u

.env.example

@@ -36,6 +36,8 @@ KAT_LISTS=0
 # Tax zone name for VAT rates (tSteuerzone.cName)
 STEUER_ZONE=Zone-EU
 
+# OpenRouter (leave OPENROUTER_API_URL unset for default OpenRouter endpoint)
+# OPENROUTER_API_URL=http://127.0.0.1:8787/api/v1/chat/completions
 OPENROUTER_FREE_KEY=sk-or-v1-3d3c4e
 OPENROUTER_FREE_MODEL=openrouter/free
 OPENROUTER_PAID_MODEL=openai/gpt-6-luna:floor

docs/api/routes/modelconversation/_id/stream.md

@@ -5,11 +5,11 @@ Protected. Requires a valid session cookie (`credentials: 'include'`).
 Stream conversation events (SSE)
 
 Opens a Server-Sent Events stream for the specified conversation.
-Streams `message` events (persisted user/assistant messages), live `delta`
-text chunks of the model answer, and `state` events (`thinking`, `working`,
-`done`). The SSE states are not stored; the database only knows `idle`/`busy`.
-`done` carries `usage` (`input`, `cached`, `output`, `cost`) on success or
-`error` (string) if the run failed.
+Streams live `delta` text chunks of the model answer, `usage` stats per turn,
+and `state` events (`thinking`, `connected`, `firstevent`, `reasoning`,
+`content`, `working`, `toolcall`, `done`). The SSE states are not stored;
+the database only knows `idle`/`busy`.
+`done` indicates generation has finished, carrying `error` (string) only if the run failed.
 
 
 ## Path parameters
@@ -20,7 +20,7 @@ text chunks of the model answer, and `state` events (`thinking`, `working`,
 
 ## Response `200`
 
-SSE stream (`text/event-stream`). Emits `state`, `delta`, and `message` events.
+SSE stream (`text/event-stream`). Emits `state`, `delta`, and `usage` events.
 
 
 ## Errors

src/api/routes/deliveryParse.js

@@ -296,7 +296,7 @@ async function parseDocumentManifest({ req, res, pages, sourceLabel }) {
 
     // ── LLMOrbiter setup ─────────────────────────────────────────────
     const model = await resolveDefaultModel(req.user.id);
-    const provider = new OpenRouterProvider({ apiKey });
+    const provider = new OpenRouterProvider({ apiKey, apiUrl: env.openrouterApiUrl });
     const orbiter = new LLMOrbiter(provider, {
         onStatus: (status) => {
             console.log(`[llmorbiter] Status: ${status}`);

src/api/routes/modelconversation.js

@@ -313,11 +313,11 @@ export function register(app, basePath = '') {
      *     summary: Stream conversation events (SSE)
      *     description: |
      *       Opens a Server-Sent Events stream for the specified conversation.
-     *       Streams `message` events (persisted user/assistant messages), live `delta`
-     *       text chunks of the model answer, and `state` events (`thinking`, `working`,
-     *       `done`). The SSE states are not stored; the database only knows `idle`/`busy`.
-     *       `done` carries `usage` (`input`, `cached`, `output`, `cost`) on success or
-     *       `error` (string) if the run failed.
+     *       Streams live `delta` text chunks of the model answer, `usage` stats per turn,
+     *       and `state` events (`thinking`, `connected`, `firstevent`, `reasoning`,
+     *       `content`, `working`, `toolcall`, `done`). The SSE states are not stored;
+     *       the database only knows `idle`/`busy`.
+     *       `done` indicates generation has finished, carrying `error` (string) only if the run failed.
      *     security:
      *       - cookieAuth: []
      *     parameters:
@@ -330,7 +330,7 @@ export function register(app, basePath = '') {
      *     responses:
      *       200:
      *         description: |
-     *           SSE stream (`text/event-stream`). Emits `state`, `delta`, and `message` events.
+     *           SSE stream (`text/event-stream`). Emits `state`, `delta`, and `usage` events.
      *         content:
      *           text/event-stream:
      *             schema:
@@ -340,16 +340,22 @@ export function register(app, basePath = '') {
      *                 data: {"state":"thinking"}
      *
      *                 event: state
-     *                 data: {"state":"working"}
+     *                 data: {"state":"connected"}
+     *
+     *                 event: state
+     *                 data: {"state":"firstevent"}
+     *
+     *                 event: state
+     *                 data: {"state":"content"}
      *
      *                 event: delta
      *                 data: {"delta":"Hel","text":"Hel"}
      *
-     *                 event: message
-     *                 data: {"id":2,"conversation_id":1,"created_at":"...","role":"assistant","message":"Hello there"}
+     *                 event: usage
+     *                 data: {"input":20,"cached":0,"output":3,"cost":0.0001}
      *
      *                 event: state
-     *                 data: {"state":"done","usage":{"input":20,"cached":0,"output":3,"cost":0.0001}}
+     *                 data: {"state":"done"}
      *       400:
      *         description: Invalid ID or no company selected
      *         content:

src/config/env.js

@@ -26,6 +26,7 @@ export default {
     implicitCreate: process.env.IMPLICIT_CREATE === 'true',
     jwtSecret: process.env.JWT_SECRET || 'fallback_secret_please_change_in_production',
     dbProvider: process.env.DB_PROVIDER || (process.env.SQLITE_PATH ? 'sqlite' : 'mssql'),
+    openrouterApiUrl: process.env.OPENROUTER_API_URL,
     openrouterFreeKey: process.env.OPENROUTER_FREE_KEY,
     openrouterFreeModel: process.env.OPENROUTER_FREE_MODEL || 'openrouter/free',
     openrouterPaidModel: process.env.OPENROUTER_PAID_MODEL || 'openai/gpt-6-luna:floor',

src/services/conversationManager.js

@@ -303,7 +303,6 @@ export class ConversationManager {
             throw err;
         }
 
-        this.broadcast('message', userMsg);
         this.broadcast('state', { state: 'thinking' });
         emitCompanyEvent('modelconversation', { companyId: this.companyId });
 
@@ -343,7 +342,6 @@ export class ConversationManager {
             throw err;
         }
 
-        this.broadcast('message', toolMsg);
         this.broadcast('state', { state: 'working' });
         emitCompanyEvent('modelconversation', { companyId: this.companyId });
 
@@ -361,7 +359,6 @@ export class ConversationManager {
      * @param {number} googleAuthId
      */
     async runGeneration(googleAuthId) {
-        let usage = null;
         let error = null;
 
         try {
@@ -397,33 +394,23 @@ export class ConversationManager {
                 ...history.map(toLlmMessage),
             ];
 
-            let working = false;
             let text = '';
             let currentReasoning = '';
             let savedToolCallsTurn = null;
 
-            const markWorking = () => {
-                if (working) return;
-                working = true;
-                this.broadcast('state', { state: 'working' });
-            };
-
-            this.orbiter = new LLMOrbiter(new OpenRouterProvider({ apiKey }), {
-                onStatus: (status) => {
-                    if (status === 'working') markWorking();
+            this.orbiter = new LLMOrbiter(new OpenRouterProvider({ apiKey, apiUrl: env.openrouterApiUrl }), {
+                onStatus: (status, meta) => {
+                    this.broadcast('state', { state: status, ...(meta || {}) });
                 },
                 onReasoning: (reasoning) => {
-                    markWorking();
                     currentReasoning = reasoning;
                     this.broadcast('reasoning', { text: reasoning });
                 },
                 onContent: (chunk) => {
-                    markWorking();
                     text += chunk;
                     this.broadcast('delta', { delta: chunk, text });
                 },
                 onToolCall: async ({ name, args, id, callItem, tool_calls, reasoning_details, reasoning, content }) => {
-                    markWorking();
                     try {
                         const effectiveToolCalls = tool_calls || (callItem ? [callItem] : [{
                             id: id || randomUUID(),
@@ -449,12 +436,11 @@ export class ConversationManager {
                                 ...(effectiveReasoning ? { reasoning: effectiveReasoning } : {}),
                             };
 
