clawmatrix 0.1.13 → 0.1.15

package/src/model-proxy.ts

@@ -11,12 +11,17 @@ import type {
 import { debug } from "./debug.ts";
 
 const MODEL_TIMEOUT = 120_000; // 2 minutes
+const MAX_STREAM_BUFFER = 1_048_576; // 1MB — guard against upstream not sending newlines
+
+type ResponseFormat = "chat" | "responses";
 
 interface PendingModelReq {
   resolve: (value: unknown) => void;
   reject: (error: Error) => void;
   timer: ReturnType<typeof setTimeout>;
   stream: boolean;
+  responseFormat: ResponseFormat;
+  model?: string;
   controller?: ReadableStreamDefaultController;
   encoder?: TextEncoder;
 }
@@ -36,35 +41,125 @@ export class ModelProxy {
     this.openclawConfig = openclawConfig;
   }
 
+  /**
+   * Normalize Responses API `input` to OpenAI chat messages for WS transport.
+   *
+   * Converts:
+   *   - string → [{role: "user", content: "..."}]
+   *   - shorthand {role, content: "..."} → pass through
+   *   - full {type: "message", content: [{type: "input_text"}, {type: "input_image"}]} → chat format
+   *   - {type: "function_call_output"} → {role: "tool", ...}
+   *
+   * Returns chat-completions compatible messages (text + image_url content parts).
+   */
+  private static normalizeResponsesInput(input: unknown): unknown[] {
+    if (typeof input === "string") {
+      return [{ role: "user", content: input }];
+    }
+    if (!Array.isArray(input)) return [];
+
+    const messages: unknown[] = [];
+    for (const item of input) {
+      if (!item || typeof item !== "object") continue;
+      const obj = item as Record<string, unknown>;
+
+      // function_call_output → tool message
+      if (obj.type === "function_call_output") {
+        messages.push({
+          role: "tool",
+          tool_call_id: obj.call_id,
+          content: typeof obj.output === "string" ? obj.output : JSON.stringify(obj.output),
+        });
+        continue;
+      }
+
+      const role = typeof obj.role === "string" ? obj.role : "user";
+
+      // Simple shorthand: {role: "user", content: "hello"}
+      if (typeof obj.content === "string") {
+        messages.push({ role, content: obj.content });
+        continue;
+      }
+
+      // Full format: {type: "message", role, content: [{type: "input_text"|"input_image"|...}]}
+      if (Array.isArray(obj.content)) {
+        const parts: unknown[] = [];
+        for (const part of obj.content) {
+          if (!part || typeof part !== "object") continue;
+          const p = part as Record<string, unknown>;
+
+          if (p.type === "input_text" || p.type === "output_text") {
+            // Text content → chat text part
+            if (typeof p.text === "string") {
+              parts.push({ type: "text", text: p.text });
+            }
+          } else if (p.type === "input_image") {
+            // Image content → chat image_url part
+            if (typeof p.image_url === "string") {
+              parts.push({ type: "image_url", image_url: { url: p.image_url } });
+            } else if (p.image_url && typeof p.image_url === "object") {
+              parts.push({ type: "image_url", image_url: p.image_url });
+            }
+          } else if (p.type === "text" && typeof p.text === "string") {
+            // Already chat format
+            parts.push(p);
+          } else if (p.type === "image_url") {
+            // Already chat format
+            parts.push(p);
+          }
+        }
+
+        if (parts.length === 1 && (parts[0] as Record<string, unknown>).type === "text") {
+          // Single text part → simplify to string content
+          messages.push({ role, content: ((parts[0] as Record<string, unknown>).text as string) });
+        } else if (parts.length > 0) {
+          messages.push({ role, content: parts });
+        }
+        continue;
+      }
+
+      // Fallback
+      if (typeof obj.text === "string") {
+        messages.push({ role, content: obj.text });
+      }
+    }
+    return messages;
+  }
+
   /** Resolve API endpoint for a model: explicit config > OpenClaw provider > gateway fallback */
-  private resolveModelEndpoint(model: { id: string; provider: string; baseUrl?: string; apiKey?: string }): { url: string; apiKey?: string; direct: boolean } {
+  private resolveModelEndpoint(model: { id: string; provider: string; baseUrl?: string; apiKey?: string; api?: string }): { baseUrl: string; apiKey?: string; direct: boolean; api: string } {
+    const defaultApi = "openai-completions";
+
     // 1. Explicit baseUrl in ClawMatrix model config
     if (model.baseUrl) {
       return {
-        url: `${model.baseUrl.replace(/\/$/, "")}/chat/completions`,
+        baseUrl: model.baseUrl.replace(/\/$/, ""),
         apiKey: model.apiKey,
         direct: true,
+        api: model.api ?? defaultApi,
       };
     }
 
     // 2. Read from OpenClaw's models.providers[provider]
     const providers = (this.openclawConfig as Record<string, unknown>).models as
-      { providers?: Record<string, { baseUrl?: string; apiKey?: string }> } | undefined;
+      { providers?: Record<string, { baseUrl?: string; apiKey?: string; api?: string }> } | undefined;
     const providerConfig = providers?.providers?.[model.provider];
     if (providerConfig?.baseUrl) {
       return {
-        url: `${providerConfig.baseUrl.replace(/\/$/, "")}/chat/completions`,
+        baseUrl: providerConfig.baseUrl.replace(/\/$/, ""),
         apiKey: typeof providerConfig.apiKey === "string" ? providerConfig.apiKey : undefined,
         direct: true,
+        api: model.api ?? providerConfig.api ?? defaultApi,
       };
     }
 
