titan-agent 6.4.3 → 6.5.0

package/dist/gateway/openai-compat.js

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 import { Router } from "express";
-import { chat, chatStream } from "../providers/router.js";
+import { chat, chatStream, tryResolveModel, discoverAllModels, getKnownProviderNames } from "../providers/router.js";
 import { loadConfig } from "../config/config.js";
 import { v4 as uuid } from "uuid";
 import logger from "../utils/logger.js";
@@ -24,7 +24,7 @@ function createOpenAICompatRouter() {
     res.json({ object: "list", data: models });
   });
   router.post("/chat/completions", async (req, res) => {
-    const { model, messages, stream, temperature, max_tokens, top_p, stop } = req.body;
+    const { model, messages, stream, temperature, max_tokens } = req.body;
     if (!messages || !Array.isArray(messages)) {
       res.status(400).json({ error: { message: "messages is required", type: "invalid_request_error" } });
       return;
@@ -32,20 +32,59 @@ function createOpenAICompatRouter() {
     const config = loadConfig();
     const effectiveModel = model || config.agent.model;
     const chatId = `chatcmpl-${uuid().slice(0, 12)}`;
-    logger.info(COMPONENT, `Chat completion: model=${effectiveModel}, messages=${messages.length}, stream=${!!stream}`);
+    if (model) {
+      const aliases = config.agent.modelAliases || {};
+      const aliased = aliases[effectiveModel] || effectiveModel;
+      const bareName = aliased.includes("/") ? aliased.split("/").slice(1).join("/") : aliased;
+      let known = false;
+      let availableIds = [];
+      try {
+        const discovered = await discoverAllModels();
+        availableIds = discovered.map((m) => m.id).slice(0, 50);
+        known = // a real, servable model — matched by full id or bare model name
+        discovered.some((m) => m.id === aliased || m.model === aliased || m.model === bareName) || aliased.includes("/") && getKnownProviderNames().includes(aliased.split("/")[0]);
+      } catch {
+        known = !!tryResolveModel(aliased);
+      }
+      if (!known) {
+        res.status(404).json({
+          error: {
+            message: `Model '${model}' is not recognized by TITAN: no configured alias, not a servable model, and no known provider prefix. See 'available_models'.`,
+            type: "model_not_found",
+            param: "model",
+            code: "model_not_found",
+            available_models: availableIds
+          }
+        });
+        return;
+      }
+    }
+    const routeInfo = tryResolveModel(effectiveModel);
+    const actualProvider = routeInfo?.provider?.name ?? "unknown";
+    const actualModelId = routeInfo ? `${routeInfo.provider.name}/${routeInfo.model}` : effectiveModel;
+    logger.info(COMPONENT, `Chat completion: requested=${effectiveModel} \u2192 routed=${actualModelId}, messages=${messages.length}, stream=${!!stream}`);
+    const ac = new AbortController();
+    req.on("close", () => ac.abort());
     try {
       if (stream) {
         res.writeHead(200, {
           "Content-Type": "text/event-stream",
           "Cache-Control": "no-cache",
-          Connection: "keep-alive"
+          Connection: "keep-alive",
+          // v6.5 P0: cost transparency — surface the actual route
+          // even when the body says model='gpt-4'. Clients can read
+          // this header in their network logs / SDK middleware.
+          "X-Titan-Provider": actualProvider,
+          "X-Titan-Routed-Model": actualModelId,
+          "X-Titan-Requested-Model": effectiveModel
         });
         const sseWrite = setupSSEFlush(res);
         for await (const chunk of chatStream({
           model: effectiveModel,
           messages,
           temperature: temperature ?? config.agent.temperature,
-          maxTokens: max_tokens ?? config.agent.maxTokens
+          maxTokens: max_tokens ?? config.agent.maxTokens,
+          signal: ac.signal
         })) {
           if (chunk.type === "text" && chunk.content) {
             const data = {
@@ -80,8 +119,12 @@ function createOpenAICompatRouter() {
           model: effectiveModel,
           messages,
           temperature: temperature ?? config.agent.temperature,
-          maxTokens: max_tokens ?? config.agent.maxTokens
+          maxTokens: max_tokens ?? config.agent.maxTokens,
+          signal: ac.signal
         });
+        res.setHeader("X-Titan-Provider", actualProvider);
+        res.setHeader("X-Titan-Routed-Model", actualModelId);
+        res.setHeader("X-Titan-Requested-Model", effectiveModel);
         res.json({
           id: chatId,
           object: "chat.completion",
@@ -100,10 +143,26 @@ function createOpenAICompatRouter() {
         });
       }
     } catch (err) {
-      logger.error(COMPONENT, `Chat completion error: ${err.message}`);
-      res.status(500).json({
-        error: { message: err.message, type: "server_error" }
-      });
+      const errMsg = err.message;
+      const errName = err.name;
