@springbrand/agent-runtime 0.2.0-alpha.43 → 0.2.0-alpha.45

package/src/runtime.ts

@@ -66,6 +66,7 @@ import {
 } from "./telemetry";
 import { connectConfiguredMcpServers } from "./lib/mcp";
 import {
+  ContextOverflowError,
   PiRuntimeAdapter,
   MODEL_STREAM_STALL_TIMEOUT_MS,
   readModelStreamStallDetails,
@@ -109,6 +110,7 @@ import {
   isTerminalSubmissionStatus,
   RuntimeDatabase,
   type StoredSubmissionAdmission,
+  type SubmissionRecoveryReason,
   type SubmissionStatus,
 } from "./db/index";
 import {
@@ -150,6 +152,8 @@ export const CHAT_RECOVERY_MAX_ATTEMPTS = 5;
 export const CHAT_STALL_MAX_ATTEMPTS = 3;
 const CHAT_RECOVERY_TERMINAL_MESSAGE =
   "Recovery failed after multiple attempts. This generation has stopped and your progress has been saved. Send a new message to continue.";
+const CONTEXT_OVERFLOW_TERMINAL_MESSAGE =
+  "The model context is still too large after forced compaction. Start a new conversation or reduce the attached context.";
 
 type RuntimeEventOutboxPayload =
   | { readonly type: "model-usage"; readonly event: RuntimeModelUsageEvent }
@@ -172,7 +176,10 @@ interface StoredSubmission extends SubmissionReceipt {
   userMessageId?: string | null;
   regenerateMessageId?: string | null;
   recoveryErrorCount: number;
-  recoveryReason?: "no_meaningful_model_progress" | "transient_model_error";
+  recoveryReason?:
+    | "no_meaningful_model_progress"
+    | "transient_model_error"
+    | "context_overflow";
   runId: string | null;
   accountId: string | null;
   rateVersion: number | null;
@@ -532,6 +539,8 @@ export abstract class AgentRuntimeKernel<
         this.db.submissions.findRunning() as StoredSubmission | null,
       findNextPending: () =>
         this.db.submissions.findNextPending() as StoredSubmission | null,
+      listPending: () =>
+        this.db.submissions.listPending() as StoredSubmission[],
       updateAbortReason: (submissionId, reason) =>
         this.db.submissions.updateAbortReason(submissionId, reason),
     };
@@ -999,6 +1008,26 @@ export abstract class AgentRuntimeKernel<
         });
         this.projectNestedToolSettlement(submission.submissionId, event);
       },
+      onUnknownToolStarted: (event) => {
+        this.telemetry.capture("toolStarted", {
+          submissionId: submission.submissionId,
+          toolCallId: event.toolCallId,
+          toolName: event.toolName,
+          input: event.input,
+          occurredAt: event.occurredAt,
+        });
+      },
+      onUnknownToolFinished: (event) => {
+        this.telemetry.capture("toolFinished", {
+          submissionId: submission.submissionId,
+          toolCallId: event.toolCallId,
+          toolName: event.toolName,
+          outcome: event.outcome,
+          output: event.output,
+          outputBytes: json(event.output).length,
+          occurredAt: event.occurredAt,
+        });
+      },
       ...(toolExecutors && Object.keys(toolExecutors).length > 0
         ? { toolExecutors }
         : {}),
@@ -1662,7 +1691,7 @@ export abstract class AgentRuntimeKernel<
 
