@springbrand/agent-runtime 0.2.0-alpha.44 → 0.2.0-alpha.46

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

@@ -15,13 +15,17 @@ import type {
   ToolResultMessage,
   UserMessage,
 } from "@earendil-works/pi-ai";
+import { isContextOverflow } from "@earendil-works/pi-ai";
 import { transformMessages } from "@earendil-works/pi-ai/api/transform-messages";
 import {
   parkPiToolApproval,
   parkPiToolInteraction,
   requiresPiToolApproval,
 } from "../turn";
-import { PiChunkEncoder } from "../message";
+import {
+  classifyPiTerminalStopReason,
+  PiChunkEncoder,
+} from "../message";
 import type { UIMessageChunk } from "ai";
 import { ChatStreamStalledError } from "agents/chat";
 import {
@@ -42,7 +46,6 @@ import {
 import {
   isRecoverableAssistantError,
   isModelStreamStallMessage,
-  resolvePiApiKey,
   withProviderRetry,
   type PiGenerationLifecycleObserver,
 } from "./models";
@@ -79,6 +82,8 @@ const MAX_WRAP_UP_TURNS = 2;
 
 export class RetryableModelError extends Error {}
 
+export class ContextOverflowError extends RetryableModelError {}
+
 function normalizePlanToolCalls(message: AgentMessage): void {
   if (message.role !== "assistant") return;
   for (const part of message.content) {
@@ -88,42 +93,6 @@ function normalizePlanToolCalls(message: AgentMessage): void {
   }
 }
 
-/**
- * 把 Pi 的终止原因翻译成本仓的回合结局。
- *
- * 调用：终态提交(`onTerminal`)与流式记录的 `turnStatus` 共用这一处,
- * 保证服务端只有一份判据 —— 两处各写一遍三元链正是它们悄悄分叉的原因。
- *
- * 为什么未知取值判失败而不是成功:`stopReason` 是上游会扩的联合体
- * (0.83 就加了 `"pending"`)。把认不出的结局宣称成「成功」，等于让一次
- * 异常结束伪装成正常完成 —— 这是本仓明确禁止的假完成，
- * 也是 Pi 自己在 0.83 的选择(认不出的终止原因直接抛错而非压平成 stop)。
- * 前端对未知取值取 `"running"`(未知即未定)，最终由服务端这份权威判据校正。
- */
-function classifyPiStopReason(stopReason: string | undefined): {
-  outcome: PiTurnTerminalIntent["outcome"];
-  turnStatus: "completed" | "error" | "aborted";
-  message?: string;
-} {
-  switch (stopReason) {
-    case "stop":
-    case "length":
-      return { outcome: "succeeded", turnStatus: "completed" };
-    case "aborted":
-      return { outcome: "aborted", turnStatus: "aborted" };
-    case "error":
-      return { outcome: "failed", turnStatus: "error" };
-    default:
-      return {
-        outcome: "failed",
-        turnStatus: "error",
-        message: `SpringBrand ended the turn with an unrecognized stop reason: ${
-          stopReason ?? "(none)"
-        }`,
-      };
-  }
-}
-
 function visibleAssistantContent(message: AssistantMessage) {
   return message.content.filter((part) =>
     part.type === "text"
@@ -140,7 +109,6 @@ interface PiTurnAdapterOptions {
     readonly thinkingLevel?: ThinkingLevel;
   };
   readonly models: Models;
-  readonly apiKey: string;
   readonly modelSessionId: string;
   // 读取这次执行真正要交给 Pi 的 canonical transcript。
   // PiTurnAdapter.run 在创建 PiCore 前调用它，宿主应返回当时最新的 transcript。
@@ -331,7 +299,6 @@ class PiTurnAdapter {
         this.opts.modelSessionId,
         this.opts.onGeneration,
       ),
-      getApiKey: () => this.opts.apiKey,
       // transformMessages 在宿主上下文变换之后运行，确保跨 provider 的 tool call ID 格式兼容。
       // OpenAI Responses API 生成含 `|` 的 450+ 字符 ID，Anthropic 只接受 ^[a-zA-Z0-9_-]+$(64 字符上限)；
       // 切换 provider 或消息跨 provider 回放时若不规范化，provider 会静默拒绝。
@@ -564,6 +531,21 @@ export interface CreatePreparedPiTurnOptions {
     error?: unknown;
     occurredAt: number;
   }>) => void;
+  /** Reports model-requested Tools absent from the pinned Runtime surface. */
+  readonly onUnknownToolStarted?: (input: Readonly<{
+    toolCallId: string;
+    toolName: string;
+    input: unknown;
+    occurredAt: number;
+  }>) => void;
+  /** Reports the immediate Pi failure for a Tool absent from the pinned Runtime surface. */
+  readonly onUnknownToolFinished?: (input: Readonly<{
+    toolCallId: string;
+    toolName: string;
+    outcome: "failed";
+    output: import("@earendil-works/pi-agent-core").AgentToolResult<unknown>;
+    occurredAt: number;
+  }>) => void;
   /** Per-Submission executors for tools whose metadata is fixed at assembly time. */
   readonly toolExecutors?: Readonly<Record<
     string,
@@ -659,6 +641,7 @@ export class PreparedPiTurnAdapter {
   private readonly encoder: PiChunkEncoder;
   private readonly steerMessageIds: string[] = [];
   private terminalIntent?: PiTurnTerminalIntent;
+  private readonly contextWindow: number;
 
