SAAS,QAAQ,SAAS,oBAAoB,QAAQ,WAAW,EAAE;AACzE,UAAM,SAAS,KAAK;AACpB,QAAI,CAAC,QAAQ;AAAE,YAAM,EAAE,MAAM,SAAS,OAAO,gCAAgC;AAAG;AAAA,IAAQ;AAExF,UAAM,oBAAoB,QAAQ,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AAC7E,UAAM,EAAE,UAAU,YAAY,IAAI,cAAc,QAAQ,QAAQ;AAChE,QAAI,cAAc,GAAG;AACjB,aAAO,KAAK,WAAW,WAAW,WAAW,yDAAyD;AAAA,IAC1G;AAEA,UAAM,OAAgC;AAAA,MAClC;AAAA,MACA,kBAAkB,EAAE,iBAAiB,eAAe,QAAQ,SAAS,2BAA2B,QAAQ,SAAS,GAAG,aAAa,QAAQ,eAAe,IAAI;AAAA,IAChK;AACA,QAAI,kBAAmB,MAAK,oBAAoB,EAAE,OAAO,CAAC,EAAE,MAAM,kBAAkB,CAAC,EAAE;AACvF,QAAI,QAAQ,SAAS,QAAQ,MAAM,SAAS,GAAG;AAC3C,WAAK,QAAQ,CAAC,EAAE,sBAAsB,QAAQ,MAAM,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,SAAS,MAAM,aAAa,EAAE,SAAS,aAAa,YAAY,EAAE,SAAS,WAAW,EAAE,EAAE,CAAC;AAAA,IACzK;AAEA,QAAI;AACA,YAAM,MAAM,2DAA2D,KAAK;AAC5E,YAAM,WAAW,MAAM,MAAM,KAAK;AAAA,QAC9B,QAAQ;AAAA,QACR,SAAS,EAAE,gBAAgB,oBAAoB,kBAAkB,OAAO;AAAA,QACxE,MAAM,KAAK,UAAU,IAAI;AAAA,MAC7B,CAAC;AAED,UAAI,CAAC,SAAS,MAAM,CAAC,SAAS,MAAM;AAChC,cAAM,YAAY,MAAM,SAAS,KAAK;AACtC,wBAAgB,eAAe,SAAS,MAAM,IAAI,MAAM,EAAE,WAAW,MAAM,CAAC;AAC5E,cAAM,EAAE,MAAM,SAAS,OAAO,qBAAqB,SAAS,MAAM,MAAM,SAAS,GAAG;AACpF;AAAA,MACJ;AAEA,YAAM,SAAS,SAAS,KAAK,UAAU;AACvC,YAAM,UAAU,IAAI,YAAY;AAChC,UAAI,SAAS;AAEb,aAAO,MAAM;AACT,cAAM,EAAE,MAAM,MAAM,IAAI,MAAM,OAAO,KAAK;AAC1C,YAAI,KAAM;AACV,kBAAU,QAAQ,OAAO,OAAO,EAAE,QAAQ,KAAK,CAAC;AAEhD,cAAM,QAAQ,OAAO,MAAM,IAAI;AAC/B,iBAAS,MAAM,IAAI,KAAK;AAExB,mBAAW,QAAQ,OAAO;AACtB,cAAI,CAAC,KAAK,WAAW,QAAQ,EAAG;AAChC,gBAAM,OAAO,KAAK,MAAM,CAAC,EAAE,KAAK;AAChC,cAAI,CAAC,KAAM;AAEX,cAAI;AACA,kBAAM,QAAQ,KAAK,MAAM,IAAI;AAC7B,kBAAM,aAAa,MAAM;AACzB,gBAAI,cAAc,WAAW,SAAS,GAAG;AACrC,oBAAM,QAAS,WAAW,CAAC,EAAE,SAAqC,SAA2C,CAAC;AAC9G,yBAAW,QAAQ,OAAO;AACtB,oBAAI,KAAK,KAAM,OAAM,EAAE,MAAM,QAAQ,SAAS,KAAK,KAAe;AAClE,oBAAI,KAAK,cAAc;AACnB,wBAAM,KAAK,KAAK;AAChB,wBAAM;AAAA,oBACF,MAAM;AAAA,oBACN,UAAU,EAAE,IAAI,KAAK,GAAG,MAAM,YAAY,UAAU,EAAE,MAAM,GAAG,MAAgB,WAAW,KAAK,UAAU,GAAG,IAAI,EAAE,EAAE;AAAA,kBACxH;AAAA,gBACJ;AAAA,cACJ;AAAA,YACJ;AAAA,UACJ,QAAQ;AAAA,UAAiC;AAAA,QAC7C;AAAA,MACJ;AACA,YAAM,EAAE,MAAM,OAAO;AAAA,IACzB,SAAS,OAAO;AACZ,YAAM,EAAE,MAAM,SAAS,OAAQ,MAAgB,QAAQ;AAAA,IAC3D;AAAA,EACJ;AAAA,EAEA,MAAM,aAAgC;AAClC,WAAO,CAAC,kBAAkB,oBAAoB,oBAAoB,gBAAgB;AAAA,EACtF;AAAA,EAEA,MAAM,cAAgC;AAClC,QAAI;AACA,UAAI,CAAC,KAAK,OAAQ,QAAO;AACzB,YAAM,MAAM;AACZ,YAAM,WAAW,MAAM,MAAM,KAAK;AAAA,QAC9B,SAAS,EAAE,kBAAkB,KAAK,OAAO;AAAA,MAC7C,CAAC;AACD,aAAO,SAAS;AAAA,IACpB,QAAQ;AACJ,aAAO;AAAA,IACX;AAAA,EACJ;AACJ;","names":[]}

package/dist/providers/modelCapabilities.js

@@ -0,0 +1,59 @@
+#!/usr/bin/env node
+import { loadConfig } from "../config/config.js";
+const STATIC_TABLE = {
+  "anthropic/claude-sonnet-4-20250514": { contextWindow: 2e5, maxOutput: 64e3, supportsThinking: true },
+  "anthropic/claude-opus-4": { contextWindow: 2e5, maxOutput: 32e3, supportsThinking: true },
+  "anthropic/claude-3-5-sonnet-20241022": { contextWindow: 2e5, maxOutput: 8192 },
