@springbrand/agent-runtime 0.2.0-alpha.61 → 0.2.0-alpha.62

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@springbrand/agent-runtime",
-  "version": "0.2.0-alpha.61",
+  "version": "0.2.0-alpha.62",
   "type": "module",
   "files": [
     "src",

package/src/db/submission.repo.ts

@@ -11,6 +11,7 @@ export type SubmissionStatus =
   | "error";
 export type SubmissionRecoveryReason =
   | "no_meaningful_model_progress"
+  | "server_tool_timeout"
   | "transient_model_error"
   | "context_overflow";
 

package/src/kernel/state.ts

@@ -62,7 +62,10 @@ export interface RuntimeTurnState {
   activeMessageId?: string;
   recoveryAttempt?: number;
   recoveryMax?: number;
-  recoveryReason?: "no_meaningful_model_progress" | "transient_model_error";
+  recoveryReason?:
+    | "no_meaningful_model_progress"
+    | "server_tool_timeout"
+    | "transient_model_error";
   steerable: boolean;
   hasPendingSteer: boolean;
   queued: RuntimeQueuedSubmission[];

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

@@ -83,6 +83,8 @@ const MAX_WRAP_UP_TURNS = 2;
 
 export class RetryableModelError extends Error {}
 
+export class RetryableModelSessionError extends RetryableModelError {}
+
 export class ContextOverflowError extends RetryableModelError {}
 
 function normalizePlanToolCalls(message: AgentMessage): void {
@@ -610,7 +612,6 @@ export type PiCanonicalMessageCommit =
 export interface PiTurnTerminalIntent {
   readonly outcome: "succeeded" | "failed" | "aborted";
   readonly message?: string;
-  readonly resetModelSession?: boolean;
 }
 
 export interface PiPreparedTurnRunOptions {
@@ -1058,7 +1059,8 @@ export class PreparedPiTurnAdapter {
       ) || isModelStreamStallMessage(
         event.assistantMessageEvent.error.errorMessage,
       ) ||
-        isRecoverableAssistantError(event.assistantMessageEvent.error))
+        isRecoverableAssistantError(event.assistantMessageEvent.error) ||
+        isOpenRouterServerToolTimeout(event.assistantMessageEvent.error))
     ) {
       return;
     }
@@ -1073,10 +1075,11 @@ export class PreparedPiTurnAdapter {
         }
         const stalled = message.stopReason === "error" &&
           isModelStreamStallMessage(message.errorMessage);
+        const resetModelSession = isOpenRouterServerToolTimeout(message);
         const retryable = isRecoverableAssistantError(message);
-        if (stalled || retryable) {
+        if (stalled || retryable || resetModelSession) {
           const content = visibleAssistantContent(message);
-          if (content.length > 0) {
+          if (!resetModelSession && content.length > 0) {
             this.assistantOrdinal += 1;
             await this.options.onCanonicalMessage({
               kind: "commit-turn",
@@ -1090,6 +1093,9 @@ export class PreparedPiTurnAdapter {
           if (stalled) {
             throw new ChatStreamStalledError(message.errorMessage!);
           }
+          if (resetModelSession) {
+            throw new RetryableModelSessionError(message.errorMessage);
+          }
           throw new RetryableModelError(message.errorMessage);
         }
       }
@@ -1136,9 +1142,6 @@ export class PreparedPiTurnAdapter {
           );
           this.terminalIntent = {
             outcome: terminal.outcome,
-            ...(isOpenRouterServerToolTimeout(authoritativeMessage)
-              ? { resetModelSession: true }
-              : {}),
             ...(authoritativeMessage.errorMessage
               ? { message: authoritativeMessage.errorMessage }
               : terminal.message

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

@@ -2,7 +2,11 @@ import {
   PreparedPiTurnAdapter,
   type CreatePreparedPiTurnOptions,
 } from "./execution";
-export { ContextOverflowError, RetryableModelError } from "./execution";
+export {
+  ContextOverflowError,
+  RetryableModelError,
+  RetryableModelSessionError,
+} from "./execution";
 import {
   uiUserMessageToPi,
 } from "../message";

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

@@ -215,6 +215,7 @@ export function isRecoverableAssistantError(
     message.stopReason === "error" &&
     /thought[_ ]signature/i.test(message.errorMessage ?? "")
   ) return false;
+  // Runtime resets the model Session and gives this timeout its own one-retry budget.
   if (isOpenRouterServerToolTimeout(message)) return false;
   return isRetryableAssistantError(message) ||
     (message.stopReason === "error" &&