   /**
    * 把 Prepared Runtime 和 Submission 持久化端口绑定成一个可执行 Turn。
@@ -679,6 +662,7 @@ export class PreparedPiTurnAdapter {
       dependencies.owner,
       options.pinnedDescriptor,
     );
+    this.contextWindow = state.snapshot.pi.model.contextWindow;
     this.assistantOrdinal = options.submission.assistantOrdinal;
     this.encoder = new PiChunkEncoder({
       messageId: options.submission.messageId,
@@ -903,10 +887,6 @@ export class PreparedPiTurnAdapter {
         thinkingLevel: state.snapshot.pi.thinkingLevel,
       },
       models: dependencies.models,
-      apiKey: resolvePiApiKey(
-        state.snapshot.bindings.provider,
-        state.snapshot.pi.model.id,
-      ),
       modelSessionId: options.modelSessionId,
       ...(options.onGeneration ? { onGeneration: options.onGeneration } : {}),
       canonicalMessages: options.canonicalMessages,
@@ -1038,6 +1018,28 @@ export class PreparedPiTurnAdapter {
   // PiTurnAdapter 会对每个 AgentEvent 调用它，PiCore 会等待该 Promise 后再越过订阅事件屏障。
   // canonical message 先于流投影提交，恢复才不会依赖仅供浏览器消费的记录；调整顺序必须复核 recovery effect。
   private async handleEvent(event: AgentEvent): Promise<void> {
+    if (
+      event.type === "tool_execution_start" &&
+      !this.candidates.some(({ tool }) => tool.name === event.toolName)
+    ) {
+      this.options.onUnknownToolStarted?.({
+        toolCallId: event.toolCallId,
+        toolName: event.toolName,
+        input: event.args,
+        occurredAt: Date.now(),
+      });
+    } else if (
+      event.type === "tool_execution_end" &&
+      !this.candidates.some(({ tool }) => tool.name === event.toolName)
+    ) {
+      this.options.onUnknownToolFinished?.({
+        toolCallId: event.toolCallId,
+        toolName: event.toolName,
+        outcome: "failed",
+        output: event.result,
+        occurredAt: Date.now(),
+      });
+    }
     if (
       event.type === "message_start" ||
       event.type === "message_update" ||
@@ -1048,7 +1050,10 @@ export class PreparedPiTurnAdapter {
     if (
       event.type === "message_update" &&
       event.assistantMessageEvent.type === "error" &&
-      (isModelStreamStallMessage(
+      (isContextOverflow(
+        event.assistantMessageEvent.error,
+        this.contextWindow,
+      ) || isModelStreamStallMessage(
         event.assistantMessageEvent.error.errorMessage,
       ) ||
         isRecoverableAssistantError(event.assistantMessageEvent.error))
@@ -1059,6 +1064,11 @@ export class PreparedPiTurnAdapter {
     if (event.type === "message_end") {
       const message = event.message;
       if (message.role === "assistant") {
+        if (isContextOverflow(message, this.contextWindow)) {
+          throw new ContextOverflowError(
+            message.errorMessage ?? "Model context window overflow",
+          );
+        }
         const stalled = message.stopReason === "error" &&
           isModelStreamStallMessage(message.errorMessage);
         const retryable = isRecoverableAssistantError(message);
@@ -1119,7 +1129,7 @@ export class PreparedPiTurnAdapter {
           authoritativeMessage.role === "assistant" &&
           authoritativeMessage.stopReason !== "toolUse"
         ) {
-          const terminal = classifyPiStopReason(
+          const terminal = classifyPiTerminalStopReason(
             authoritativeMessage.stopReason,
           );
           this.terminalIntent = {

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

@@ -2,13 +2,13 @@ import {
   PreparedPiTurnAdapter,
   type CreatePreparedPiTurnOptions,
 } from "./execution";
-export { RetryableModelError } from "./execution";
+export { ContextOverflowError, RetryableModelError } from "./execution";
 import {
   uiUserMessageToPi,
 } from "../message";
 import type { UIMessage } from "ai";
 import { createModels, type MutableModels } from "@earendil-works/pi-ai";
-import { configurePiModels, createPiModels, resolvePiApiKey } from "./models";
+import { configurePiModels, createPiModels } from "./models";
 export {
   MODEL_STREAM_STALL_TIMEOUT_MS,
   modelRequestUrl,
@@ -120,12 +120,6 @@ export class PiRuntimeAdapter {
     return uiUserMessageToPi(input, privateContext);
   }
 
-  resolveApiKey(
-    ...args: Parameters<typeof resolvePiApiKey>
-  ): ReturnType<typeof resolvePiApiKey> {
-    return resolvePiApiKey(...args);
-  }
-
   /**
    * 加载并预检一份可供后续准入使用的 Prepared Runtime。
    *

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

@@ -111,6 +111,38 @@ export type PiGenerationLifecycleObserver = (
   event: PiGenerationLifecycleEvent,
 ) => void;
 
+export function structuredOutputStream(
+  streamFn: StreamFn,
+  model: Model<Api>,
+  schema: Record<string, unknown>,
+): StreamFn | undefined {
+  if (
+    model.api !== "anthropic-messages" ||
+    !model.provider.startsWith("model-api-openrouter-messages-")
+  ) return undefined;
+
+  return (activeModel, context, options) =>
+    streamFn(activeModel, context, {
+      ...options,
+      onPayload: async (payload, payloadModel) => {
+        const value = await options?.onPayload?.(payload, payloadModel) ?? payload;
+        if (!value || typeof value !== "object" || Array.isArray(value)) {
+          throw new Error("Structured model payload must be an object");
+        }
+        const current = (value as Record<string, unknown>).output_config;
+        return {
+          ...value,
+          output_config: {
+            ...(current && typeof current === "object" && !Array.isArray(current)
+              ? current
+              : {}),
+            format: { type: "json_schema", schema },
+          },
+        };
+      },
+    });
+}
+
 export function isModelStreamStallMessage(message?: string): message is string {
   return message === MODEL_STREAM_STALL_MESSAGE ||
     message?.startsWith(MODEL_STREAM_STALL_DETAILS_PREFIX) === true;
@@ -534,8 +566,8 @@ export function withProviderRetry(
 }
 
 // 找出部署把这个模型放在哪个端点，并保留该端点在配置中的位置。
-// resolvePiModel 和 resolvePiApiKey 在组装运行快照、发起模型请求前调用它，调用方只需传部署配置和 modelId。
-// 这里复用同一条查找路径，避免模型与密钥落到不同端点；同一 modelId 若配置多次会取第一个，是否应改为拒绝重复配置仍待确认。
+// resolvePiModel 在组装运行快照前调用它，调用方只需传部署配置和 modelId。
+// 同一 modelId 若配置多次会取第一个，是否应改为拒绝重复配置仍待确认。
 function endpointFor(
   provider: RuntimeProviderPort,
   modelId: string,
@@ -585,6 +617,20 @@ function piApiFor(protocol: RuntimeModelProtocol) {
   }
 }
 
