@springbrand/agent-runtime 0.1.0 → 0.1.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@springbrand/agent-runtime",
-  "version": "0.1.0",
+  "version": "0.1.2",
   "type": "module",
   "files": [
     "src",

package/src/kernel/bindings.ts

@@ -759,6 +759,7 @@ export interface RuntimeBrowserPort {
  */
 export type RuntimeModelProtocol =
   | "openai-chat"
+  | "openrouter-chat"
   | "anthropic-messages"
   | "google-generative-ai"
   | "openai-codex-responses";

package/src/kernel/profile.ts

@@ -84,7 +84,6 @@ export interface RuntimeProfile {
   denyPolicy?: RuntimeDenyPolicy;
   enabledSubagents: readonly string[];
   mcpServers: readonly RuntimeMcpServer[];
-  enabledExtensions: readonly string[];
   executionLevel: ExecutionLevel;
   memory: RuntimeMemoryProfile;
 }
@@ -109,7 +108,6 @@ export function freezeRuntimeProfile(profile: RuntimeProfile): RuntimeProfile {
       : undefined,
     enabledSubagents: Object.freeze([...profile.enabledSubagents]),
     mcpServers: Object.freeze(profile.mcpServers.map((server) => Object.freeze({ ...server }))),
-    enabledExtensions: Object.freeze([...profile.enabledExtensions]),
     memory: Object.freeze({ ...profile.memory }),
   };
   return Object.freeze(frozen);

package/src/pi/runtime-adapter/assembly.ts

@@ -45,7 +45,7 @@ interface PiAssemblySnapshot {
   >;
   readonly profile: Pick<
     RuntimeProfile,
-    "denyPolicy" | "enabledExtensions" | "mcpServers"
+    "denyPolicy" | "mcpServers"
   >;
   readonly bindings: {
     readonly platform: Pick<RuntimePlatformPort, "loader">;
@@ -453,7 +453,7 @@ async function preparePiAssembly(
   const extensions = await assemblePiExtensions({
     extensions: snapshot.pi.extensions,
     published: extensionNames,
-    enabled: new Set(snapshot.profile.enabledExtensions),
+    enabled: extensionNames,
     authorized: extensionNames,
     load: async (extension) => {
       const workspacePermission =

package/src/pi/runtime-adapter/models.ts

@@ -2,7 +2,9 @@ import type { StreamFn } from "@earendil-works/pi-agent-core";
 import {
   createModels,
   createProvider,
+  lazyStream,
   type Api,
+  type AssistantMessageEvent,
   type Model,
   type MutableModels,
 } from "@earendil-works/pi-ai";
@@ -24,6 +26,9 @@ import {
 import {
   openaiProvider,
 } from "@earendil-works/pi-ai/providers/openai";
+import {
+  openrouterProvider,
+} from "@earendil-works/pi-ai/providers/openrouter";
 import {
   googleProvider,
 } from "@earendil-works/pi-ai/providers/google";
@@ -38,25 +43,64 @@ import type {
 
 const CATALOGS = {
   "openai-chat": openaiProvider().getModels(),
+  "openrouter-chat": openrouterProvider().getModels(),
   "anthropic-messages": anthropicProvider().getModels(),
   "google-generative-ai": googleProvider().getModels(),
   "openai-codex-responses": openaiCodexProvider().getModels(),
 } satisfies Record<RuntimeModelProtocol, readonly Model<Api>[]>;
 
 const PROVIDER_MAX_RETRIES = 2;
+const MODEL_STREAM_STALL_TIMEOUT_MS = 120_000;
+
+async function* stopStalledModelStream(
+  source: AsyncIterable<AssistantMessageEvent>,
+  watchdog: AbortController,
+): AsyncGenerator<AssistantMessageEvent> {
+  const iterator = source[Symbol.asyncIterator]();
+  while (true) {
+    let timer: ReturnType<typeof setTimeout> | undefined;
+    const next = await Promise.race([
+      iterator.next(),
+      new Promise<never>((_, reject) => {
+        timer = setTimeout(() => {
+          const error = new Error(
+            `Model stream stalled for ${MODEL_STREAM_STALL_TIMEOUT_MS} ms`,
+          );
+          watchdog.abort(error);
+          reject(error);
+        }, MODEL_STREAM_STALL_TIMEOUT_MS);
+      }),
+    ]).finally(() => {
+      if (timer !== undefined) clearTimeout(timer);
+    });
+    if (next.done) {
+      throw new Error("Model stream ended without a terminal event");
+    }
+    yield next.value;
+    if (next.value.type === "done" || next.value.type === "error") return;
+  }
+}
 