-    // 3. Fallback: OpenClaw gateway (goes through agent system — not recommended)
+    // 3. Fallback: OpenClaw gateway
     const { port } = this.gatewayInfo;
     return {
-      url: `http://127.0.0.1:${port}/v1/chat/completions`,
+      baseUrl: `http://127.0.0.1:${port}/v1`,
       apiKey: undefined,
       direct: false,
+      api: model.api ?? defaultApi,
     };
   }
 
@@ -79,7 +174,12 @@ export class ModelProxy {
 
         if (p === "/chat/completions" && req.method === "POST") {
           const body = await this.readBody(req);
-          const response = await this.handleChatCompletion(body);
+          const response = await this.handleChatCompletion(body, "openai-completions");
+          debug("proxy", `response status=${response.status}`);
+          this.sendResponse(res, response);
+        } else if (p === "/responses" && req.method === "POST") {
+          const body = await this.readBody(req);
+          const response = await this.handleResponses(body);
           debug("proxy", `response status=${response.status}`);
           this.sendResponse(res, response);
         } else if (p === "/models" && req.method === "GET") {
@@ -152,55 +252,56 @@ export class ModelProxy {
   }
 
   // ── HTTP handlers ──────────────────────────────────────────────
-  private async handleChatCompletion(rawBody: string): Promise<{ status: number; headers: Record<string, string>; body: string | ReadableStream }> {
-    let body: {
-      model: string;
-      messages: unknown[];
-      stream?: boolean;
-      temperature?: number;
-      max_tokens?: number;
-    };
 
-    try {
-      body = JSON.parse(rawBody);
-    } catch {
-      return {
-        status: 400,
-        headers: { "Content-Type": "application/json" },
-        body: JSON.stringify({ error: "Invalid JSON" }),
-      };
-    }
-
-    const rawModelId = body.model;
-    // Parse "nodeId/model" format: first segment is nodeId, rest is model ID.
-    // OpenClaw sends "providerId/modelId" where providerId = nodeId, so this
-    // naturally handles both OpenClaw calls and direct curl calls.
-    // If no "/" present, treat entire string as model ID and auto-resolve.
-    let nodeId: string | undefined;
-    let modelId: string;
+  /** Resolve model ID → proxyModel + route. Shared by chat completions and responses handlers. */
+  private resolveModelRoute(rawModelId: string): {
+    nodeId: string; modelId: string;
+    proxyModel: (typeof this.config.proxyModels)[number] | undefined;
+    routeNodeId: string;
+  } | { error: { status: number; message: string } } {
     const slashIdx = rawModelId.indexOf("/");
+    let nodeId: string;
+    let modelId: string;
+    let proxyModel: (typeof this.config.proxyModels)[number] | undefined;
+
     if (slashIdx > 0) {
       nodeId = rawModelId.slice(0, slashIdx);
       modelId = rawModelId.slice(slashIdx + 1);
+      proxyModel = this.config.proxyModels.find((m) => m.id === modelId && m.nodeId === nodeId);
     } else {
       modelId = rawModelId;
+      proxyModel = this.config.proxyModels.find((m) => m.id === modelId);
+      if (!proxyModel) {
+        return { error: { status: 404, message: `Model "${rawModelId}" not found in proxy models` } };
+      }
+      nodeId = proxyModel.nodeId;
     }
-    debug("proxy", `model raw="${rawModelId}" nodeId=${nodeId ?? "auto"} modelId="${modelId}" stream=${body.stream ?? false}`);
-    const proxyModel = this.config.proxyModels.find((m) => m.id === modelId && (!nodeId || m.nodeId === nodeId))
-      ?? this.config.proxyModels.find((m) => m.id === modelId);
-    const route = nodeId
-      ? this.peerManager.router.getRoute(nodeId)
-      : this.peerManager.router.resolveModel(modelId);
-    debug("proxy", `proxyModel=${proxyModel?.id ?? "none"} route=${route?.nodeId ?? "none"} reachable=${route ? this.peerManager.canReach(route.nodeId) : false}`);
+
+    const route = this.peerManager.router.getRoute(nodeId);
+    debug("proxy", `model raw="${rawModelId}" nodeId=${nodeId} modelId="${modelId}" route=${route?.nodeId ?? "none"}`);
     if (!route) {
-      return {
-        status: 404,
-        headers: { "Content-Type": "application/json" },
-        body: JSON.stringify({ error: { message: `Model "${modelId}" not found in cluster (proxyModels: [${this.config.proxyModels.map(m => m.id).join(", ")}])` } }),
-      };
+      return { error: { status: 404, message: `Node "${nodeId}" not found in cluster` } };
     }
+    if (!this.peerManager.canReach(route.nodeId)) {
+      return { error: { status: 502, message: `Cannot reach model node "${route.nodeId}"` } };
+    }
+    return { nodeId, modelId, proxyModel, routeNodeId: route.nodeId };
+  }
 