+      const aborted = ac.signal.aborted || errName === "AbortError" || errName === "TimeoutError";
+      if (aborted) {
+        logger.info(COMPONENT, `Chat completion cancelled (client disconnect or timeout): ${errMsg}`);
+      } else {
+        logger.error(COMPONENT, `Chat completion error: ${errMsg}`);
+      }
+      if (res.headersSent) {
+        try {
+          res.end();
+        } catch {
+        }
+      } else {
+        res.setHeader("X-Titan-Provider", actualProvider);
+        res.setHeader("X-Titan-Error-Class", aborted ? "request_cancelled" : "server_error");
+        res.status(aborted ? 499 : 500).json({
+          error: { message: errMsg, type: aborted ? "request_cancelled" : "server_error" }
+        });
+      }
     }
   });
   router.post("/embeddings", async (req, res) => {

package/dist/gateway/openai-compat.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/gateway/openai-compat.ts"],"sourcesContent":["/**\n * TITAN — OpenAI API Compatibility Layer\n * Serves /v1/models, /v1/chat/completions, /v1/embeddings so any OpenAI-compatible\n * client (OpenWebUI, Continue.dev, LiteLLM, etc.) can connect to TITAN as a provider.\n */\nimport { Router, type Request, type Response } from 'express';\nimport { chat, chatStream } from '../providers/router.js';\nimport { loadConfig } from '../config/config.js';\nimport { TITAN_VERSION } from '../utils/constants.js';\nimport { v4 as uuid } from 'uuid';\nimport logger from '../utils/logger.js';\nimport { setupSSEFlush } from '../utils/sseFlush.js';\n\nconst COMPONENT = 'OpenAI-Compat';\n\nexport function createOpenAICompatRouter(): Router {\n    const router = Router();\n\n    // GET /v1/models — list available models\n    router.get('/models', (_req: Request, res: Response) => {\n        const config = loadConfig();\n        const currentModel = config.agent.model || 'unknown';\n        const aliases = config.agent.modelAliases || {};\n\n        const models = [\n            { id: currentModel, object: 'model', created: Math.floor(Date.now() / 1000), owned_by: 'titan' },\n            ...Object.entries(aliases)\n                .filter(([, v]) => v && v !== currentModel)\n                .map(([_key, value]) => ({\n                    id: value as string,\n                    object: 'model' as const,\n                    created: Math.floor(Date.now() / 1000),\n                    owned_by: 'titan',\n                })),\n        ];\n\n        res.json({ object: 'list', data: models });\n    });\n\n    // POST /v1/chat/completions — chat completion (streaming and non-streaming)\n    router.post('/chat/completions', async (req: Request, res: Response) => {\n        const { model, messages, stream, temperature, max_tokens, top_p, stop } = req.body;\n\n        if (!messages || !Array.isArray(messages)) {\n            res.status(400).json({ error: { message: 'messages is required', type: 'invalid_request_error' } });\n            return;\n        }\n\n        const config = loadConfig();\n        const effectiveModel = model || config.agent.model;\n        const chatId = `chatcmpl-${uuid().slice(0, 12)}`;\n\n        logger.info(COMPONENT, `Chat completion: model=${effectiveModel}, messages=${messages.length}, stream=${!!stream}`);\n\n        try {\n            if (stream) {\n                // SSE streaming response\n                res.writeHead(200, {\n                    'Content-Type': 'text/event-stream',\n                    'Cache-Control': 'no-cache',\n                    Connection: 'keep-alive',\n                });\n                const sseWrite = setupSSEFlush(res);\n\n                for await (const chunk of chatStream({\n                    model: effectiveModel,\n                    messages,\n                    temperature: temperature ?? config.agent.temperature,\n                    maxTokens: max_tokens ?? config.agent.maxTokens,\n                })) {\n                    if (chunk.type === 'text' && chunk.content) {\n                        const data = {\n                            id: chatId,\n                            object: 'chat.completion.chunk',\n                            created: Math.floor(Date.now() / 1000),\n                            model: effectiveModel,\n                            choices: [{\n                                index: 0,\n                                delta: { content: chunk.content },\n                                finish_reason: null,\n                            }],\n                        };\n                        sseWrite(`data: ${JSON.stringify(data)}\\n\\n`);\n                    }\n                }\n\n                // Send final chunk\n                sseWrite(`data: ${JSON.stringify({\n                    id: chatId,\n                    object: 'chat.completion.chunk',\n                    created: Math.floor(Date.now() / 1000),\n                    model: effectiveModel,\n                    choices: [{ index: 0, delta: {}, finish_reason: 