 /**
- * 给模型请求补上可中断的 provider 瞬时错误重试。
+ * 给模型请求补上可中断的 provider 瞬时错误重试和空闲终止边界。
  *
  * Runtime Turn 和 SubAgent 在把 `Models.streamSimple` 交给 Pi 前调用；显式传入的重试次数优先。
  *
- * Pi 0.83 默认 `maxRetries` 为 0，这里只补 2 次默认值，避免瞬时网络错误直接终止，同时不覆盖调用方策略。
+ * Pi 0.83 默认 `maxRetries` 为 0，且流无事件时会无限等待。这里补 2 次默认重试，
+ * 并在连续 120 秒没有模型事件时中止 provider、返回明确失败终态。
  */
 export function withProviderRetry(streamFn: StreamFn): StreamFn {
   return (model, context, options) =>
-    streamFn(model, context, {
-      ...options,
-      maxRetries: options?.maxRetries ?? PROVIDER_MAX_RETRIES,
+    lazyStream(model, async () => {
+      const watchdog = new AbortController();
+      const source = await streamFn(model, context, {
+        ...options,
+        signal: options?.signal
+          ? AbortSignal.any([options.signal, watchdog.signal])
+          : watchdog.signal,
+        maxRetries: options?.maxRetries ?? PROVIDER_MAX_RETRIES,
+      });
+      return stopStalledModelStream(source, watchdog);
     });
 }
 
@@ -86,6 +130,7 @@ function providerId(endpoint: RuntimeModelEndpoint, index: number): string {
 function apiFor(protocol: RuntimeModelProtocol): Api {
   switch (protocol) {
     case "openai-chat":
+    case "openrouter-chat":
       return "openai-completions";
     case "anthropic-messages":
       return "anthropic-messages";
@@ -99,6 +144,7 @@ function apiFor(protocol: RuntimeModelProtocol): Api {
 function piApiFor(protocol: RuntimeModelProtocol) {
   switch (protocol) {
     case "openai-chat":
+    case "openrouter-chat":
       return openAICompletionsApi();
     case "anthropic-messages":
       return anthropicMessagesApi();
@@ -125,12 +171,11 @@ function configuredModel(
       `Model is not in Pi's built-in catalog for ${endpoint.protocol}: ${modelId}`,
     );
   }
-  // 待确认：当前统一丢弃 catalogModel.compat；现有提交历史没有说明这样做对 Anthropic 兼容参数的影响，修改前需先验证 Pi 的流式请求行为。
   const {
     api: _api,
     provider: _provider,
     baseUrl: _baseUrl,
-    compat: _compat,
+    compat,
     headers: catalogHeaders,
     ...metadata
   } = catalogModel;
@@ -143,6 +188,9 @@ function configuredModel(
     api: apiFor(endpoint.protocol),
     provider: providerId(endpoint, index),
     baseUrl: endpoint.baseURL,
+    ...(endpoint.protocol === "openrouter-chat" && compat
+      ? { compat }
+      : {}),
     ...(Object.keys(headers).length > 0 ? { headers } : {}),
   };
 }
@@ -188,10 +236,10 @@ export function resolvePiApiKey(
  * 用当前部署端点完整重建一个已有的 Pi 模型集合。
  *
  * PiRuntimeAdapter 激活新运行快照时调用它；需要复用 MutableModels 的调用方也可以调用，并继续使用返回的同一实例。
- * 该函数会先清空集合，所以调用方必须把传入对象视为由 Runtime 独占管理，不能期待自定义 provider 被保留。
+ * 该函数在全部新 provider 校验通过后才清空并替换集合，所以失败时旧集合保持可用。
  *
  * Pi 当前用 provider.id 在 Map 中注册和路由请求，因此每个部署端点都必须同时注册唯一 provider、对应模型、认证解析器和协议实现。
- * 先 clearProviders 是为了避免配置更新后旧端点继续可用；所有模型在注册时都会经过 catalog 校验，因此任一未知 modelId 会让整次配置失败。
+ * 替换时先构建新 provider，确保任一未知 modelId 都不会让旧端点失效；校验通过后再清空旧集合，避免配置更新后旧端点继续可用。
  *
  * @returns 传入并已重建完成的 `models` 实例。
  * @throws 当任一部署模型不在 Pi 内置 OpenAI、Anthropic catalog 中时抛出错误。
@@ -200,9 +248,8 @@ export function configurePiModels(
   models: MutableModels,
   provider: RuntimeProviderPort,
 ): MutableModels {
-  models.clearProviders();
-  provider.endpoints.forEach((endpoint, index) => {
-    models.setProvider(createProvider({
+  const configuredProviders = provider.endpoints.map((endpoint, index) =>
+    createProvider({
       id: providerId(endpoint, index),
       name: endpoint.protocol,
       baseUrl: endpoint.baseURL,
@@ -219,8 +266,12 @@ export function configurePiModels(
         configuredModel(endpoint, index, modelId),
       ),
       api: piApiFor(endpoint.protocol),
-    }));
-  });
+    })
+  );
+  models.clearProviders();
+  configuredProviders.forEach((configuredProvider) =>
+    models.setProvider(configuredProvider)
+  );
   return models;
 }
 