+  "anthropic/claude-3-5-haiku-20241022": { contextWindow: 2e5, maxOutput: 8192 },
+  "openai/gpt-4o": { contextWindow: 128e3, maxOutput: 16384 },
+  "openai/gpt-4o-mini": { contextWindow: 128e3, maxOutput: 16384 },
+  "openai/gpt-4.1": { contextWindow: 1e6, maxOutput: 32768 },
+  "openai/o1": { contextWindow: 2e5, maxOutput: 1e5, supportsThinking: true },
+  "openai/o1-mini": { contextWindow: 128e3, maxOutput: 65536, supportsThinking: true },
+  "openai/o3-mini": { contextWindow: 2e5, maxOutput: 1e5, supportsThinking: true },
+  "kimi/kimi-k2.6": { contextWindow: 262144, maxOutput: 128e3, supportsThinking: true },
+  "moonshot/kimi-k2-0905-preview": { contextWindow: 262144, maxOutput: 128e3 },
+  "zhipu/glm-5.1": { contextWindow: 198e3, maxOutput: 128e3 },
+  "ollama/qwen3.5:cloud": { contextWindow: 128e3, maxOutput: 32e3 },
+  "ollama/qwen3:32b": { contextWindow: 32768, maxOutput: 8192 },
+  "ollama/llama3.3:70b": { contextWindow: 128e3, maxOutput: 8192 },
+  "ollama/deepseek-v3": { contextWindow: 128e3, maxOutput: 16384 },
+  "mistral/mistral-large": { contextWindow: 128e3, maxOutput: 8192 },
+  "gemini/gemini-2.0-flash": { contextWindow: 1e6, maxOutput: 8192 },
+  "gemini/gemini-2.5-pro": { contextWindow: 2e6, maxOutput: 65536, supportsThinking: true },
+  "cohere/command-r-plus": { contextWindow: 128e3, maxOutput: 4096 },
+  "groq/llama-3.3-70b-versatile": { contextWindow: 128e3, maxOutput: 32768 }
+};
+const FAMILY_DEFAULTS = [
+  { pattern: /^anthropic\//, caps: { contextWindow: 2e5, maxOutput: 8192 } },
+  { pattern: /^openai\/o\d/, caps: { contextWindow: 2e5, maxOutput: 1e5, supportsThinking: true } },
+  { pattern: /^openai\//, caps: { contextWindow: 128e3, maxOutput: 16384 } },
+  { pattern: /^kimi\/|^moonshot\//, caps: { contextWindow: 262144, maxOutput: 128e3 } },
+  { pattern: /^zhipu\/|^glm/, caps: { contextWindow: 198e3, maxOutput: 128e3 } },
+  { pattern: /^gemini\//, caps: { contextWindow: 1e6, maxOutput: 8192 } },
+  { pattern: /^ollama\//, caps: { contextWindow: 32e3, maxOutput: 8192 } }
+];
+const DEFAULT_FALLBACK = { contextWindow: 32e3, maxOutput: 8192 };
+function getModelCapabilities(model) {
+  if (STATIC_TABLE[model]) return STATIC_TABLE[model];
+  try {
+    const cfg = loadConfig();
+    const override = cfg.providers?.modelCapabilities;
+    if (override?.[model]) return override[model];
+  } catch {
+  }
+  for (const f of FAMILY_DEFAULTS) {
+    if (f.pattern.test(model)) return f.caps;
+  }
+  return DEFAULT_FALLBACK;
+}
+function clampMaxTokens(model, requested) {
+  const caps = getModelCapabilities(model);
+  const want = requested ?? caps.maxOutput;
+  return Math.max(1, Math.min(want, caps.maxOutput));
+}
+export {
+  clampMaxTokens,
+  getModelCapabilities
+};
+//# sourceMappingURL=modelCapabilities.js.map