'stop' }],\n                })}\\n\\n`);\n                sseWrite('data: [DONE]\\n\\n');\n                res.end();\n            } else {\n                // Non-streaming response\n                const response = await chat({\n                    model: effectiveModel,\n                    messages,\n                    temperature: temperature ?? config.agent.temperature,\n                    maxTokens: max_tokens ?? config.agent.maxTokens,\n                });\n\n                res.json({\n                    id: chatId,\n                    object: 'chat.completion',\n                    created: Math.floor(Date.now() / 1000),\n                    model: response.model || effectiveModel,\n                    choices: [{\n                        index: 0,\n                        message: { role: 'assistant', content: response.content },\n                        finish_reason: response.finishReason || 'stop',\n                    }],\n                    usage: {\n                        prompt_tokens: response.usage?.promptTokens ?? 0,\n                        completion_tokens: response.usage?.completionTokens ?? 0,\n                        total_tokens: response.usage?.totalTokens ?? 0,\n                    },\n                });\n            }\n        } catch (err) {\n            logger.error(COMPONENT, `Chat completion error: ${(err as Error).message}`);\n            res.status(500).json({\n                error: { message: (err as Error).message, type: 'server_error' },\n            });\n        }\n    });\n\n    // POST /v1/embeddings — embedding generation (proxy to Ollama)\n    router.post('/embeddings', async (req: Request, res: Response) => {\n        const { model, input } = req.body;\n        const config = loadConfig();\n        const embeddingModel = model || (config.memory as Record<string, unknown>)?.embeddingModel || 'nomic-embed-text';\n\n        try {\n            // Proxy to Ollama embeddings endpoint\n            const ollamaUrl = config.providers?.ollama?.baseUrl || 'http://localhost:11434';\n            const texts = Array.isArray(input) ? input : [input];\n            const embeddings = [];\n\n            for (let i = 0; i < texts.length; i++) {\n                const resp = await fetch(`${ollamaUrl}/api/embed`, {\n                    method: 'POST',\n                    headers: { 'Content-Type': 'application/json' },\n                    body: JSON.stringify({ model: embeddingModel, input: texts[i] }),\n                });\n\n                if (!resp.ok) throw new Error(`Ollama embed failed: ${resp.status}`);\n                const data = await resp.json() as { embeddings?: number[][] };\n                embeddings.push({\n                    object: 'embedding',\n                    index: i,\n                    embedding: data.embeddings?.[0] || [],\n                });\n            }\n\n            res.json({\n                object: 'list',\n                data: embeddings,\n                model: embeddingModel,\n                usage: { prompt_tokens: 0, total_tokens: 0 },\n            });\n        } catch (err) {\n            res.status(500).json({\n                error: { message: (err as Error).message, type: 'server_error' },\n            });\n        }\n    });\n\n    return router;\n}\n"],"mappings":";AAKA,SAAS,cAA2C;AACpD,SAAS,MAAM,kBAAkB;AACjC,SAAS,kBAAkB;AAE3B,SAAS,MAAM,YAAY;AAC3B,OAAO,YAAY;AACnB,SAAS,qBAAqB;AAE9B,MAAM,YAAY;AAEX,SAAS,2BAAmC;AAC/C,QAAM,SAAS,OAAO;AAGtB,SAAO,IAAI,WAAW,CAAC,MAAe,QAAkB;AACpD,UAAM,SAAS,WAAW;AAC1B,UAAM,eAAe,OAAO,MAAM,SAAS;AAC3C,UAAM,UAAU,OAAO,MAAM,gBAAgB,CAAC;AAE9C,UAAM,SAAS;AAAA,MACX,EAAE,IAAI,cAAc,QAAQ,SAAS,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI,GAAG,UAAU,QAAQ;AAAA,MAC/F,GAAG,OAAO,QAAQ,OAAO,EACpB,OAAO,CAAC,CAAC,EAAE,CAAC,MAAM,KAAK,MAAM,YAAY,EACzC,IAAI,CAAC,CAAC,MAAM,KAAK,OAAO;AAAA,QACrB,IAAI;AAAA,QACJ,QAAQ;AAAA,QACR,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAAA,QACrC,UAAU;AAAA,MACd,EAAE;AAAA,IACV;AAEA,QAAI,KAAK,EAAE,QAAQ,QAAQ,MAAM,OAAO,CAAC;AAAA,EAC7C,CAAC;AAGD,SAAO,KAAK,qBAAqB,OAAO,KAAc,QAAkB;AACpE,UAAM,EAAE,OAAO,UAAU,QAAQ,aAAa,YAAY,OAAO,KAAK,IAAI,IAAI;AAE9E,QAAI,CAAC,YAAY,CAAC,MAAM,QAAQ,QAAQ,GAAG;AACvC,UAAI,OAAO,GAAG,EAAE,KAAK,EAAE,OAAO,EAAE,SAAS,wBAAwB,MAAM,wBAAwB,EAAE,CAAC;AAClG;AAAA,IACJ;AAEA,UAAM,SAAS,WAAW;AAC1B,UAAM,iBAAiB,SAAS,OAAO,MAAM;AAC7C,UAAM,SAAS,YAAY,KAAK,EAAE,MAAM,GAAG,EAAE,CAAC;AAE9C,WAAO,KAAK,WAAW,0BAA0B,cAAc,cAAc,SAAS,MAAM,YAAY,CAAC,CAAC,MAAM,EAAE;AAElH,QAAI;AACA,UAAI,QAAQ;AAER,YAAI,UAAU,KAAK;AAAA,UACf,gBAAgB;AAAA,UAChB,iBAAiB;AAAA,UACjB,YAAY;AAAA,QAChB,CAAC;AACD,cAAM,WAAW,cAAc,GAAG;AAElC,yBAAiB,SAAS,WAAW;AAAA,UACjC,OAAO;AAAA,UACP;AAAA,UACA,aAAa,eAAe,OAAO,MAAM;AAAA,UACzC,WAAW,cAAc,OAAO,MAAM;AAAA,QAC1C,CAAC,GAAG;AACA,cAAI,MAAM,SAAS,UAAU,MAAM,SAAS;AACxC,kBAAM,OAAO;AAAA,cACT,IAAI;AAAA,cACJ,QAAQ;AAAA,cACR,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAAA,cACrC,OAAO;AAAA,cACP,SAAS,CAAC;AAAA,gBACN,OAAO;AAAA,gBACP,OAAO,EAAE,SAAS,MAAM,QAAQ;AAAA,gBAChC,eAAe;AAAA,cACnB,CAAC;AAAA,YACL;AACA,qBAAS,SAAS,KAAK,UAAU,IAAI,CAAC;AAAA;AAAA,CAAM;AAAA,UAChD;AAAA,QACJ;AAGA,iBAAS,SAAS,KAAK,UAAU;AAAA,UAC7B,IAAI;AAAA,UACJ,QAAQ;AAAA,UACR,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAAA,UACrC,OAAO;AAAA,UACP,SAAS,CAAC,EAAE,OAAO,GAAG,OAAO,CAAC,GAAG,eAAe,OAAO,CAAC;AAAA,QAC5D,CAAC,CAAC;AAAA;AAAA,CAAM;AACR,iBAAS,kBAAkB;AAC3B,YAAI,IAAI;AAAA,MACZ,OAAO;AAEH,cAAM,WAAW,MAAM,KAAK;AAAA,UACxB,OAAO;AAAA,UACP;AAAA,UACA,aAAa,eAAe,OAAO,MAAM;AAAA,UACzC,WAAW,cAAc,OAAO,MAAM;AAAA,QAC1C,CAAC;AAED,YAAI,KAAK;AAAA,UACL,IAAI;AAAA,UACJ,QAAQ;AAAA,UACR,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAAA,UACrC,OAAO,SAAS,SAAS;AAAA,UACzB,SAAS,CAAC;AAAA,YACN,OAAO;AAAA,YACP,SAAS,EAAE,MAAM,aAAa,SAAS,SAAS,QAAQ;AAAA,YACxD,eAAe,SAAS,gBAAgB;AAAA,UAC5C,CAAC;AAAA,UACD,OAAO;AAAA,YACH,eAAe,SAAS,OAAO,gBAAgB;AAAA,YAC/C,mBAAmB,SAAS,OAAO,oBAAoB;AAAA,YACvD,cAAc,SAAS,OAAO,eAAe;AAAA,UACjD;AAAA,QACJ,CAAC;AAAA,MACL;AAAA,IACJ,SAAS,KAAK;AACV,aAAO,MAAM,WAAW,0BAA2B,IAAc,OAAO,EAAE;AAC1E,UAAI,OAAO,GAAG,EAAE,KAAK;AAAA,QACjB,OAAO,EAAE,SAAU,IAAc,SAAS,MAAM,eAAe;AAAA,MACnE,CAAC;AAAA,IACL;AAAA,EACJ,CAAC;AAGD,SAAO,KAAK,eAAe,OAAO,KAAc,QAAkB;AAC9D,UAAM,EAAE,OAAO,MAAM,IAAI,IAAI;AAC7B,UAAM,SAAS,WAAW;AAC1B,UAAM,iBAAiB,SAAU,OAAO,QAAoC,kBAAkB;AAE9F,QAAI;AAEA,YAAM,YAAY,OAAO,WAAW,QAAQ,WAAW;AACvD,YAAM,QAAQ,MAAM,QAAQ,KAAK,IAAI,QAAQ,CAAC,KAAK;AACnD,YAAM,aAAa,CAAC;AAEpB,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACnC,cAAM,OAAO,MAAM,MAAM,GAAG,SAAS,cAAc;AAAA,UAC/C,QAAQ;AAAA,UACR,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,UAC9C,MAAM,KAAK,UAAU,EAAE,OAAO,gBAAgB,OAAO,MAAM,CAAC,EAAE,CAAC;AAAA,QACnE,CAAC;AAED,YAAI,CAAC,KAAK,GAAI,OAAM,IAAI,MAAM,wBAAwB,KAAK,MAAM,EAAE;AACnE,cAAM,OAAO,MAAM,KAAK,KAAK;AAC7B,mBAAW,KAAK;AAAA,UACZ,QAAQ;AAAA,UACR,OAAO;AAAA,UACP,WAAW,KAAK,aAAa,CAAC,KAAK,CAAC;AAAA,QACxC,CAAC;AAAA,MACL;AAEA,UAAI,KAAK;AAAA,QACL,QAAQ;AAAA,QACR,MAAM;AAAA,QACN,OAAO;AAAA,QACP,OAAO,EAAE,eAAe,GAAG,cAAc,EAAE;AAAA,MAC/C,CAAC;AAAA,IACL,SAAS,KAAK;AACV,UAAI,OAAO,GAAG,EAAE,KAAK;AAAA,QACjB,OAAO,EAAE,SAAU,IAAc,SAAS,MAAM,eAAe;AAAA,MACnE,CAAC;AAAA,IACL;AAAA,EACJ,CAAC;AAED,SAAO;AACX;","names":[]}
+{"version":3,"sources":["../../src/gateway/openai-compat.ts"],"sourcesContent":["/**\n * TITAN — OpenAI API Compatibility Layer\n * Serves /v1/models, /v1/chat/completions, /v1/embeddings so any OpenAI-compatible\n * client (OpenWebUI, Continue.dev, LiteLLM, etc.) can connect to TITAN as a provider.\n */\nimport { Router, type Request, type Response } from 'express';\nimport { chat, chatStream, tryResolveModel, discoverAllModels, getKnownProviderNames } from '../providers/router.js';\nimport { loadConfig } from '../config/config.js';\nimport { v4 as uuid } from 'uuid';\nimport logger from '../utils/logger.js';\nimport { setupSSEFlush } from '../utils/sseFlush.js';\n\nconst COMPONENT = 'OpenAI-Compat';\n\nexport function createOpenAICompatRouter(): Router {\n    const router = Router();\n\n    // GET /v1/models — list available models\n    router.get('/models', (_req: Request, res: Response) => {\n        const config = loadConfig();\n        const currentModel = config.agent.model || 'unknown';\n        const aliases = config.agent.modelAliases || {};\n\n        const models = [\n            { id: currentModel, object: 'model', created: Math.floor(Date.now() / 1000), owned_by: 'titan' },\n            ...Object.entries(aliases)\n                .filter(([, v]) => v && v !== currentModel)\n                .map(([_key, value]) => ({\n                    id: value as string,\n                    object: 'model' as const,\n                    created: Math.floor(Date.now() / 1000),\n                    owned_by: 'titan',\n                })),\n        ];\n\n        res.json({ object: 'list', data: models });\n    });\n\n    // POST /v1/chat/completions — chat completion (streaming and non-streaming)\n    router.post('/chat/completions', async (req: Request, res: Response) => {\n        const { model, messages, stream, temperature, max_tokens } = req.body;\n\n        if (!messages || !Array.isArray(messages)) {\n            res.status(400).json({ error: { message: 'messages is required', type: 'invalid_request_error' } });\n            return;\n        }\n\n        const config = loadConfig();\n        const effectiveModel = model || config.agent.model;\n        const chatId = `chatcmpl-${uuid().slice(0, 12)}`;\n\n        // ── v6.5 P2: validate model is recognized BEFORE generating ──\n        // Sage flagged this during the v6.5 Council session: silent fallback to\n        // config.agent.model when the client sends a model TITAN doesn't know\n        // (e.g. 'claude-3.7-sonnet') is a trust violation — the client thinks\n        // they got 3.7-sonnet's output. Return 404 with the recognized models\n        // list instead. Skip this check when the client omitted `model`\n        // entirely (using TITAN's default is fine).