-    // Inject model identity so the LLM knows what it is
+  private async handleChatCompletion(rawBody: string, _api: string): Promise<{ status: number; headers: Record<string, string>; body: string | ReadableStream }> {
+    let body: { model: string; messages: unknown[]; stream?: boolean; temperature?: number; max_tokens?: number };
+    try {
+      body = JSON.parse(rawBody);
+    } catch {
+      return { status: 400, headers: { "Content-Type": "application/json" }, body: JSON.stringify({ error: "Invalid JSON" }) };
+    }
+
+    const resolved = this.resolveModelRoute(body.model);
+    if ("error" in resolved) {
+      return { status: resolved.error.status, headers: { "Content-Type": "application/json" }, body: JSON.stringify({ error: { message: resolved.error.message } }) };
+    }
+
+    const { modelId, proxyModel, routeNodeId } = resolved;
     const messages = body.messages;
     if (proxyModel?.description) {
       const first = messages[0] as { role?: string; content?: string } | undefined;
@@ -213,35 +314,66 @@ export class ModelProxy {
 
     const stream = body.stream ?? false;
     const requestId = crypto.randomUUID();
-
     const frame: ModelRequest = {
-      type: "model_req",
-      id: requestId,
-      from: this.config.nodeId,
-      to: route.nodeId,
-      timestamp: Date.now(),
-      payload: {
-        model: modelId,
-        messages,
-        temperature: body.temperature,
-        maxTokens: body.max_tokens,
-        stream,
-      },
+      type: "model_req", id: requestId, from: this.config.nodeId, to: routeNodeId, timestamp: Date.now(),
+      payload: { model: modelId, provider: proxyModel?.provider, api: proxyModel?.api, messages, temperature: body.temperature, maxTokens: body.max_tokens, stream },
     };
 
-    // Pre-check reachability before starting a stream (avoids silent empty response)
-    if (!this.peerManager.canReach(route.nodeId)) {
-      return {
-        status: 502,
-        headers: { "Content-Type": "application/json" },
-        body: JSON.stringify({ error: { message: `Cannot reach model node "${route.nodeId}"` } }),
-      };
+    if (stream) {
+      return this.handleStreamRequest(requestId, routeNodeId, frame, "chat");
+    } else {
+      return this.handleNonStreamRequest(requestId, routeNodeId, frame, "chat");
+    }
+  }
+
+  private async handleResponses(rawBody: string): Promise<{ status: number; headers: Record<string, string>; body: string | ReadableStream }> {
+    let body: { model: string; input: unknown; stream?: boolean; temperature?: number; max_output_tokens?: number; instructions?: string };
+    try {
+      body = JSON.parse(rawBody);
+    } catch {
+      return { status: 400, headers: { "Content-Type": "application/json" }, body: JSON.stringify({ error: "Invalid JSON" }) };
+    }
+
+    const resolved = this.resolveModelRoute(body.model);
+    if ("error" in resolved) {
+      return { status: resolved.error.status, headers: { "Content-Type": "application/json" }, body: JSON.stringify({ error: { message: resolved.error.message } }) };
+    }
+
+    const { modelId, proxyModel, routeNodeId } = resolved;
+
+    // Normalize responses API input → simple chat messages for WS transport.
+    // Responses API items use {type: "message", role, content: [{type: "input_text", text}]}
+    // but WS protocol carries simple {role, content} chat messages.
+    const messages = ModelProxy.normalizeResponsesInput(body.input);
+
+    // Prepend instructions as system/developer message
+    if (body.instructions) {
+      messages.unshift({ role: "developer", content: body.instructions });
+    }
+
+    if (proxyModel?.description) {
+      const first = messages[0] as { role?: string; content?: string } | undefined;
+      if (first?.role === "system" && typeof first.content === "string") {
+        first.content = `[Model: ${proxyModel.description}]\n${first.content}`;
+      } else if (first?.role === "developer" && typeof first.content === "string") {
+        first.content = `[Model: ${proxyModel.description}]\n${first.content}`;
+      } else {
+        messages.unshift({ role: "system", content: `[Model: ${proxyModel.description}]` });
+      }
     }
 
+    const stream = body.stream ?? false;
+    const requestId = crypto.randomUUID();
+    debug("proxy", `responses: stream=${stream} messages=${messages.length} input_type=${typeof body.input}${Array.isArray(body.input) ? `[${body.input.length}]` : ""}`);
+    const frame: ModelRequest = {
+      type: "model_req", id: requestId, from: this.config.nodeId, to: routeNodeId, timestamp: Date.now(),
+      payload: { model: modelId, provider: proxyModel?.provider, api: proxyModel?.api, messages, temperature: body.temperature, maxTokens: body.max_output_tokens, stream },
+    };
+
     if (stream) {
-      return this.handleStreamRequest(requestId, route.nodeId, frame);
+      return this.handleStreamRequest(requestId, routeNodeId, frame, "responses");
     } else {
-      return this.handleNonStreamRequest(requestId, route.nodeId, frame);
+      return this.handleNonStreamRequest(requestId, routeNodeId, frame, "responses");
     }
   }
 
@@ -249,8 +381,10 @@ export class ModelProxy {
     requestId: string,
     targetNodeId: string,
     frame: ModelRequest,
+    responseFormat: ResponseFormat,
   ): { status: number; headers: Record<string, string>; body: ReadableStream } {
     const encoder = new TextEncoder();
+    const model = frame.payload.model;
 