+function anthropicMessagesModelForOpenRouter(
+  modelId: string,
+): Model<"anthropic-messages"> | undefined {
+  if (!modelId.startsWith("anthropic/")) return;
+  const anthropicModelId = modelId
+    .slice("anthropic/".length)
+    .replace(/:batch$/, "")
+    .replace(/-fast$/, "")
+    .replaceAll(".", "-");
+  return CATALOGS["anthropic-messages"].find(
+    (model) => model.id === anthropicModelId,
+  );
+}
+
 // 把部署声明的模型改造成可由对应自定义端点执行的 Pi 模型。
 // resolvePiModel 为单个运行快照调用它，configurePiModels 则在注册端点时为端点中的每个 modelId 调用它。
 // Pi 的内置 OpenAI、Anthropic provider 提供上下文窗口、费用和推理能力等元数据；这里保留这些元数据，但以部署协议、provider 标识和 baseURL 覆盖内置路由，未知模型会在初始化阶段直接报错，而不是拖到请求阶段。
@@ -602,10 +648,10 @@ function configuredModel(
     );
   }
   const {
-    api: _api,
+    api: catalogApi,
     provider: _provider,
     baseUrl: _baseUrl,
-    compat,
+    compat: catalogCompat,
     headers: catalogHeaders,
     ...metadata
   } = catalogModel;
@@ -614,17 +660,24 @@ function configuredModel(
     ...endpoint.headers,
   };
   const openRouterProviderPin = endpoint.openRouterProviderPins?.[modelId];
-  const supportsStrictTools = modelId.startsWith("anthropic/claude-");
+  const configuredApi = apiFor(endpoint.protocol);
+  // OpenRouter uses dotted versions and optional transport suffixes while the
+  // Anthropic catalog uses dashed base model ids.
+  const anthropicMessagesCompat = endpoint.protocol === "openrouter-messages"
+    ? anthropicMessagesModelForOpenRouter(modelId)?.compat
+    : undefined;
   return {
     ...metadata,
-    api: apiFor(endpoint.protocol),
+    api: configuredApi,
     provider: providerId(endpoint, index),
     baseUrl: endpoint.baseURL,
     ...(endpoint.protocol === "openrouter-messages"
       ? {
           compat: {
+            ...anthropicMessagesCompat,
             supportsEagerToolInputStreaming: false,
-            supportsStrictTools,
+            supportsStrictTools:
+              anthropicMessagesCompat?.supportsStrictTools ?? false,
             supportsToolReferences: true,
             openRouterRouting: {
               require_parameters: true,
@@ -637,7 +690,9 @@ function configuredModel(
             },
           },
         }
-      : {}),
+      : catalogApi === configuredApi && catalogCompat
+        ? { compat: catalogCompat }
+        : {}),
     ...(Object.keys(headers).length > 0 ? { headers } : {}),
   };
 }
@@ -661,24 +716,6 @@ export function resolvePiModel(
   return configuredModel(endpoint, index, modelId);
 }
 
-/**
- * 取出负责指定模型的部署端点密钥。
- *
- * Runtime 在创建一次模型执行或压缩上下文前调用它；辅助的标题生成器也在请求 Pi 时使用它。
- * 调用方应传入与 {@link resolvePiModel} 相同的 provider 和 modelId，并只把返回值用于当前请求。
- *
- * 它与模型解析共用 endpointFor，确保密钥和模型总是选择同一个端点，也让未配置的模型在密钥离开部署边界前失败。
- * 不要另写按协议选密钥的分支，否则同协议的多个端点会拿错凭据。
- *
- * @throws 当部署未配置该 modelId 时抛出错误。
- */
-export function resolvePiApiKey(
-  provider: RuntimeProviderPort,
-  modelId: string,
-): string {
-  return endpointFor(provider, modelId).endpoint.apiKey;
-}
-
 /**
  * 用当前部署端点完整重建一个已有的 Pi 模型集合。
  *