\n        if (model) {\n            // v6.5 fix — the old guard used tryResolveModel(), which RESOLVES a\n            // bare id like 'gpt-4'/'claude-3.7-sonnet' to anthropic/<literal> and\n            // returns truthy, so the 404 only fired for ids with an explicit\n            // UNKNOWN provider prefix ('foo/bar'). The exact deception P2 targets\n            // (bare 'claude-3.7-sonnet' → silent mis-route → Anthropic 500) slipped\n            // through. Validate against the REAL servable-model set instead.\n            const aliases = (config.agent.modelAliases || {}) as Record<string, string>;\n            const aliased = aliases[effectiveModel] || effectiveModel;\n            const bareName = aliased.includes('/') ? aliased.split('/').slice(1).join('/') : aliased;\n\n            let known = false;\n            let availableIds: string[] = [];\n            try {\n                const discovered = await discoverAllModels();\n                availableIds = discovered.map(m => m.id).slice(0, 50);\n                known =\n                    // a real, servable model — matched by full id or bare model name\n                    discovered.some(m => m.id === aliased || m.model === aliased || m.model === bareName)\n                    // OR an explicit, KNOWN provider prefix — the client deliberately\n                    // chose a provider, so let that provider validate the model name.\n                    || (aliased.includes('/') && getKnownProviderNames().includes(aliased.split('/')[0]));\n            } catch {\n                // Discovery failed — we can't prove the model is unknown, so fall\n                // back to the old resolve check rather than block a valid request.\n                known = !!tryResolveModel(aliased);\n            }\n\n            if (!known) {\n                res.status(404).json({\n                    error: {\n                        message: `Model '${model}' is not recognized by TITAN: no configured alias, not a servable model, and no known provider prefix. See 'available_models'.`,\n                        type: 'model_not_found',\n                        param: 'model',\n                        code: 'model_not_found',\n                        available_models: availableIds,\n                    },\n                });\n                return;\n            }\n        }\n\n        // ── v6.5 P0: cost transparency headers ──\n        // Sage flagged the cost-deception bug: if client sends 'gpt-4' + OpenAI\n        // key and TITAN routes to ollama/qwen, the client's OpenAI dashboard\n        // never moves. Set x-titan-provider so the actual route is at least\n        // VISIBLE in network logs even when the body says model='gpt-4'.\n        const routeInfo = tryResolveModel(effectiveModel);\n        const actualProvider = routeInfo?.provider?.name ?? 'unknown';\n        const actualModelId = routeInfo ? `${routeInfo.provider.name}/${routeInfo.model}` : effectiveModel;\n\n        logger.info(COMPONENT, `Chat completion: requested=${effectiveModel} → routed=${actualModelId}, messages=${messages.length}, stream=${!!stream}`);\n\n        // v6.5 P1 — cancel upstream generation when the client disconnects.\n        // Cline/Continue/Aider give up after 30-60s while a TITAN skill loop can\n        // run for minutes; without this we keep burning provider tokens to a dead\n        // socket. The signal flows chat/chatStream → router → provider fetch().\n        const ac = new AbortController();\n        req.on('close', () => ac.abort());\n\n        try {\n            if (stream) {\n                // SSE streaming response\n                res.writeHead(200, {\n                    'Content-Type': 'text/event-stream',\n                    'Cache-Control': 'no-cache',\n                    Connection: 'keep-alive',\n                    // v6.5 P0: cost transparency — surface the actual route\n                    // even when the body says model='gpt-4'. Clients can read\n                    // this header in their network logs / SDK middleware.\n                    'X-Titan-Provider': actualProvider,\n                    'X-Titan-Routed-Model': actualModelId,\n                    'X-Titan-Requested-Model': effectiveModel,\n                });\n                const sseWrite = setupSSEFlush(res);\n\n                for await (const chunk of chatStream({\n                    model: effectiveModel,\n                    messages,\n                    temperature: temperature ?? config.agent.temperature,\n                    maxTokens: max_tokens ?? config.agent.maxTokens,\n                    signal: ac.signal,\n                })) {\n                    if (chunk.type === 'text' && chunk.content) {\n                        const data = {\n                            id: chatId,\n                            object: 'chat.completion.chunk',\n                            created: Math.floor(Date.now() / 1000),\n                            model: effectiveModel,\n                            choices: [{\n                                index: 