     const readable = new ReadableStream({
       start: (controller) => {
@@ -258,63 +392,89 @@ export class ModelProxy {
           this.pending.delete(requestId);
           this.peerManager.router.markFailed(requestId);
           try {
-            const errorChunk = {
-              id: `chatcmpl-${requestId}`,
-              object: "chat.completion.chunk",
-              choices: [{ index: 0, delta: { content: "\n\n[ClawMatrix] Error: model request timed out" }, finish_reason: "stop" }],
-            };
-            controller.enqueue(
-              encoder.encode(`data: ${JSON.stringify(errorChunk)}\n\n`),
-            );
-            controller.enqueue(encoder.encode("data: [DONE]\n\n"));
+            if (responseFormat === "responses") {
+              controller.enqueue(encoder.encode(`event: response.output_text.delta\ndata: ${JSON.stringify({ type: "response.output_text.delta", delta: "\n\n[ClawMatrix] Error: model request timed out" })}\n\n`));
+              this.enqueueResponsesStreamDone(controller, encoder, requestId, model);
+            } else {
+              controller.enqueue(encoder.encode(`data: ${JSON.stringify({ id: `chatcmpl-${requestId}`, object: "chat.completion.chunk", choices: [{ index: 0, delta: { content: "\n\n[ClawMatrix] Error: model request timed out" }, finish_reason: "stop" }] })}\n\n`));
+              controller.enqueue(encoder.encode("data: [DONE]\n\n"));
+            }
             controller.close();
-          } catch {
-            // controller may already be closed
-          }
+          } catch { /* controller may already be closed */ }
         }, MODEL_TIMEOUT);
 
         this.pending.set(requestId, {
-          resolve: () => {},
-          reject: () => {},
-          timer,
-          stream: true,
-          controller,
-          encoder,
+          resolve: () => {}, reject: () => {},
+          timer, stream: true, responseFormat, model,
+          controller, encoder,
         });
 
+        // Emit setup events for responses API
+        if (responseFormat === "responses") {
+          this.enqueueResponsesStreamSetup(controller, encoder, requestId, model);
+        }
+
         const sent = this.peerManager.sendTo(targetNodeId, frame);
         if (!sent) {
           this.pending.delete(requestId);
           clearTimeout(timer);
-          const errChunk = {
-            id: `chatcmpl-${requestId}`,
-            object: "chat.completion.chunk",
-            choices: [{ index: 0, delta: { content: `[ClawMatrix] Cannot reach model node "${targetNodeId}"` }, finish_reason: "stop" }],
-          };
-          controller.enqueue(
-            encoder.encode(`data: ${JSON.stringify(errChunk)}\n\n`),
-          );
-          controller.enqueue(encoder.encode("data: [DONE]\n\n"));
-          controller.close();
+          try {
+            if (responseFormat === "responses") {
+              controller.enqueue(encoder.encode(`event: response.output_text.delta\ndata: ${JSON.stringify({ type: "response.output_text.delta", delta: `[ClawMatrix] Cannot reach model node "${targetNodeId}"` })}\n\n`));
+              this.enqueueResponsesStreamDone(controller, encoder, requestId, model);
+            } else {
+              controller.enqueue(encoder.encode(`data: ${JSON.stringify({ id: `chatcmpl-${requestId}`, object: "chat.completion.chunk", choices: [{ index: 0, delta: { content: `[ClawMatrix] Cannot reach model node "${targetNodeId}"` }, finish_reason: "stop" }] })}\n\n`));
+              controller.enqueue(encoder.encode("data: [DONE]\n\n"));
+            }
+            controller.close();
+          } catch { /* controller may already be closed */ }
         }
       },
     });
 
     return {
       status: 200,
-      headers: {
-        "Content-Type": "text/event-stream",
-        "Cache-Control": "no-cache",
-        "Connection": "keep-alive",
-      },
+      headers: { "Content-Type": "text/event-stream", "Cache-Control": "no-cache", "Connection": "keep-alive" },
       body: readable,
     };
   }
 
+  /** Emit responses API stream setup events (response.created → content_part.added). */
+  private enqueueResponsesStreamSetup(controller: ReadableStreamDefaultController, encoder: TextEncoder, id: string, model: string) {
+    const respId = `resp_${id}`;
+    const msgId = `msg_${id}`;
+    const now = Math.floor(Date.now() / 1000);
+    const baseResp = { id: respId, object: "response", created_at: now, status: "in_progress", model, output: [] };
+    const msgItem = { type: "message", id: msgId, role: "assistant", content: [], status: "in_progress" };
+    const textPart = { type: "output_text", text: "" };
+
+    controller.enqueue(encoder.encode(`event: response.created\ndata: ${JSON.stringify({ type: "response.created", response: baseResp })}\n\n`));
+    controller.enqueue(encoder.encode(`event: response.in_progress\ndata: ${JSON.stringify({ type: "response.in_progress", response: baseResp })}\n\n`));
+    controller.enqueue(encoder.encode(`event: response.output_item.added\ndata: ${JSON.stringify({ type: "response.output_item.added", output_index: 0, item: msgItem })}\n\n`));
+    controller.enqueue(encoder.encode(`event: response.content_part.added\ndata: ${JSON.stringify({ type: "response.content_part.added", item_id: msgId, output_index: 0, content_index: 0, part: textPart })}\n\n`));
+  }
+
+  /** Emit responses API stream completion events (output_text.done → response.completed). */
+  private enqueueResponsesStreamDone(controller: ReadableStreamDefaultController, encoder: TextEncoder, id: string, model: string, content?: string, usage?: { inputTokens: number; outputTokens: number }) {
+    const respId = `resp_${id}`;
+    const msgId = `msg_${id}`;
+    const now = Math.floor(Date.now() / 1000);
+    const textPart = { type: "output_text", text: content ?? "" };
+    const msgItem = { type: "message", id: msgId, role: "assistant", content: [textPart], status: "completed" };
+    const usageObj = usage ? { input_tokens: usage.inputTokens, output_tokens: usage.outputTokens, total_tokens: usage.inputTokens + usage.outputTokens } : { input_tokens: 0, output_tokens: 0, total_tokens: 0 };
+    const completedResp = { id: respId, object: "response", created_at: now, status: "completed", model, output: [msgItem], usage: usageObj };
+
+    controller.enqueue(encoder.encode(`event: response.output_text.done\ndata: ${JSON.stringify({ type: "response.output_text.done", item_id: msgId, output_index: 0, content_index: 0, text: content ?? "" })}\n\n`));
+    controller.enqueue(encoder.encode(`event: response.content_part.done\ndata: ${JSON.stringify({ type: "response.content_part.done", item_id: msgId, output_index: 0, content_index: 0, part: textPart })}\n\n`));
+    controller.enqueue(encoder.encode(`event: response.output_item.done\ndata: ${JSON.stringify({ type: "response.output_item.done", output_index: 0, item: msgItem })}\n\n`));
+    controller.enqueue(encoder.encode(`event: response.completed\ndata: ${JSON.stringify({ type: "response.completed", response: completedResp })}\n\n`));
+  }
+
   private async handleNonStreamRequest(
     requestId: string,
     targetNodeId: string,
     frame: ModelRequest,
+    responseFormat: ResponseFormat,
   ): Promise<{ status: number; headers: Record<string, string>; body: string }> {
     try {
       const result = await new Promise<ModelResponse["payload"]>(
@@ -327,9 +487,7 @@ export class ModelProxy {
 