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

@@ -440,7 +440,7 @@ export class PiRuntimeTranscript {
   /**
    * 在 Pi 请求模型前按 Runtime 阈值压缩当前上下文，并把压缩 entry 写回 Session。
    *
-   * `AgentRuntimeKernel.transformPiContext` 作为 Pi 的 `transformContext` 调用它，并提供本次固定的模型、密钥和取消信号。
+   * `AgentRuntimeKernel.transformPiContext` 作为 Pi 的 `transformContext` 调用它，并提供本次固定的模型和取消信号。
    *
    * 方法先逐条核对调用参数与耐久分支；不一致或压缩不可用时保留原消息，避免把漂移的内存上下文提交成新的规范分支。
    */
@@ -448,8 +448,8 @@ export class PiRuntimeTranscript {
     messages: AgentMessage[],
     options: {
       compactAfterTokens: number;
+      force?: boolean;
       model: Model<Api>;
-      apiKey: string;
       signal?: AbortSignal;
       submissionId?: string;
       onCompactionPersisted?: (event: RuntimeModelUsageEvent) => void;
@@ -474,11 +474,14 @@ export class PiRuntimeTranscript {
     const result = await compactPiContext({
       branch,
       compactAfterTokens: options.compactAfterTokens,
+      force: options.force,
       models: this.models,
       model: options.model,
-      apiKey: options.apiKey,
       signal: options.signal,
     });
+    if (options.force && !result.didCompact) {
+      throw new Error("Forced context compaction did not reduce the model input");
+    }
     if (result.compactionEntry) {
       // 整对象写入:assembly 侧已经造好了完整的 compaction 条目,这里不再拆成位置
       // 参数。旧写法拆 5 个参数喂 pi 的 Session.appendCompaction,pi 0.81 把它加到

package/src/pi/tool/base.ts

@@ -2,7 +2,12 @@ import type {
   AgentTool,
   AgentToolResult,
 } from "@earendil-works/pi-agent-core";
-import { Type, type Static, type TSchema } from "@earendil-works/pi-ai";
+import {
+  StringEnum,
+  Type,
+  type Static,
+  type TSchema,
+} from "@earendil-works/pi-ai";
 import { estimateStringTokens } from "agents/experimental/memory/utils";
 import type { RuntimeMemoryPort } from "../../kernel/bindings";
 import type { RuntimeMemoryProfile } from "../../kernel/profile";
@@ -145,11 +150,7 @@ const updatePlanParameters = Type.Object({
     text: Type.String({
       description: "The step, phrased as a short imperative.",
     }),
-    status: Type.Union([
-      Type.Literal("pending"),
-      Type.Literal("in_progress"),
-      Type.Literal("done"),
-    ], {
+    status: StringEnum(["pending", "in_progress", "done"] as const, {
       description:
         'Current status. Use "done" for a finished step; never use "completed".',
     }),
@@ -214,18 +215,12 @@ export function normalizeUpdatePlanArguments(
 }
 
 const setContextParameters = Type.Object({
-  label: Type.Union([
-    Type.Literal("memory"),
-    Type.Literal("preferences"),
-  ], {
+  label: StringEnum(["memory", "preferences"] as const, {
     description:
       "Context block to update: memory for durable facts and active context, preferences for tone and workflow choices.",
   }),
   content: Type.String({ description: "Text to store in the selected context block." }),
-  action: Type.Optional(Type.Union([
-    Type.Literal("replace"),
-    Type.Literal("append"),
-  ], {
+  action: Type.Optional(StringEnum(["replace", "append"] as const, {
     description: 'Whether to replace the block or append to it. Defaults to "replace".',
   })),
 }, { additionalProperties: false });
@@ -318,7 +313,7 @@ export function basePiToolCandidates(
         name: "update_plan",
         label: "Update plan",
         description:
-          "Maintain the user-visible plan for the current task. Call this whenever a task involves 2 or more steps, and again every time the plan or a step's status changes. Always pass the FULL plan — it replaces the previous plan entirely (idempotent overwrite), so omitted steps disappear.",
+          "Maintain the user-visible plan for the current task. Required input: steps. Call this whenever a task involves 2 or more steps, and again every time the plan or a step's status changes. Always pass the FULL plan — it replaces the previous plan entirely (idempotent overwrite), so omitted steps disappear.",
         parameters: updatePlanParameters,
         prepareArguments: normalizeUpdatePlanArguments,
         async execute(_toolCallId, input) {