package/src/pi/tool/base.ts

@@ -66,9 +66,9 @@ function candidate<T extends TSchema>(tool: AgentTool<T>): PiToolCandidate {
   };
 }
 
-/** Runtime-permanent Pi tools, independent of optional Host plugins. */
+/** Runtime base tools, including native web search when the model supports it. */
 export function basePiToolCandidates(
-  webSearch: WebSearch,
+  webSearch?: WebSearch,
 ): PiToolCandidate[] {
   return [
     candidate({
@@ -105,6 +105,6 @@ export function basePiToolCandidates(
         });
       },
     }),
-    webSearchPiToolCandidate(webSearch),
+    ...(webSearch ? [webSearchPiToolCandidate(webSearch)] : []),
   ];
 }

package/src/pi/tool/web-search/api.ts

@@ -1106,6 +1106,8 @@ export interface WebSearchOptions {
 
 function apiFor(endpoint: RuntimeModelEndpoint): NativeApi {
     switch (endpoint.protocol) {
+        case "openrouter-chat":
+            throw new Error("Native web search is unavailable for openrouter-chat");
         case "google-generative-ai":
             return "google-generative-ai";
         case "openai-chat":

package/src/plugins.ts

@@ -837,7 +837,6 @@ class RuntimeBuilder implements RuntimeContributionContext {
       denyPolicy: this.denyPolicy,
       enabledSubagents: [...this.enabledSubagents],
       mcpServers: [...this.connectors.values()],
-      enabledExtensions: [...this.extensions.keys()],
       executionLevel: this.profile.executionLevel,
       memory: this.memoryProfile,
     });
@@ -864,12 +863,22 @@ class RuntimeBuilder implements RuntimeContributionContext {
       );
     }
     const resolvedModel = resolvePiModel(provider, profile.model);
-    const webSearch = createWebSearch({
-      endpoint: endpoints[0]!,
-      model: profile.model,
-      maxTokens: resolvedModel.maxTokens,
-      reasoning: resolvedModel.reasoning,
-    });
+    const endpoint = endpoints[0]!;
+    const webSearch = endpoint.protocol === "openrouter-chat"
+      ? undefined
+      : createWebSearch({
+        endpoint,
+        model: profile.model,
+        maxTokens: resolvedModel.maxTokens,
+        reasoning: resolvedModel.reasoning,
+      });
+    if (!webSearch) {
+      this.reportDegradation({
+        capability: "web_search",
+        reason: "unavailable",
+        detail: "Native web search is unavailable for openrouter-chat",
+      });
+    }
     for (const candidate of basePiToolCandidates(webSearch)) {
       this.addPiTool(candidate);
     }