           this.pending.set(requestId, {
             resolve: resolve as (v: unknown) => void,
-            reject,
-            timer,
-            stream: false,
+            reject, timer, stream: false, responseFormat,
           });
 
           const sent = this.peerManager.sendTo(targetNodeId, frame);
@@ -349,6 +507,41 @@ export class ModelProxy {
         };
       }
 
+      if (responseFormat === "responses") {
+        const msgId = `msg_${requestId}`;
+        const usageObj = result.usage
+          ? { input_tokens: result.usage.inputTokens, output_tokens: result.usage.outputTokens, total_tokens: result.usage.inputTokens + result.usage.outputTokens }
+          : { input_tokens: 0, output_tokens: 0, total_tokens: 0 };
+        // If upstream sent full output array (responses API), use it directly
+        const output = Array.isArray(result.message)
+          ? result.message
+          : [{
+              type: "message", id: msgId, role: "assistant",
+              content: [{ type: "output_text", text: result.content ?? "" }],
+              status: "completed",
+            }];
+        return {
+          status: 200,
+          headers: { "Content-Type": "application/json" },
+          body: JSON.stringify({
+            id: `resp_${requestId}`,
+            object: "response",
+            created_at: Math.floor(Date.now() / 1000),
+            status: "completed",
+            model: frame.payload.model,
+            output,
+            usage: usageObj,
+          }),
+        };
+      }
+
+      // Chat completions format — use full message object when available (has tool_calls etc.)
+      const msg = result.message as Record<string, unknown> | undefined;
+      const message = msg
+        ? { role: "assistant", ...msg }
+        : { role: "assistant", content: result.content };
+      const finishReason = msg?.tool_calls ? "tool_calls" : "stop";
+
       return {
         status: 200,
         headers: { "Content-Type": "application/json" },
@@ -357,19 +550,13 @@ export class ModelProxy {
           object: "chat.completion",
           created: Math.floor(Date.now() / 1000),
           model: frame.payload.model,
-          choices: [
-            {
-              index: 0,
-              message: { role: "assistant", content: result.content },
-              finish_reason: "stop",
-            },
-          ],
+          choices: [{
+            index: 0,
+            message,
+            finish_reason: finishReason,
+          }],
           usage: result.usage
-            ? {
-                prompt_tokens: result.usage.inputTokens,
-                completion_tokens: result.usage.outputTokens,
-                total_tokens: result.usage.inputTokens + result.usage.outputTokens,
-              }
+            ? { prompt_tokens: result.usage.inputTokens, completion_tokens: result.usage.outputTokens, total_tokens: result.usage.inputTokens + result.usage.outputTokens }
             : undefined,
         }),
       };
@@ -383,16 +570,35 @@ export class ModelProxy {
   }
 
   private handleListModels(): { status: number; headers: Record<string, string>; body: string } {
-    const models = this.peerManager.router
-      .getAllPeers()
-      .flatMap((p) =>
-        p.models.map((m) => ({
-          id: m.id,
-          object: "model",
-          created: 0,
-          owned_by: m.provider,
-        })),
-      );
+    // Build from proxyModels config (has full detail) and enrich with
+    // connectivity info from the router so consumers know what's reachable.
+    const reachable = new Set(
+      this.peerManager.router.getAllPeers()
+        .filter((p) => p.connection?.isOpen || p.reachableVia)
+        .map((p) => p.nodeId),
+    );
+
+    const models = this.config.proxyModels.map((m) => {
+      const entry: Record<string, unknown> = {
+        id: m.id,
+        object: "model",
+        created: 0,
+        owned_by: m.provider ?? "unknown",
+        // Extended fields
+        ...(m.description && { description: m.description }),
+        ...(m.contextWindow && { context_window: m.contextWindow }),
+        ...(m.maxTokens && { max_tokens: m.maxTokens }),
+        ...(m.reasoning !== undefined && { reasoning: m.reasoning }),
+        ...(m.input && { input: m.input }),
+        ...(m.api && { api: m.api }),
+        ...(m.cost && { cost: m.cost }),
+        ...(m.compat && { compat: m.compat }),
+        // Cluster info
+        node_id: m.nodeId,
+        reachable: reachable.has(m.nodeId),
+      };
+      return entry;
+    });
 