package/src/pi/tool/compiler.ts

@@ -66,7 +66,7 @@ export interface PiToolCandidate {
    * 那是失效不是降级。`deny` 仍然优先，被 deny 的工具根本不会注入。
    */
   readonly alwaysRequiresApproval?: boolean;
-  /** Recovery may replay this Tool Call only when the candidate guarantees convergence. */
+  /** Recovery may replay this Tool Call only when the candidate guarantees convergence. Omitted means non-idempotent. */
   readonly retry?: "idempotent" | "non-idempotent";
   readonly summary?: string;
   readonly source?: ApprovalReceipt["source"];

package/src/pi/tool/core.ts

@@ -196,6 +196,7 @@ export function listExtensionsPiToolCandidate(
   return {
     owner: "core:extensions",
     requiredExecutionLevel: "low",
+    retry: "idempotent",
     tool,
   };
 }
@@ -278,6 +279,7 @@ export function codeExecutionPiToolCandidate(
   return {
     owner: "core:codemode",
     requiredExecutionLevel: "safe",
+    retry: "idempotent",
     outputBudget: { kind: "structure" },
     source: "codemode",
     summary: "Run JavaScript with network and configured connector access",

package/src/pi/tool/schedule.ts

@@ -92,7 +92,7 @@ function candidate<T extends TSchema>(
   options: Partial<
     Pick<
       PiToolCandidate,
-      "alwaysRequiresApproval" | "exposureMode" | "owner" | "requiredExecutionLevel" | "summary"
+      "alwaysRequiresApproval" | "exposureMode" | "owner" | "requiredExecutionLevel" | "retry" | "summary"
     >
   > = {},
 ): PiToolCandidate {
@@ -106,6 +106,7 @@ function candidate<T extends TSchema>(
       ? { alwaysRequiresApproval: true }
       : {}),
     ...(options.summary ? { summary: options.summary } : {}),
+    ...(options.retry ? { retry: options.retry } : {}),
   };
 }
 
@@ -138,17 +139,20 @@ export function schedulePiToolCandidates(
         summary: "Create a scheduled task",
       },
     ),
-    candidate({
-      name: "list_schedules",
-      label: "List schedules",
-      description:
-        "List the user's existing scheduled tasks. Call this when the user asks what they have scheduled, or before cancelling one — you need the id, and the user will refer to a task by what it does, not by its id.",
-      parameters: noParameters,
-      async execute(_toolCallId, _input, signal) {
-        signal?.throwIfAborted();
-        return result({ schedules: await schedule.list() });
+    candidate(
+      {
+        name: "list_schedules",
+        label: "List schedules",
+        description:
+          "List the user's existing scheduled tasks. Call this when the user asks what they have scheduled, or before cancelling one — you need the id, and the user will refer to a task by what it does, not by its id.",
+        parameters: noParameters,
+        async execute(_toolCallId, _input, signal) {
+          signal?.throwIfAborted();
+          return result({ schedules: await schedule.list() });
+        },
       },
-    }),
+      { retry: "idempotent" },
+    ),
     candidate({
       name: "update_schedule",
       label: "Update schedule",

package/src/pi/tool/skill.ts

@@ -534,6 +534,9 @@ export async function skillPiToolCandidates(
       ...(name === "activate_skill"
         ? { exposureMode: "direct" as const }
         : { exposureMode: "codemode" as const }),
+      ...(name === "activate_skill" || name === "read_skill_resource"
+        ? { retry: "idempotent" as const }
+        : {}),
       tool: adapted,
     };
   });

package/src/pi/tool/subagent.ts

@@ -182,6 +182,7 @@ export function subagentPiToolCandidates(
       owner: `subagent:${type.name}`,
       exposureMode: "direct",
       requiredExecutionLevel: "safe",
+      retry: "idempotent",
       tool,
     };
   });
@@ -265,6 +266,7 @@ export function subagentPiToolCandidates(
     owner: "subagent:background",
     exposureMode: "direct",
     requiredExecutionLevel: "low",
+    retry: "idempotent",
     summary: "Dispatch a background sub-agent",
     source: "action",
     tool: background,

package/src/pi/tool/time.ts

@@ -11,6 +11,7 @@ export function getTimePiToolCandidate(owner = "utilities"): PiToolCandidate {
   return {
     owner,
     requiredExecutionLevel: "safe",
+    retry: "idempotent",
     source: "action",
     summary: "Get the current server time as an ISO string.",
     tool: {