0,\n                                delta: { content: chunk.content },\n                                finish_reason: null,\n                            }],\n                        };\n                        sseWrite(`data: ${JSON.stringify(data)}\\n\\n`);\n                    }\n                }\n\n                // Send final chunk\n                sseWrite(`data: ${JSON.stringify({\n                    id: chatId,\n                    object: 'chat.completion.chunk',\n                    created: Math.floor(Date.now() / 1000),\n                    model: effectiveModel,\n                    choices: [{ index: 0, delta: {}, finish_reason: 'stop' }],\n                })}\\n\\n`);\n                sseWrite('data: [DONE]\\n\\n');\n                res.end();\n            } else {\n                // Non-streaming response\n                const response = await chat({\n                    model: effectiveModel,\n                    messages,\n                    temperature: temperature ?? config.agent.temperature,\n                    maxTokens: max_tokens ?? config.agent.maxTokens,\n                    signal: ac.signal,\n                });\n\n                // v6.5 P0: cost transparency headers on the non-stream path too\n                res.setHeader('X-Titan-Provider', actualProvider);\n                res.setHeader('X-Titan-Routed-Model', actualModelId);\n                res.setHeader('X-Titan-Requested-Model', effectiveModel);\n\n                res.json({\n                    id: chatId,\n                    object: 'chat.completion',\n                    created: Math.floor(Date.now() / 1000),\n                    model: response.model || effectiveModel,\n                    choices: [{\n                        index: 0,\n                        message: { role: 'assistant', content: response.content },\n                        finish_reason: response.finishReason || 'stop',\n                    }],\n                    usage: {\n                        prompt_tokens: response.usage?.promptTokens ?? 0,\n                        completion_tokens: response.usage?.completionTokens ?? 0,\n                        total_tokens: response.usage?.totalTokens ?? 0,\n                    },\n                });\n            }\n        } catch (err) {\n            // v6.5 P1/P3 — distinguish client-cancellation from real failures, and\n            // never throw ERR_HTTP_HEADERS_SENT on the streaming path. If the client\n            // disconnected (ac aborted) or the stream timed out, that's expected\n            // cancellation, not a server error — there's nothing to write to a dead\n            // socket. Once SSE headers are sent we can't change the status code, so\n            // just close the stream cleanly. X-Titan-Error-Class gives clients a\n            // back-channel for the non-streaming case.\n            const errMsg = (err as Error).message;\n            const errName = (err as Error).name;\n            const aborted = ac.signal.aborted || errName === 'AbortError' || errName === 'TimeoutError';\n            if (aborted) {\n                logger.info(COMPONENT, `Chat completion cancelled (client disconnect or timeout): ${errMsg}`);\n            } else {\n                logger.error(COMPONENT, `Chat completion error: ${errMsg}`);\n            }\n            if (res.headersSent) {\n                // streaming path — terminate the SSE stream without a status change\n                try { res.end(); } catch { /* socket already gone */ }\n            } else {\n                res.setHeader('X-Titan-Provider', actualProvider);\n                res.setHeader('X-Titan-Error-Class', aborted ? 'request_cancelled' : 'server_error');\n                res.status(aborted ? 499 : 500).json({\n                    error: { message: errMsg, type: aborted ? 'request_cancelled' : 'server_error' },\n                });\n            }\n        }\n    });\n\n    // POST /v1/embeddings — embedding generation (proxy to Ollama)\n    router.post('/embeddings', async (req: Request, res: Response) => {\n        const { model, input } = req.body;\n        const config = loadConfig();\n        const embeddingModel = model || (config.memory as Record<string, unknown>)?.embeddingModel || 'nomic-embed-text';\n\n        try {\n            // Proxy to Ollama embeddings endpoint\n            const ollamaUrl = config.providers?.ollama?.baseUrl || 'http://localhost:11434';\n            const texts = Array.isArray(input) ? input : [input];\n            const embeddings = [];\n\n            for (let i = 0; i < texts.length; i++) {\n                const resp = await fetch(`${ollamaUrl}/api/embed`, {\n                    method: 'POST',\n                    headers: { 'Content-Type': 'application/json' },\n                    body: JSON.stringify({ model: embeddingModel, input: texts[i] }),\n                });\n\n                if (!resp.ok) throw new Error(`Ollama embed failed: ${resp.status}`);\n                const data = await resp.json() as { embeddings?: number[][] };\n                embeddings.push({\n                    object: 'embedding',\n                    index: i,\n                    embedding: data.embeddings?.