package/dist/providers/modelCapabilities.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/providers/modelCapabilities.ts"],"sourcesContent":["/**\n * TITAN — Per-Model Capability Registry\n *\n * Single source of truth for output-token caps, context-window sizes,\n * and feature flags per model.  Providers call clampMaxTokens() so\n * DEFAULT_MAX_TOKENS can safely be a high user-preference ceiling\n * (200 K) without causing 400s on capped providers.\n */\nimport { loadConfig } from '../config/config.js';\n\nexport interface ModelCapabilities {\n    contextWindow: number;\n    maxOutput: number;\n    supportsThinking?: boolean;\n}\n\n// Exact-match table.  Keep in alphabetical order by provider / model.\nconst STATIC_TABLE: Record<string, ModelCapabilities> = {\n    'anthropic/claude-sonnet-4-20250514':   { contextWindow: 200_000, maxOutput: 64_000, supportsThinking: true },\n    'anthropic/claude-opus-4':              { contextWindow: 200_000, maxOutput: 32_000, supportsThinking: true },\n    'anthropic/claude-3-5-sonnet-20241022': { contextWindow: 200_000, maxOutput: 8_192 },\n    'anthropic/claude-3-5-haiku-20241022':  { contextWindow: 200_000, maxOutput: 8_192 },\n    'openai/gpt-4o':                        { contextWindow: 128_000, maxOutput: 16_384 },\n    'openai/gpt-4o-mini':                   { contextWindow: 128_000, maxOutput: 16_384 },\n    'openai/gpt-4.1':                       { contextWindow: 1_000_000, maxOutput: 32_768 },\n    'openai/o1':                            { contextWindow: 200_000, maxOutput: 100_000, supportsThinking: true },\n    'openai/o1-mini':                       { contextWindow: 128_000, maxOutput: 65_536, supportsThinking: true },\n    'openai/o3-mini':                       { contextWindow: 200_000, maxOutput: 100_000, supportsThinking: true },\n    'kimi/kimi-k2.6':                       { contextWindow: 262_144, maxOutput: 128_000, supportsThinking: true },\n    'moonshot/kimi-k2-0905-preview':        { contextWindow: 262_144, maxOutput: 128_000 },\n    'zhipu/glm-5.1':                        { contextWindow: 198_000, maxOutput: 128_000 },\n    'ollama/qwen3.5:cloud':                 { contextWindow: 128_000, maxOutput: 32_000 },\n    'ollama/qwen3:32b':                     { contextWindow: 32_768,  maxOutput: 8_192 },\n    'ollama/llama3.3:70b':                  { contextWindow: 128_000, maxOutput: 8_192 },\n    'ollama/deepseek-v3':                   { contextWindow: 128_000, maxOutput: 16_384 },\n    'mistral/mistral-large':                { contextWindow: 128_000, maxOutput: 8_192 },\n    'gemini/gemini-2.0-flash':              { contextWindow: 1_000_000, maxOutput: 8_192 },\n    'gemini/gemini-2.5-pro':                { contextWindow: 2_000_000, maxOutput: 65_536, supportsThinking: true },\n    'cohere/command-r-plus':                { contextWindow: 128_000, maxOutput: 4_096 },\n    'groq/llama-3.3-70b-versatile':         { contextWindow: 128_000, maxOutput: 32_768 },\n};\n\n// Family heuristic for unknown specific versions.