-                            const callMsg = await db.addMessage(googleAuthId, this.conversationId, {
+                            await db.addMessage(googleAuthId, this.conversationId, {
                                 role: 'assistant',
                                 message: JSON.stringify(assistantPayload),
                             });
                             text = '';
-                            this.broadcast('message', callMsg);
                             emitCompanyEvent('modelconversation', { companyId: this.companyId });
                         }
                     } catch (err) {
@@ -462,15 +448,23 @@ export class ConversationManager {
                     }
                 },
                 onToolResult: async ({ name, result, id, callItem }) => {
+                    const rawArgs = callItem?.function?.arguments || '';
+                    this.broadcast('state', {
+                        state: 'toolcall',
+                        id: id || callItem?.id || '',
+                        name,
+                        arguments: rawArgs,
+                        content: rawArgs,
+                        result,
+                    });
                     try {
-                        const resMsg = await db.addMessage(googleAuthId, this.conversationId, {
+                        await db.addMessage(googleAuthId, this.conversationId, {
                             role: 'tool',
                             message: JSON.stringify({
                                 content: typeof result === 'string' ? result : JSON.stringify(result),
                                 tool_call_id: id || callItem?.id || '',
                             }),
                         });
-                        this.broadcast('message', resMsg);
                         emitCompanyEvent('modelconversation', { companyId: this.companyId });
                     } catch (err) {
                         console.error(`[ConversationManager ${this.conversationId}] failed to save tool result message:`, err);
@@ -478,21 +472,37 @@ export class ConversationManager {
                 },
                 onToolError: async ({ name, error, id, callItem }) => {
                     console.error(`[ConversationManager ${this.conversationId}] Tool "${name}" errored:`, error);
+                    const rawArgs = callItem?.function?.arguments || '';
+                    this.broadcast('state', {
+                        state: 'toolcall',
+                        id: id || callItem?.id || '',
+                        name,
+                        arguments: rawArgs,
+                        content: rawArgs,
+                        error: String(error),
+                    });
                     try {
-                        const errMsg = await db.addMessage(googleAuthId, this.conversationId, {
+                        await db.addMessage(googleAuthId, this.conversationId, {
                             role: 'tool',
                             message: JSON.stringify({
                                 content: `Tool error [${name}]: ${error}`,
                                 tool_call_id: id || callItem?.id || '',
                             }),
                         });
-                        this.broadcast('message', errMsg);
                         emitCompanyEvent('modelconversation', { companyId: this.companyId });
                     } catch (err) {
                         console.error(`[ConversationManager ${this.conversationId}] failed to save tool error message:`, err);
                     }
                 },
                 onUsage: (u) => {
+                    if (u) {
+                        this.broadcast('usage', {
+                            input: u.input ?? 0,
+                            cached: u.cached ?? 0,
+                            output: u.output ?? 0,
+                            ...(u.cost != null ? { cost: u.cost } : {}),
+                        });
+                    }
                     // A call that consumed 0 input tokens AND produced 0 output means the
                     // provider returned nothing at all — not a real generation. LLMOrbiter
                     // treats "no content and no tool calls" as "keep looping", so several of
@@ -577,7 +587,7 @@ export class ConversationManager {
             }
 
             if (content || finalReasoningDetails || finalReasoning) {
-                const assistantMsg = await db.addMessage(googleAuthId, this.conversationId, {
+                await db.addMessage(googleAuthId, this.conversationId, {
                     message: JSON.stringify({
                         content: content || '',
                         ...(finalReasoningDetails ? { reasoning_details: finalReasoningDetails } : {}),
@@ -585,15 +595,7 @@ export class ConversationManager {
                     }),
                     role: 'assistant',
                 });
-                this.broadcast('message', assistantMsg);
             }
-
-            usage = {
-                input: result.usage.input,
-                cached: result.usage.cached,
-                output: result.usage.output,
-                cost: result.usage.cost,
-            };
         } catch (err) {
             if (this.destroyed) return;
             error = err.message;
@@ -609,7 +611,7 @@ export class ConversationManager {
             console.error(`[ConversationManager ${this.conversationId}] failed to reset state:`, err);
         }
         emitCompanyEvent('modelconversation', { companyId: this.companyId });
-        this.broadcast('state', error ? { state: 'done', error } : { state: 'done', usage });
+        this.broadcast('state', error ? { state: 'done', error } : { state: 'done' });
     }
 
     destroy() {

src/tools/analyze-image.js

@@ -373,7 +373,7 @@ export function createAnalyzeImageTool({ googleAuthId, model }) {
                 : question || 'Describe this image, judge its visual quality, flag any watermark, text overlay or defect, and say whether it is usable as a product or category image.';
 
             const apiKey = await resolveOpenRouterKey(googleAuthId);
-            const provider = new OpenRouterProvider({ apiKey });
+            const provider = new OpenRouterProvider({ apiKey, apiUrl: env.openrouterApiUrl });
             const response = await provider.chatCompletions({
                 model,
                 messages: [