[0] || [],\n                });\n            }\n\n            res.json({\n                object: 'list',\n                data: embeddings,\n                model: embeddingModel,\n                usage: { prompt_tokens: 0, total_tokens: 0 },\n            });\n        } catch (err) {\n            res.status(500).json({\n                error: { message: (err as Error).message, type: 'server_error' },\n            });\n        }\n    });\n\n    return router;\n}\n"],"mappings":";AAKA,SAAS,cAA2C;AACpD,SAAS,MAAM,YAAY,iBAAiB,mBAAmB,6BAA6B;AAC5F,SAAS,kBAAkB;AAC3B,SAAS,MAAM,YAAY;AAC3B,OAAO,YAAY;AACnB,SAAS,qBAAqB;AAE9B,MAAM,YAAY;AAEX,SAAS,2BAAmC;AAC/C,QAAM,SAAS,OAAO;AAGtB,SAAO,IAAI,WAAW,CAAC,MAAe,QAAkB;AACpD,UAAM,SAAS,WAAW;AAC1B,UAAM,eAAe,OAAO,MAAM,SAAS;AAC3C,UAAM,UAAU,OAAO,MAAM,gBAAgB,CAAC;AAE9C,UAAM,SAAS;AAAA,MACX,EAAE,IAAI,cAAc,QAAQ,SAAS,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI,GAAG,UAAU,QAAQ;AAAA,MAC/F,GAAG,OAAO,QAAQ,OAAO,EACpB,OAAO,CAAC,CAAC,EAAE,CAAC,MAAM,KAAK,MAAM,YAAY,EACzC,IAAI,CAAC,CAAC,MAAM,KAAK,OAAO;AAAA,QACrB,IAAI;AAAA,QACJ,QAAQ;AAAA,QACR,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAAA,QACrC,UAAU;AAAA,MACd,EAAE;AAAA,IACV;AAEA,QAAI,KAAK,EAAE,QAAQ,QAAQ,MAAM,OAAO,CAAC;AAAA,EAC7C,CAAC;AAGD,SAAO,KAAK,qBAAqB,OAAO,KAAc,QAAkB;AACpE,UAAM,EAAE,OAAO,UAAU,QAAQ,aAAa,WAAW,IAAI,IAAI;AAEjE,QAAI,CAAC,YAAY,CAAC,MAAM,QAAQ,QAAQ,GAAG;AACvC,UAAI,OAAO,GAAG,EAAE,KAAK,EAAE,OAAO,EAAE,SAAS,wBAAwB,MAAM,wBAAwB,EAAE,CAAC;AAClG;AAAA,IACJ;AAEA,UAAM,SAAS,WAAW;AAC1B,UAAM,iBAAiB,SAAS,OAAO,MAAM;AAC7C,UAAM,SAAS,YAAY,KAAK,EAAE,MAAM,GAAG,EAAE,CAAC;AAS9C,QAAI,OAAO;AAOP,YAAM,UAAW,OAAO,MAAM,gBAAgB,CAAC;AAC/C,YAAM,UAAU,QAAQ,cAAc,KAAK;AAC3C,YAAM,WAAW,QAAQ,SAAS,GAAG,IAAI,QAAQ,MAAM,GAAG,EAAE,MAAM,CAAC,EAAE,KAAK,GAAG,IAAI;AAEjF,UAAI,QAAQ;AACZ,UAAI,eAAyB,CAAC;AAC9B,UAAI;AACA,cAAM,aAAa,MAAM,kBAAkB;AAC3C,uBAAe,WAAW,IAAI,OAAK,EAAE,EAAE,EAAE,MAAM,GAAG,EAAE;AACpD;AAAA,QAEI,WAAW,KAAK,OAAK,EAAE,OAAO,WAAW,EAAE,UAAU,WAAW,EAAE,UAAU,QAAQ,KAGhF,QAAQ,SAAS,GAAG,KAAK,sBAAsB,EAAE,SAAS,QAAQ,MAAM,GAAG,EAAE,CAAC,CAAC;AAAA,MAC3F,QAAQ;AAGJ,gBAAQ,CAAC,CAAC,gBAAgB,OAAO;AAAA,MACrC;AAEA,UAAI,CAAC,OAAO;AACR,YAAI,OAAO,GAAG,EAAE,KAAK;AAAA,UACjB,OAAO;AAAA,YACH,SAAS,UAAU,KAAK;AAAA,YACxB,MAAM;AAAA,YACN,OAAO;AAAA,YACP,MAAM;AAAA,YACN,kBAAkB;AAAA,UACtB;AAAA,QACJ,CAAC;AACD;AAAA,MACJ;AAAA,IACJ;AAOA,UAAM,YAAY,gBAAgB,cAAc;AAChD,UAAM,iBAAiB,WAAW,UAAU,QAAQ;AACpD,UAAM,gBAAgB,YAAY,GAAG,UAAU,SAAS,IAAI,IAAI,UAAU,KAAK,KAAK;AAEpF,WAAO,KAAK,WAAW,8BAA8B,cAAc,kBAAa,aAAa,cAAc,SAAS,MAAM,YAAY,CAAC,CAAC,MAAM,EAAE;AAMhJ,UAAM,KAAK,IAAI,gBAAgB;AAC/B,QAAI,GAAG,SAAS,MAAM,GAAG,MAAM,CAAC;AAEhC,QAAI;AACA,UAAI,QAAQ;AAER,YAAI,UAAU,KAAK;AAAA,UACf,gBAAgB;AAAA,UAChB,iBAAiB;AAAA,UACjB,YAAY;AAAA;AAAA;AAAA;AAAA,UAIZ,oBAAoB;AAAA,UACpB,wBAAwB;AAAA,UACxB,2BAA2B;AAAA,QAC/B,CAAC;AACD,cAAM,WAAW,cAAc,GAAG;AAElC,yBAAiB,SAAS,WAAW;AAAA,UACjC,OAAO;AAAA,UACP;AAAA,UACA,aAAa,eAAe,OAAO,MAAM;AAAA,UACzC,WAAW,cAAc,OAAO,MAAM;AAAA,UACtC,QAAQ,GAAG;AAAA,QACf,CAAC,GAAG;AACA,cAAI,MAAM,SAAS,UAAU,MAAM,SAAS;AACxC,kBAAM,OAAO;AAAA,cACT,IAAI;AAAA,cACJ,QAAQ;AAAA,cACR,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAAA,cACrC,OAAO;AAAA,cACP,SAAS,CAAC;AAAA,gBACN,OAAO;AAAA,gBACP,OAAO,EAAE,SAAS,MAAM,QAAQ;AAAA,gBAChC,eAAe;AAAA,cACnB,CAAC;AAAA,YACL;AACA,qBAAS,SAAS,KAAK,UAAU,IAAI,CAAC;AAAA;AAAA,CAAM;AAAA,UAChD;AAAA,QACJ;AAGA,iBAAS,SAAS,KAAK,UAAU;AAAA,UAC7B,IAAI;AAAA,UACJ,QAAQ;AAAA,UACR,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAAA,UACrC,OAAO;AAAA,UACP,SAAS,CAAC,EAAE,OAAO,GAAG,OAAO,CAAC,GAAG,eAAe,OAAO,CAAC;AAAA,QAC5D,CAAC,CAAC;AAAA;AAAA,CAAM;AACR,iBAAS,kBAAkB;AAC3B,YAAI,IAAI;AAAA,MACZ,OAAO;AAEH,cAAM,WAAW,MAAM,KAAK;AAAA,UACxB,OAAO;AAAA,UACP;AAAA,UACA,aAAa,eAAe,OAAO,MAAM;AAAA,UACzC,WAAW,cAAc,OAAO,MAAM;AAAA,UACtC,QAAQ,GAAG;AAAA,QACf,CAAC;AAGD,YAAI,UAAU,oBAAoB,cAAc;AAChD,YAAI,UAAU,wBAAwB,aAAa;AACnD,YAAI,UAAU,2BAA2B,cAAc;AAEvD,YAAI,KAAK;AAAA,UACL,IAAI;AAAA,UACJ,QAAQ;AAAA,UACR,SAAS,KAAK,MAAM,KAAK,IAAI,IAAI,GAAI;AAAA,UACrC,OAAO,SAAS,SAAS;AAAA,UACzB,SAAS,CAAC;AAAA,YACN,OAAO;AAAA,YACP,SAAS,EAAE,MAAM,aAAa,SAAS,SAAS,QAAQ;AAAA,YACxD,eAAe,SAAS,gBAAgB;AAAA,UAC5C,CAAC;AAAA,UACD,OAAO;AAAA,YACH,eAAe,SAAS,OAAO,gBAAgB;AAAA,YAC/C,mBAAmB,SAAS,OAAO,oBAAoB;AAAA,YACvD,cAAc,SAAS,OAAO,eAAe;AAAA,UACjD;AAAA,QACJ,CAAC;AAAA,MACL;AAAA,IACJ,SAAS,KAAK;AAQV,YAAM,SAAU,IAAc;AAC9B,YAAM,UAAW,IAAc;AAC/B,YAAM,UAAU,GAAG,OAAO,WAAW,YAAY,gBAAgB,YAAY;AAC7E,UAAI,SAAS;AACT,eAAO,KAAK,WAAW,6DAA6D,MAAM,EAAE;AAAA,MAChG,OAAO;AACH,eAAO,MAAM,WAAW,0BAA0B,MAAM,EAAE;AAAA,MAC9D;AACA,UAAI,IAAI,aAAa;AAEjB,YAAI;AAAE,cAAI,IAAI;AAAA,QAAG,QAAQ;AAAA,QAA4B;AAAA,MACzD,OAAO;AACH,YAAI,UAAU,oBAAoB,cAAc;AAChD,YAAI,UAAU,uBAAuB,UAAU,sBAAsB,cAAc;AACnF,YAAI,OAAO,UAAU,MAAM,GAAG,EAAE,KAAK;AAAA,UACjC,OAAO,EAAE,SAAS,QAAQ,MAAM,UAAU,sBAAsB,eAAe;AAAA,QACnF,CAAC;AAAA,MACL;AAAA,IACJ;AAAA,EACJ,CAAC;AAGD,SAAO,KAAK,eAAe,OAAO,KAAc,QAAkB;AAC9D,UAAM,EAAE,OAAO,MAAM,IAAI,