feat: wire chat to /api/chat/stream SSE endpoint

apps/web/src/lib/api/chat.ts

@@ -1,6 +1,6 @@
 /**
  * Chat API client
- * Handles LLM chat interactions via /api/llm/chat
+ * Handles LLM chat interactions via /api/chat/stream (streaming) and /api/llm/chat (fallback)
  */
 
 import { apiPost, fetchCsrfToken, getCsrfToken } from "./client";
@@ -33,9 +33,28 @@ export interface ChatResponse {
 }
 
 /**
- * Parsed SSE data chunk from the LLM stream
+ * Parsed SSE data chunk from OpenAI-compatible stream
  */
-interface SseChunk {
+interface OpenAiSseChunk {
+  id?: string;
+  object?: string;
+  created?: number;
+  model?: string;
+  choices?: {
+    index: number;
+    delta?: {
+      role?: string;
+      content?: string;
+    };
+    finish_reason?: string | null;
+  }[];
+  error?: string;
+}
+
+/**
+ * Parsed SSE data chunk from legacy /api/llm/chat stream
+ */
+interface LegacySseChunk {
   error?: string;
   message?: {
     role: string;
@@ -46,7 +65,17 @@ interface SseChunk {
 }
 
 /**
- * Send a chat message to the LLM
+ * Parsed SSE data chunk with simple token format
+ */
+interface SimpleTokenChunk {
+  token?: string;
+  done?: boolean;
+  error?: string;
+}
+
+/**
+ * Send a chat message to the LLM (non-streaming fallback)
+ * Uses /api/llm/chat endpoint which supports both streaming and non-streaming
  */
 export async function sendChatMessage(request: ChatRequest): Promise<ChatResponse> {
   return apiPost<ChatResponse>("/api/llm/chat", request);
@@ -66,11 +95,20 @@ async function ensureCsrfTokenForStream(): Promise<string> {
 /**
  * Stream a chat message from the LLM using SSE over fetch.
  *
- * The backend accepts stream: true in the request body and responds with
- * Server-Sent Events:
- *   data: {"message":{"content":"token"},...}\n\n  for each token
- *   data: [DONE]\n\n  when the stream is complete
- *   data: {"error":"message"}\n\n  on error
+ * Uses /api/chat/stream endpoint which proxies to OpenClaw.
+ * The backend responds with Server-Sent Events in one of these formats:
+ *
+ * OpenAI-compatible format:
+ *   data: {"choices":[{"delta":{"content":"token"}}],...}\n\n
+ *   data: [DONE]\n\n
+ *
+ * Legacy format (from /api/llm/chat):
+ *   data: {"message":{"content":"token"},...}\n\n
+ *   data: [DONE]\n\n
+ *
+ * Simple token format:
+ *   data: {"token":"..."}\n\n
+ *   data: {"done":true}\n\n
  *
  * @param request - Chat request (stream field will be forced to true)
  * @param onChunk - Called with each token string as it arrives
@@ -89,14 +127,14 @@ export function streamChatMessage(
     try {
       const csrfToken = await ensureCsrfTokenForStream();
 
-      const response = await fetch(`${API_BASE_URL}/api/llm/chat`, {
+      const response = await fetch(`${API_BASE_URL}/api/chat/stream`, {
         method: "POST",
         headers: {
           "Content-Type": "application/json",
           "X-CSRF-Token": csrfToken,
         },
         credentials: "include",
-        body: JSON.stringify({ ...request, stream: true }),
+        body: JSON.stringify({ messages: request.messages, stream: true }),
         signal: signal ?? null,
       });
 
@@ -132,6 +170,25 @@ export function streamChatMessage(
           const trimmed = part.trim();
           if (!trimmed) continue;
 
+          // Handle event: error format
+          const eventMatch = /^event:\s*(\S+)\n/i.exec(trimmed);
+          const dataMatch = /^data:\s*(.+)$/im.exec(trimmed);
+
+          if (eventMatch?.[1] === "error" && dataMatch?.[1]) {
+            try {
+              const errorData = JSON.parse(dataMatch[1].trim()) as {
+                error?: string;
+              };
+              throw new Error(errorData.error ?? "Stream error occurred");
+            } catch (parseErr) {
+              if (parseErr instanceof SyntaxError) {
+                throw new Error("Stream error occurred");
+              }
+              throw parseErr;
+            }
+          }
+
+          // Standard SSE format: data: {...}
           for (const line of trimmed.split("\n")) {
             if (!line.startsWith("data: ")) continue;
 
@@ -143,14 +200,39 @@ export function streamChatMessage(
             }
 
             try {
-              const parsed = JSON.parse(data) as SseChunk;
+              const parsed: unknown = JSON.parse(data);
 
-              if (parsed.error) {
-                throw new Error(parsed.error);
+              // Handle OpenAI format (from /api/chat/stream via OpenClaw)
+              const openAiChunk = parsed as OpenAiSseChunk;
+              if (openAiChunk.choices?.[0]?.delta?.content) {
+                onChunk(openAiChunk.choices[0].delta.content);
+                continue;
               }
 
-              if (parsed.message?.content) {
-                onChunk(parsed.message.content);
+              // Handle legacy format (from /api/llm/chat)
+              const legacyChunk = parsed as LegacySseChunk;
+              if (legacyChunk.message?.content) {
+                onChunk(legacyChunk.message.content);
+                continue;
+              }
+
+              // Handle simple token format
+              const simpleChunk = parsed as SimpleTokenChunk;
+              if (simpleChunk.token) {
+                onChunk(simpleChunk.token);
+                continue;
+              }
+
+              // Handle done flag in simple format
+              if (simpleChunk.done === true) {
+                onComplete();
+                return;
+              }
+
+              // Handle error in any format
+              const error = openAiChunk.error ?? legacyChunk.error ?? simpleChunk.error;
+              if (error) {
+                throw new Error(error);
               }
             } catch (parseErr) {
               if (parseErr instanceof SyntaxError) {
@@ -162,7 +244,7 @@ export function streamChatMessage(
         }
       }
 
-      // Natural end of stream without [DONE]
+      // Natural end of stream without [DONE] or done flag
       onComplete();
     } catch (err: unknown) {
       if (err instanceof DOMException && err.name === "AbortError") {