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

@@ -707,6 +707,12 @@ async function callOpenAIStream(
         tools: isOpenRouter
             ? [{ type: "openrouter:web_search", parameters: { max_uses: 10 } }]
             : [{ type: "web_search" }],
+        ...(isOpenRouter
+            ? {
+                max_output_tokens: Math.min(model.maxTokens, OPENROUTER_MAX_OUTPUT_TOKENS),
+                reasoning: { effort: "none" },
+            }
+            : {}),
         ...(!isOpenRouter
             ? {
                 include: isCodex
@@ -1108,6 +1114,8 @@ export function applyCitations(text: string, groundingMetadata: any): { text: st
 
 const PROVIDER_ERROR_MAX_LENGTH = 2_048;
 const NATIVE_TIMEOUT_MS = 60_000;
+const OPENROUTER_MAX_OUTPUT_TOKENS = 4_096;
+export const WEB_SEARCH_MODEL = "deepseek/deepseek-v4-flash";
 
 export interface WebSearchOptions {
     endpoint: RuntimeModelEndpoint;

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

@@ -121,6 +121,7 @@ export function webSearchPiToolCandidate(
     owner: "runtime-base",
     exposureMode: "direct",
     requiredExecutionLevel: "safe",
+    retry: "idempotent",
     source: "action",
     tool,
   };

package/src/pi/tool/workspace-sandbox.ts

@@ -402,6 +402,7 @@ export function workspacePiToolCandidates(
     {
       owner: "workspace",
       requiredExecutionLevel: "safe" as const,
+      retry: "idempotent" as const,
       // read 是大输出外置协议的**出口**：溢出结果的提示语指向它。出口自己再外置，
       // 就会变成 read → 新 artifact → read 的死循环，模型只能靠猜逃出去。
       // 分页上限由 `pageReadResult` 自己保证，不需要外置也不会撑爆一轮。
@@ -411,6 +412,7 @@ export function workspacePiToolCandidates(
     ...[write, edit, list, find, grep, remove].map((tool) => ({
       owner: "workspace",
       requiredExecutionLevel: "safe" as const,
+      retry: "idempotent" as const,
       tool,
     })),
   ];
@@ -621,6 +623,7 @@ export function sandboxPiToolCandidates(
     {
       owner: "sandbox",
       requiredExecutionLevel: "safe",
+      retry: "idempotent",
       tool: processLogs,
     },
     {

package/src/runtime-agent.ts

@@ -301,6 +301,7 @@ export interface RuntimeAgentControls {
     requestId?: string,
     reason?: string,
   ): Promise<{ ok: boolean }>;
+  stopAllSubmissions(reason?: string): Promise<{ ok: boolean }>;
 }
 
 export interface RuntimeAgentConfigControls<Command, Change> {
@@ -525,6 +526,10 @@ export function defineRuntimeAgent<
         return super.stopTurn(requestId, reason);
       }
 
+      stopAllSubmissions(reason?: string): Promise<{ ok: boolean }> {
+        return super.stopAllSubmissions(reason);
+      }
+
       protected override async prepareAgentToolRun(
         input: unknown,
       ): Promise<string> {
@@ -922,6 +927,10 @@ export function defineRuntimeAgent<
     callableContext,
   );
   callable()(GeneratedRuntimeAgent.prototype.stopTurn, callableContext);
+  callable()(
+    GeneratedRuntimeAgent.prototype.stopAllSubmissions,
+    callableContext,
+  );
 
   return GeneratedRuntimeAgent as unknown as RuntimeAgentClass<
     Env,