\nconst FAMILY_DEFAULTS: Array<{ pattern: RegExp; caps: ModelCapabilities }> = [\n    { pattern: /^anthropic\\//,                caps: { contextWindow: 200_000, maxOutput: 8_192 } },\n    { pattern: /^openai\\/o\\d/,                caps: { contextWindow: 200_000, maxOutput: 100_000, supportsThinking: true } },\n    { pattern: /^openai\\//,                   caps: { contextWindow: 128_000, maxOutput: 16_384 } },\n    { pattern: /^kimi\\/|^moonshot\\//,         caps: { contextWindow: 262_144, maxOutput: 128_000 } },\n    { pattern: /^zhipu\\/|^glm/,               caps: { contextWindow: 198_000, maxOutput: 128_000 } },\n    { pattern: /^gemini\\//,                   caps: { contextWindow: 1_000_000, maxOutput: 8_192 } },\n    { pattern: /^ollama\\//,                   caps: { contextWindow: 32_000,  maxOutput: 8_192 } },\n];\n\nconst DEFAULT_FALLBACK: ModelCapabilities = { contextWindow: 32_000, maxOutput: 8_192 };\n\nexport function getModelCapabilities(model: string): ModelCapabilities {\n    if (STATIC_TABLE[model]) return STATIC_TABLE[model];\n\n    // Config override path\n    try {\n        const cfg = loadConfig();\n        const override = (cfg.providers as Record<string, unknown> | undefined)?.modelCapabilities as\n            Record<string, ModelCapabilities> | undefined;\n        if (override?.[model]) return override[model];\n    } catch { /* config not loaded yet during early boot */ }\n\n    // Family fallback\n    for (const f of FAMILY_DEFAULTS) {\n        if (f.pattern.test(model)) return f.caps;\n    }\n\n    return DEFAULT_FALLBACK;\n}\n\n/**\n * Clamp a requested maxTokens to the model's actual ceiling.\n * If requested is undefined, returns the model's own maxOutput.\n * Never returns less than 1.\n */\nexport function clampMaxTokens(model: string, requested?: number): number {\n    const caps = getModelCapabilities(model);\n    const want = requested ?? caps.maxOutput;\n    return Math.max(1, Math.min(want, caps.maxOutput));\n}\n"],"mappings":";AAQA,SAAS,kBAAkB;AAS3B,MAAM,eAAkD;AAAA,EACpD,sCAAwC,EAAE,eAAe,KAAS,WAAW,MAAQ,kBAAkB,KAAK;AAAA,EAC5G,2BAAwC,EAAE,eAAe,KAAS,WAAW,MAAQ,kBAAkB,KAAK;AAAA,EAC5G,wCAAwC,EAAE,eAAe,KAAS,WAAW,KAAM;AAAA,EACnF,uCAAwC,EAAE,eAAe,KAAS,WAAW,KAAM;AAAA,EACnF,iBAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AAAA,EACpF,sBAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AAAA,EACpF,kBAAwC,EAAE,eAAe,KAAW,WAAW,MAAO;AAAA,EACtF,aAAwC,EAAE,eAAe,KAAS,WAAW,KAAS,kBAAkB,KAAK;AAAA,EAC7G,kBAAwC,EAAE,eAAe,OAAS,WAAW,OAAQ,kBAAkB,KAAK;AAAA,EAC5G,kBAAwC,EAAE,eAAe,KAAS,WAAW,KAAS,kBAAkB,KAAK;AAAA,EAC7G,kBAAwC,EAAE,eAAe,QAAS,WAAW,OAAS,kBAAkB,KAAK;AAAA,EAC7G,iCAAwC,EAAE,eAAe,QAAS,WAAW,MAAQ;AAAA,EACrF,iBAAwC,EAAE,eAAe,OAAS,WAAW,MAAQ;AAAA,EACrF,wBAAwC,EAAE,eAAe,OAAS,WAAW,KAAO;AAAA,EACpF,oBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,uBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,sBAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AAAA,EACpF,yBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,2BAAwC,EAAE,eAAe,KAAW,WAAW,KAAM;AAAA,EACrF,yBAAwC,EAAE,eAAe,KAAW,WAAW,OAAQ,kBAAkB,KAAK;AAAA,EAC9G,yBAAwC,EAAE,eAAe,OAAS,WAAW,KAAM;AAAA,EACnF,gCAAwC,EAAE,eAAe,OAAS,WAAW,MAAO;AACxF;AAGA,MAAM,kBAAuE;AAAA,EACzE,EAAE,SAAS,gBAA+B,MAAM,EAAE,eAAe,KAAS,WAAW,KAAM,EAAE;AAAA,EAC7F,EAAE,SAAS,gBAA+B,MAAM,EAAE,eAAe,KAAS,WAAW,KAAS,kBAAkB,KAAK,EAAE;AAAA,EACvH,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAO,EAAE;AAAA,EAC9F,EAAE,SAAS,uBAA+B,MAAM,EAAE,eAAe,QAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,iBAA+B,MAAM,EAAE,eAAe,OAAS,WAAW,MAAQ,EAAE;AAAA,EAC/F,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,KAAW,WAAW,KAAM,EAAE;AAAA,EAC/F,EAAE,SAAS,aAA+B,MAAM,EAAE,eAAe,MAAS,WAAW,KAAM,EAAE;AACjG;AAEA,MAAM,mBAAsC,EAAE,eAAe,MAAQ,WAAW,KAAM;AAE/E,SAAS,qBAAqB,OAAkC;AACnE,MAAI,aAAa,KAAK,EAAG,QAAO,aAAa,KAAK;AAGlD,MAAI;AACA,UAAM,MAAM,WAAW;AACvB,UAAM,WAAY,IAAI,WAAmD;AAEzE,QAAI,WAAW,KAAK,EAAG,QAAO,SAAS,KAAK;AAAA,EAChD,QAAQ;AAAA,EAAgD;AAGxD,aAAW,KAAK,iBAAiB;AAC7B,QAAI,EAAE,QAAQ,KAAK,KAAK,EAAG,QAAO,EAAE;AAAA,EACxC;AAEA,SAAO;AACX;AAOO,SAAS,eAAe,OAAe,WAA4B;AACtE,QAAM,OAAO,qBAAqB,KAAK;AACvC,QAAM,OAAO,aAAa,KAAK;AAC/B,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,MAAM,KAAK,SAAS,CAAC;AACrD;","names":[]}

package/dist/providers/ollama.js

@@ -7,6 +7,7 @@ import logger from "../utils/logger.js";
 import { fetchWithRetry } from "../utils/helpers.js";
 import { v4 as uuid } from "uuid";
 import * as fs from "fs";
+import { clampMaxTokens } from "./modelCapabilities.js";
 const COMPONENT = "Ollama";
 const CLOUD_MODEL_CTX = {
   // GLM-5.1 — 198K context (newest agentic flagship, SOTA SWE-Bench Pro)
@@ -344,7 +345,7 @@ class OllamaProvider extends LLMProvider {
         // responses don't come close to that. 8K is plenty for any
         // single turn and keeps us from getting HTTP 402s when
         // credit runs low.
-        num_predict: options.maxTokens || (isCloudModel ? 8192 : 16384),
+        num_predict: clampMaxTokens(options.model || "ollama/llama3.1", options.maxTokens),
         num_ctx: getModelCtx(model),
         temperature: options.temperature ?? 0.7
       }
@@ -531,7 +532,7 @@ ${msgs[firstUserIdx].content}`;
       keep_alive: "30m",
       options: {
         // v4.10.0-local (cost cap): 8K cloud cap matches non-stream path
-        num_predict: options.maxTokens || (isCloudModel ? 8192 : 16384),
+        num_predict: clampMaxTokens(options.model || "ollama/llama3.1", options.maxTokens),
         num_ctx: getModelCtx(model),
         temperature: options.temperature ?? 0.7
       }