     return {
       status: 200,
@@ -414,19 +620,16 @@ export class ModelProxy {
         clearTimeout(pending.timer);
         this.pending.delete(frame.id);
         try {
-          const errChunk = {
-            id: `chatcmpl-${frame.id}`,
-            object: "chat.completion.chunk",
-            choices: [{ index: 0, delta: { content: `[ClawMatrix] Remote error: ${frame.payload.error}` }, finish_reason: "stop" }],
-          };
-          pending.controller.enqueue(
-            pending.encoder.encode(`data: ${JSON.stringify(errChunk)}\n\n`),
-          );
-          pending.controller.enqueue(pending.encoder.encode("data: [DONE]\n\n"));
+          const errMsg = `[ClawMatrix] Remote error: ${frame.payload.error}`;
+          if (pending.responseFormat === "responses") {
+            pending.controller.enqueue(pending.encoder.encode(`event: response.output_text.delta\ndata: ${JSON.stringify({ type: "response.output_text.delta", delta: errMsg })}\n\n`));
+            this.enqueueResponsesStreamDone(pending.controller, pending.encoder, frame.id, pending.model ?? "", errMsg);
+          } else {
+            pending.controller.enqueue(pending.encoder.encode(`data: ${JSON.stringify({ id: `chatcmpl-${frame.id}`, object: "chat.completion.chunk", choices: [{ index: 0, delta: { content: errMsg }, finish_reason: "stop" }] })}\n\n`));
+            pending.controller.enqueue(pending.encoder.encode("data: [DONE]\n\n"));
+          }
           pending.controller.close();
-        } catch {
-          // controller may already be closed
-        }
+        } catch { /* controller may already be closed */ }
       }
       return;
     }
@@ -436,6 +639,9 @@ export class ModelProxy {
     pending.resolve(frame.payload);
   }
 
+  /** Accumulated text per stream request (needed for responses API done events). */
+  private streamText = new Map<string, string>();
+
   handleModelStream(frame: ModelStreamChunk) {
     debug("stream", `id=${frame.id} done=${frame.payload.done} delta=${JSON.stringify(frame.payload.delta?.slice?.(0, 50) ?? frame.payload.delta)} failed=${this.peerManager.router.isFailed(frame.id)} hasPending=${this.pending.has(frame.id)}`);
     if (this.peerManager.router.isFailed(frame.id)) return;
@@ -443,54 +649,87 @@ export class ModelProxy {
     if (!pending?.stream || !pending.controller || !pending.encoder) return;
 
     try {
-      if (frame.payload.done) {
-        const finalChunk: Record<string, unknown> = {
-          id: `chatcmpl-${frame.id}`,
-          object: "chat.completion.chunk",
-          choices: [{ index: 0, delta: {}, finish_reason: "stop" }],
-        };
-        if (frame.payload.usage) {
-          finalChunk.usage = {
-            prompt_tokens: frame.payload.usage.inputTokens,
-            completion_tokens: frame.payload.usage.outputTokens,
-            total_tokens: frame.payload.usage.inputTokens + frame.payload.usage.outputTokens,
-          };
-        }
-        pending.controller.enqueue(
-          pending.encoder.encode(`data: ${JSON.stringify(finalChunk)}\n\n`),
-        );
-        pending.controller.enqueue(pending.encoder.encode("data: [DONE]\n\n"));
-        pending.controller.close();
-        clearTimeout(pending.timer);
-        this.pending.delete(frame.id);
+      if (pending.responseFormat === "responses") {
+        this.handleModelStreamResponses(frame, pending);
       } else {
-        const chunk = {
-          id: `chatcmpl-${frame.id}`,
-          object: "chat.completion.chunk",
-          choices: [
-            {
-              index: 0,
-              delta: { content: frame.payload.delta },
-              finish_reason: null,
-            },
-          ],
-        };
-        pending.controller.enqueue(
-          pending.encoder.encode(`data: ${JSON.stringify(chunk)}\n\n`),
-        );
+        this.handleModelStreamChat(frame, pending);
       }
     } catch {
       clearTimeout(pending.timer);
       this.pending.delete(frame.id);
+      this.streamText.delete(frame.id);
     }
   }
 