IAAI;AAC7B,UAAM,SAAS,WAAW;AAC1B,UAAM,iBAAiB,SAAU,OAAO,QAAoC,kBAAkB;AAE9F,QAAI;AAEA,YAAM,YAAY,OAAO,WAAW,QAAQ,WAAW;AACvD,YAAM,QAAQ,MAAM,QAAQ,KAAK,IAAI,QAAQ,CAAC,KAAK;AACnD,YAAM,aAAa,CAAC;AAEpB,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACnC,cAAM,OAAO,MAAM,MAAM,GAAG,SAAS,cAAc;AAAA,UAC/C,QAAQ;AAAA,UACR,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,UAC9C,MAAM,KAAK,UAAU,EAAE,OAAO,gBAAgB,OAAO,MAAM,CAAC,EAAE,CAAC;AAAA,QACnE,CAAC;AAED,YAAI,CAAC,KAAK,GAAI,OAAM,IAAI,MAAM,wBAAwB,KAAK,MAAM,EAAE;AACnE,cAAM,OAAO,MAAM,KAAK,KAAK;AAC7B,mBAAW,KAAK;AAAA,UACZ,QAAQ;AAAA,UACR,OAAO;AAAA,UACP,WAAW,KAAK,aAAa,CAAC,KAAK,CAAC;AAAA,QACxC,CAAC;AAAA,MACL;AAEA,UAAI,KAAK;AAAA,QACL,QAAQ;AAAA,QACR,MAAM;AAAA,QACN,OAAO;AAAA,QACP,OAAO,EAAE,eAAe,GAAG,cAAc,EAAE;AAAA,MAC/C,CAAC;AAAA,IACL,SAAS,KAAK;AACV,UAAI,OAAO,GAAG,EAAE,KAAK;AAAA,QACjB,OAAO,EAAE,SAAU,IAAc,SAAS,MAAM,eAAe;AAAA,MACnE,CAAC;AAAA,IACL;AAAA,EACJ,CAAC;AAED,SAAO;AACX;","names":[]}

package/dist/gateway/server.js

@@ -33,6 +33,7 @@ import { MattermostChannel } from "../channels/mattermost.js";
 import { LarkChannel } from "../channels/lark.js";
 import { EmailInboundChannel } from "../channels/email_inbound.js";
 import { LineChannel } from "../channels/line.js";
+import { kill as killSwitchFire, resume as killSwitchResume, getState as getKillSwitchState } from "../safety/killSwitch.js";
 import { ZulipChannel } from "../channels/zulip.js";
 import { MessengerChannel } from "../channels/messenger.js";
 import { initAgents, routeMessage, listAgents } from "../agent/multiAgent.js";
@@ -835,6 +836,31 @@ async function startGateway(options) {
     });
     next();
   });
+  app.use("/v1", (req, res, next) => {
+    const cfg = loadConfig();
+    const auth = cfg.gateway.auth;
+    if (!auth || auth.mode === "none") {
+      next();
+      return;
+    }
+    if (auth.mode === "token" && !auth.token) {
+      next();
+      return;
+    }
+    const header = req.headers.authorization;
+    const token = header?.startsWith("Bearer ") ? header.slice(7) : req.query.token;
+    if (isValidToken(token, cfg)) {
+      next();
+      return;
+    }
+    res.status(401).json({
+      error: {
+        message: "Unauthorized. /v1 requires the TITAN gateway token as the Bearer credential when auth is enabled.",
+        type: "invalid_request_error",
+        code: "unauthorized"
+      }
+    });
+  });
   app.use("/v1", createOpenAICompatRouter());
   app.use("/api/paperclip", createPaperclipRouter());
   app.use("/paperclip", createPaperclipUIRouter());
@@ -1002,6 +1028,14 @@ async function startGateway(options) {
       next();
       return;
     }
+    if (req.path === "/channels/line/webhook") {
+      next();
+      return;
+    }
+    if (req.path === "/channels/lark/webhook") {
+      next();
+      return;
+    }
     const header = req.headers.authorization;
     const token = header?.startsWith("Bearer ") ? header.slice(7) : req.query.token;
     if (isValidToken(token, cfg)) {
@@ -3571,6 +3605,51 @@ data: ${JSON.stringify(structured)}
     });
     logger.info(COMPONENT, `Messenger webhook: /api/messenger/webhook + /messenger/webhook (verify token: ${messengerAdapter.getVerifyToken()})`);
   }
+  const lineAdapter = channels.get("line");
+  if (lineAdapter) {
+    app.post("/api/channels/line/webhook", express.json(), (req, res) => {
+      try {
+        lineAdapter.handleWebhookEvent(req.body || {});
+      } catch (e) {
+        logger.error(COMPONENT, `LINE webhook error: ${e.message}`);
+      }
+      res.status(200).end();
+    });
+    logger.info(COMPONENT, "LINE webhook: POST /api/channels/line/webhook");
+  }
+  const larkAdapter = channels.get("lark");
+  if (larkAdapter) {
+    app.post("/api/channels/lark/webhook", express.json(), (req, res) => {
+      try {
+        const result = larkAdapter.handleWebhookEvent(req.body || {});
+        if (result?.challenge) {
+          res.json({ challenge: result.challenge });
+          return;
+        }
+      } catch (e) {
+        logger.error(COMPONENT, `Lark webhook error: ${e.message}`);
+      }
+      res.status(200).end();
+    });
+    logger.info(COMPONENT, "Lark webhook: POST /api/channels/lark/webhook");
+  }
+  app.get("/api/safety/kill-switch", (_req, res) => {
+    res.json(getKillSwitchState());
+  });
+  app.post("/api/safety/kill", express.json(), async (req, res) => {
+    const reason = req.body?.reason || "Manual kill via API";
+    try {
+      await killSwitchFire("manual", reason, { firedBy: req.body?.firedBy || "api" });
+      res.json({ ok: true, state: getKillSwitchState() });
+    } catch (e) {
+      res.status(500).json({ error: e.message });
+    }
+  });
+  app.post("/api/safety/resume", express.json(), (req, res) => {
+    const note = req.body?.resolutionNote || req.body?.reason || "Resumed via API";
+    const state = killSwitchResume(note, req.body?.resumedBy || "api");
+    res.json({ ok: true, state });
+  });
   {
     const {
       twimlPlayAndGather,