package/src/runtime.ts

@@ -71,6 +71,7 @@ import {
   MODEL_STREAM_STALL_TIMEOUT_MS,
   readModelStreamStallDetails,
   RetryableModelError,
+  RetryableModelSessionError,
   type PiCanonicalUserInput,
   type PiCanonicalTranscriptSnapshot,
   type PiChatRecoveryData,
@@ -145,6 +146,7 @@ const CONTINUATION_WATCHDOG_SECONDS = 300;
 // Host 侧的一次抖动会把 Session 永久卡在「队列已满」。
 const ADMISSION_ATTEMPT_WINDOW_MS = 60_000;
 export const CHAT_RECOVERY_MAX_ATTEMPTS = 5;
+const SERVER_TOOL_TIMEOUT_MAX_ATTEMPTS = 2;
 // stall 单独收窄。理由不是「stall 更不值得救」,而是它的重试**期望值和别的错不一样**:
 // 瞬时错(5xx / 断流)重跑一次往往就好了;stall 的重跑是拿同一份 transcript 让模型
 // 重新想同样久,如果它本来就超预算,再跑几次也一样超。把看门狗放宽到 240s 之后,
@@ -177,10 +179,7 @@ interface StoredSubmission extends SubmissionReceipt {
   userMessageId?: string | null;
   regenerateMessageId?: string | null;
   recoveryErrorCount: number;
-  recoveryReason?:
-    | "no_meaningful_model_progress"
-    | "transient_model_error"
-    | "context_overflow";
+  recoveryReason?: SubmissionRecoveryReason;
   runId: string | null;
   accountId: string | null;
   rateVersion: number | null;
@@ -893,7 +892,6 @@ export abstract class AgentRuntimeKernel<
         terminal: {
           outcome: "succeeded" | "failed" | "aborted";
           message?: string;
-          resetModelSession?: boolean;
         },
       ) => void | Promise<void>;
     } = {},
@@ -3062,7 +3060,6 @@ export abstract class AgentRuntimeKernel<
       | {
           outcome: "succeeded" | "failed" | "aborted";
           message?: string;
-          resetModelSession?: boolean;
         }
       | undefined;
     const startedAt = ready.createdAt;
@@ -3082,13 +3079,6 @@ export abstract class AgentRuntimeKernel<
           intent.outcome,
           intent.message,
         );
-        if (intent.resetModelSession) {
-          this.modelSessionId = crypto.randomUUID();
-          await this.ctx.storage.put(
-            MODEL_SESSION_ID_STORAGE_KEY,
-            this.modelSessionId,
-          );
-        }
       },
     });
     for (const pending of this.db.steers.listForSubmission(submissionId)) {
@@ -3212,13 +3202,23 @@ export abstract class AgentRuntimeKernel<
       }
       const stalled = error instanceof ChatStreamStalledError;
       const contextOverflow = error instanceof ContextOverflowError;
+      const resetModelSession = error instanceof RetryableModelSessionError;
       const abortReason = this.readSubmission(submissionId)?.abortReason;
       if (!abortReason && (stalled || error instanceof RetryableModelError)) {
+        if (resetModelSession) {
+          this.modelSessionId = crypto.randomUUID();
+          await this.ctx.storage.put(
+            MODEL_SESSION_ID_STORAGE_KEY,
+            this.modelSessionId,
+          );
+        }
         const recoveryReason: SubmissionRecoveryReason = stalled
           ? "no_meaningful_model_progress"
-          : contextOverflow
-            ? "context_overflow"
-            : "transient_model_error";
+          : resetModelSession
+            ? "server_tool_timeout"
+            : contextOverflow
+              ? "context_overflow"
+              : "transient_model_error";
         let contextOverflowRecoveryClaimed = false;
         const recoveryErrorCount = this.db.transaction(() => {
           if (contextOverflow) {
@@ -3253,7 +3253,11 @@ export abstract class AgentRuntimeKernel<
           ? contextOverflowRecoveryClaimed &&
             recoveryErrorCount < CHAT_RECOVERY_MAX_ATTEMPTS
           : recoveryErrorCount <
-            (stalled ? CHAT_STALL_MAX_ATTEMPTS : CHAT_RECOVERY_MAX_ATTEMPTS);
+            (stalled
+              ? CHAT_STALL_MAX_ATTEMPTS
+              : resetModelSession
+                ? SERVER_TOOL_TIMEOUT_MAX_ATTEMPTS
+                : CHAT_RECOVERY_MAX_ATTEMPTS);
         if (
           !stoppedDuringRecovery &&
           retryAllowed
@@ -4710,7 +4714,9 @@ export abstract class AgentRuntimeKernel<
                         : current.recoveryReason ===
                             "no_meaningful_model_progress"
                           ? CHAT_STALL_MAX_ATTEMPTS
-                          : CHAT_RECOVERY_MAX_ATTEMPTS,
+                          : current.recoveryReason === "server_tool_timeout"
+                            ? SERVER_TOOL_TIMEOUT_MAX_ATTEMPTS
+                            : CHAT_RECOVERY_MAX_ATTEMPTS,
                     ...(current.recoveryReason &&
                         current.recoveryReason !== "context_overflow"
                       ? { recoveryReason: current.recoveryReason }