+  private handleModelStreamChat(frame: ModelStreamChunk, pending: PendingModelReq) {
+    if (frame.payload.done) {
+      const finalChunk: Record<string, unknown> = {
+        id: `chatcmpl-${frame.id}`, object: "chat.completion.chunk",
+        choices: [{ index: 0, delta: {}, finish_reason: "stop" }],
+      };
+      if (frame.payload.usage) {
+        finalChunk.usage = { prompt_tokens: frame.payload.usage.inputTokens, completion_tokens: frame.payload.usage.outputTokens, total_tokens: frame.payload.usage.inputTokens + frame.payload.usage.outputTokens };
+      }
+      pending.controller!.enqueue(pending.encoder!.encode(`data: ${JSON.stringify(finalChunk)}\n\n`));
+      pending.controller!.enqueue(pending.encoder!.encode("data: [DONE]\n\n"));
+      pending.controller!.close();
+      clearTimeout(pending.timer);
+      this.pending.delete(frame.id);
+    } else {
+      // Use full deltaObj when available (carries tool_calls etc.), otherwise simple text delta
+      const delta = frame.payload.deltaObj ?? { content: frame.payload.delta };
+      const chunk = { id: `chatcmpl-${frame.id}`, object: "chat.completion.chunk", choices: [{ index: 0, delta, finish_reason: null }] };
+      pending.controller!.enqueue(pending.encoder!.encode(`data: ${JSON.stringify(chunk)}\n\n`));
+    }
+  }
+
+  private handleModelStreamResponses(frame: ModelStreamChunk, pending: PendingModelReq) {
+    if (frame.payload.done) {
+      const fullText = this.streamText.get(frame.id) ?? "";
+      this.streamText.delete(frame.id);
+      this.enqueueResponsesStreamDone(pending.controller!, pending.encoder!, frame.id, pending.model ?? "", fullText, frame.payload.usage);
+      pending.controller!.close();
+      clearTimeout(pending.timer);
+      this.pending.delete(frame.id);
+    } else {
+      // Accumulate text for done event
+      this.streamText.set(frame.id, (this.streamText.get(frame.id) ?? "") + frame.payload.delta);
+      const evt = { type: "response.output_text.delta", item_id: `msg_${frame.id}`, output_index: 0, content_index: 0, delta: frame.payload.delta };
+      pending.controller!.enqueue(pending.encoder!.encode(`event: response.output_text.delta\ndata: ${JSON.stringify(evt)}\n\n`));
+    }
+  }
+
+  private sendStreamDelta(to: string, id: string, delta: string, deltaObj?: unknown) {
+    this.peerManager.sendTo(to, {
+      type: "model_stream",
+      id,
+      from: this.config.nodeId,
+      to,
+      timestamp: Date.now(),
+      payload: { delta, ...(deltaObj !== undefined && { deltaObj }), done: false },
+    } satisfies ModelStreamChunk);
+  }
+
+  private sendStreamDone(to: string, id: string, usage?: { inputTokens: number; outputTokens: number }) {
+    this.peerManager.sendTo(to, {
+      type: "model_stream",
+      id,
+      from: this.config.nodeId,
+      to,
+      timestamp: Date.now(),
+      payload: { delta: "", done: true, usage },
+    } satisfies ModelStreamChunk);
+  }
+
   /** Handle model_req locally: call the model API directly or fall back to OpenClaw gateway. */
   async handleModelRequest(frame: ModelRequest): Promise<void> {
     const { id, from, payload } = frame;
-    debug("model_req", `handling model="${payload.model}" from=${from} stream=${payload.stream}`);
+    debug("model_req", `handling model="${payload.model}" provider=${payload.provider ?? "any"} from=${from} stream=${payload.stream}`);
 
-    const model = this.config.models.find((m) => m.id === payload.model);
+    const model = payload.provider
+      ? this.config.models.find((m) => m.id === payload.model && m.provider === payload.provider)
+        ?? this.config.models.find((m) => m.id === payload.model)
+      : this.config.models.find((m) => m.id === payload.model);
     if (!model) {
       this.peerManager.sendTo(from, {
         type: "model_res",
@@ -505,28 +744,42 @@ export class ModelProxy {
 
     try {
       const endpoint = this.resolveModelEndpoint(model);
+      const isResponsesApi = endpoint.api === "openai-responses" || endpoint.api === "openai-codex-responses";
+      const path = isResponsesApi ? "/responses" : "/chat/completions";
+      const url = `${endpoint.baseUrl}${path}`;
       const headers: Record<string, string> = { "Content-Type": "application/json" };
 
       if (endpoint.direct) {
         if (endpoint.apiKey) headers["Authorization"] = `Bearer ${endpoint.apiKey}`;
-        debug("model_req", `direct API call to ${endpoint.url}`);
+        debug("model_req", `direct API call to ${url} (api=${endpoint.api})`);
       } else {
         const { authHeader } = this.gatewayInfo;
         if (authHeader) headers["Authorization"] = authHeader;
-        debug("model_req", `gateway fallback to ${endpoint.url} (not recommended)`);
+        debug("model_req", `gateway fallback to ${url}`);
       }
 
-      const response = await fetch(endpoint.url, {
+      const modelField = endpoint.direct ? model.id : `${model.provider}/${model.id}`;
+      const requestBody = isResponsesApi
+        ? {
+            model: modelField,
+            input: payload.messages,
+            stream: payload.stream,
+            temperature: payload.temperature,
+            max_output_tokens: payload.maxTokens,
+          }
+        : {
+            model: modelField,
+            messages: payload.messages,
+            temperature: payload.temperature,
+            max_tokens: payload.maxTokens,
+            stream: payload.stream,
+            ...(payload.stream ? { stream_options: { include_usage: true } } : {}),
+          };
+
+      const response = await fetch(url, {
         method: "POST",
         headers,
-        body: JSON.stringify({
-          model: endpoint.direct ? model.id : `${model.provider}/${model.id}`,
-          messages: payload.messages,
-          temperature: payload.temperature,
-          max_tokens: payload.maxTokens,
-          stream: payload.stream,
-          ...(payload.stream ? { stream_options: { include_usage: true } } : {}),
-        }),
+        body: JSON.stringify(requestBody),
       });
 
       if (!response.ok) {
@@ -549,73 +802,122 @@ export class ModelProxy {
             if (done) break;
 
             buffer += decoder.decode(value, { stream: true });
+            if (buffer.length > MAX_STREAM_BUFFER) {
+              throw new Error("Stream buffer exceeded 1MB — upstream may be malformed");
+            }
             const lines = buffer.split("\n");
             buffer = lines.pop()!;
 
+            // Track SSE event type for responses API
+            let currentEvent = "";
             for (const line of lines) {
+              if (line.startsWith("event: ")) {
+                currentEvent = line.slice(7).trim();
+                continue;
+              }
               if (!line.startsWith("data: ")) continue;
               const data = line.slice(6).trim();
               if (data === "[DONE]") {
-                this.peerManager.sendTo(from, {
-                  type: "model_stream",
-                  id,
-                  from: this.config.nodeId,
-                  to: from,
-                  timestamp: Date.now(),
-                  payload: { delta: "", done: true, usage: lastUsage },
-                } satisfies ModelStreamChunk);
+                this.sendStreamDone(from, id, lastUsage);
                 streamDone = true;
                 break;
               }
 
               try {
                 const parsed = JSON.parse(data);
-                if (parsed.usage) {
-                  lastUsage = {
-                    inputTokens: parsed.usage.prompt_tokens,
-                    outputTokens: parsed.usage.completion_tokens,
-                  };
-                }
-                const d = parsed.choices?.[0]?.delta;
-                const delta = d?.content || d?.reasoning_content || "";
-                if (delta) {
-                  this.peerManager.sendTo(from, {
-                    type: "model_stream",
-                    id,
-                    from: this.config.nodeId,
-                    to: from,
-                    timestamp: Date.now(),
-                    payload: { delta, done: false },
-                  } satisfies ModelStreamChunk);
+
+                if (isResponsesApi) {
+                  const evtType = currentEvent || parsed.type;
+                  if (evtType === "response.output_text.delta") {
+                    const delta = parsed.delta || "";
+                    if (delta) {
+                      this.sendStreamDelta(from, id, delta);
+                    }
+                  } else if (evtType === "response.completed") {
+                    const usage = parsed.response?.usage;
+                    if (usage) {
+                      lastUsage = {
+                        inputTokens: usage.input_tokens ?? usage.prompt_tokens ?? 0,
+                        outputTokens: usage.output_tokens ?? usage.completion_tokens ?? 0,
+                      };
+                    }
+                    this.sendStreamDone(from, id, lastUsage);
+                    streamDone = true;
+                    break;
+                  }
+                } else {
+                  // Chat completions format
+                  if (parsed.usage) {
+                    lastUsage = {
+                      inputTokens: parsed.usage.prompt_tokens,
+                      outputTokens: parsed.usage.completion_tokens,
+                    };
+                  }
+                  const d = parsed.choices?.[0]?.delta;
+                  const delta = d?.content || d?.reasoning_content || "";
+                  // Pass full delta object when it contains tool_calls or other structured data
+                  const hasStructured = d?.tool_calls || d?.refusal != null;
+                  if (delta || hasStructured) {
+                    this.sendStreamDelta(from, id, delta, hasStructured ? d : undefined);
+                  }
                 }
               } catch {
                 // skip malformed chunks
               }
+              currentEvent = "";
             }
           }
-          // If the upstream closed without sending [DONE], send a completion
-          // frame so the requesting side doesn't hang until MODEL_TIMEOUT.
+          // If the upstream closed without sending [DONE] or response.completed,
+          // send a completion frame so the requesting side doesn't hang.
           if (!streamDone) {
-            this.peerManager.sendTo(from, {
-              type: "model_stream",
-              id,
-              from: this.config.nodeId,
-              to: from,
-              timestamp: Date.now(),
-              payload: { delta: "", done: true, usage: lastUsage },
-            } satisfies ModelStreamChunk);
+            this.sendStreamDone(from, id, lastUsage);
           }
         } finally {
           reader.releaseLock();
         }
       } else {
-        const result = (await response.json()) as {
-          choices?: { message?: { content?: string; reasoning_content?: string } }[];
-          usage?: { prompt_tokens: number; completion_tokens: number };
-        };
-        const msg = result.choices?.[0]?.message;
-        const content = msg?.content || msg?.reasoning_content || "";
-        const usage = result.usage;
+        // Non-streaming response
+        const result = await response.json();
+        let content: string;
+        let message: unknown | undefined;
+        let usage: { inputTokens: number; outputTokens: number } | undefined;
+
+        if (isResponsesApi) {
+          // Responses API: extract text from output[].content[].text
+          content = "";
+          const output = result.output as { type?: string; content?: { type?: string; text?: string }[] }[] | undefined;
+          if (Array.isArray(output)) {
+            for (const item of output) {
+              if (item.type === "message" && Array.isArray(item.content)) {
+                for (const part of item.content) {
+                  if (part.type === "output_text" && part.text) content += part.text;
+                }
+              }
+            }
+          }
+          // Carry full output array for structured data (function_call items, etc.)
+          message = result.output;
+          if (result.usage) {
+            usage = {
+              inputTokens: result.usage.input_tokens ?? result.usage.prompt_tokens ?? 0,
+              outputTokens: result.usage.output_tokens ?? result.usage.completion_tokens ?? 0,
+            };
+          }
+        } else {
+          // Chat completions format
+          const msg = result.choices?.[0]?.message;
+          content = msg?.content || msg?.reasoning_content || "";
+          // Carry full message object when it has tool_calls or other structured data
+          if (msg?.tool_calls || msg?.refusal != null || msg?.function_call) {
+            message = msg;
+          }
+          if (result.usage) {
+            usage = {
+              inputTokens: result.usage.prompt_tokens,
+              outputTokens: result.usage.completion_tokens,
+            };
+          }
+        }
 
         this.peerManager.sendTo(from, {
           type: "model_res",
@@ -626,9 +928,8 @@ export class ModelProxy {
           payload: {
             success: true,
             content,
-            usage: usage
-              ? { inputTokens: usage.prompt_tokens, outputTokens: usage.completion_tokens }
-              : undefined,
+            ...(message !== undefined && { message }),
+            usage,
           },
         } satisfies ModelResponse);
       }