   // 作用：按当前模型输入预算压缩 Pi 上下文。
   // 调用：Pi Turn 适配器在发模型请求前通过 `transformContext` 回调。
-  // 原因：模型窗口必须预留最大输出，压缩也必须使用本 Snapshot 的模型和密钥。
+  // 原因：模型窗口必须预留最大输出，压缩也必须使用本 Snapshot 的模型。
   private async transformPiContext(
     submissionId: string,
     messages: Parameters<PiRuntimeTranscript["compactContext"]>[0],
@@ -1671,13 +1700,12 @@ export abstract class AgentRuntimeKernel<
     const snapshot = this.assembly();
     const inputBudget = snapshot.pi.model.contextWindow -
       snapshot.pi.model.maxTokens;
+    const force = this.readSubmission(submissionId)?.recoveryReason ===
+      "context_overflow";
     const compacted = await this.transcript.compactContext(messages, {
       compactAfterTokens: inputBudget * 0.85,
+      force,
       model: snapshot.pi.model,
-      apiKey: this.pi.resolveApiKey(
-        snapshot.bindings.provider,
-        snapshot.pi.model.id,
-      ),
       signal,
       submissionId,
       onCompactionPersisted: (event) => {
@@ -1710,6 +1738,7 @@ export abstract class AgentRuntimeKernel<
         });
       },
     });
+    if (force) this.db.submissions.clearRecoveryReason(submissionId);
     await this.drainRuntimeEvents();
     return compacted;
   }
@@ -2840,7 +2869,9 @@ export abstract class AgentRuntimeKernel<
         reason: submission.recoveryReason ?? "runtime_restart",
         attempt: Math.max(1, submission.recoveryErrorCount),
       });
-      this.db.submissions.clearRecoveryReason(submissionId);
+      if (submission.recoveryReason !== "context_overflow") {
+        this.db.submissions.clearRecoveryReason(submissionId);
+      }
       await this.broadcastApprovals();
     }
     if (mode !== "fresh") await this.ensureRuntimeReady();
@@ -3135,15 +3166,25 @@ export abstract class AgentRuntimeKernel<
         return this.readSubmission(submissionId)!;
       }
       const stalled = error instanceof ChatStreamStalledError;
+      const contextOverflow = error instanceof ContextOverflowError;
       const abortReason = this.readSubmission(submissionId)?.abortReason;
       if (!abortReason && (stalled || error instanceof RetryableModelError)) {
-        const recoveryErrorCount = this.db.submissions
-          .incrementRecoveryErrorCount(
+        const recoveryReason: SubmissionRecoveryReason = stalled
+          ? "no_meaningful_model_progress"
+          : contextOverflow
+            ? "context_overflow"
+            : "transient_model_error";
+        let contextOverflowRecoveryClaimed = false;
+        const recoveryErrorCount = this.db.transaction(() => {
+          if (contextOverflow) {
+            contextOverflowRecoveryClaimed = this.db.submissions
+              .claimContextOverflowRecovery(submissionId);
+          }
+          return this.db.submissions.incrementRecoveryErrorCount(
             submissionId,
-            stalled
-              ? "no_meaningful_model_progress"
-              : "transient_model_error",
+            recoveryReason,
           );
+        });
         if (stalled) {
           const stallDetails = readModelStreamStallDetails(errorText(error)) ?? {
             lastMeaningfulActivityAt:
@@ -3163,10 +3204,14 @@ export abstract class AgentRuntimeKernel<
         await this.broadcastApprovals();
         const stoppedDuringRecovery = this.readSubmission(submissionId)
           ?.abortReason;
+        const retryAllowed = contextOverflow
+          ? contextOverflowRecoveryClaimed &&
+            recoveryErrorCount < CHAT_RECOVERY_MAX_ATTEMPTS
+          : recoveryErrorCount <
+            (stalled ? CHAT_STALL_MAX_ATTEMPTS : CHAT_RECOVERY_MAX_ATTEMPTS);
         if (
           !stoppedDuringRecovery &&
-          recoveryErrorCount <
-            (stalled ? CHAT_STALL_MAX_ATTEMPTS : CHAT_RECOVERY_MAX_ATTEMPTS)
+          retryAllowed
         ) {
           const recoveryOutcome = await this.scheduleChatRecoveryRetry(
             {
@@ -3207,7 +3252,9 @@ export abstract class AgentRuntimeKernel<
         } else if (!stoppedDuringRecovery) {
           terminalIntent = {
             outcome: "failed",
-            message: CHAT_RECOVERY_TERMINAL_MESSAGE,
+            message: contextOverflow
+              ? CONTEXT_OVERFLOW_TERMINAL_MESSAGE
+              : CHAT_RECOVERY_TERMINAL_MESSAGE,
           };
         }
       }
@@ -4105,6 +4152,35 @@ export abstract class AgentRuntimeKernel<
     return result;
   }
 
+  /** Stops every running, queued, or steered user request in this Session. */
+  async stopAllSubmissions(reason?: string): Promise<{ ok: boolean }> {
+    const current = this.db.submissions.findRunning() ??
+      this.db.submissions.findNextPending();
+    if (current) {
+      this.db.transaction(() => {
+        for (const steer of this.db.steers.listForSubmission(
+          current.submissionId,
+        )) {
+          const message = JSON.parse(
+            steer.canonicalJson,
+          ) as PiCanonicalUserInput;
+          const userMessage = steer.uiMessageJson
+            ? JSON.parse(steer.uiMessageJson) as UIMessage & { role: "user" }
+            : undefined;
+          this.transcript.append(steer.messageId, message, {
+            submissionId: current.submissionId,
+            createdAt: steer.createdAt,
+            ...(userMessage ? { userMessage } : {}),
+          });
+          this.db.steers.deleteByMessageId(steer.messageId);
+        }
+      });
+    }
+    const result = await this.submissions.stopAll(reason ?? "Stopped");
+    await this.broadcastApprovals();
+    return result;
+  }
+
   /** 返回当前已装配 User Agent 的执行档位，不提供 Session 覆盖。 */
   executionLevel(): ExecutionLevel {
     return this.assembly().profile.executionLevel;
@@ -4553,13 +4629,20 @@ export abstract class AgentRuntimeKernel<
               ...(current.recoveryErrorCount === undefined
                 ? {}
                 : {
-                    recoveryAttempt: current.recoveryErrorCount,
+                    recoveryAttempt:
+                      current.recoveryReason === "context_overflow"
+                        ? 1
+                        : current.recoveryErrorCount,
                     // 上限按当前恢复原因取,否则 stall 会显示 "2/5" 却在第 3 次就终止。
                     recoveryMax:
-                      current.recoveryReason === "no_meaningful_model_progress"
-                        ? CHAT_STALL_MAX_ATTEMPTS
-                        : CHAT_RECOVERY_MAX_ATTEMPTS,
-                    ...(current.recoveryReason
+                      current.recoveryReason === "context_overflow"
+                        ? 1
+                        : current.recoveryReason ===
+                            "no_meaningful_model_progress"
+                          ? CHAT_STALL_MAX_ATTEMPTS
+                          : CHAT_RECOVERY_MAX_ATTEMPTS,
+                    ...(current.recoveryReason &&
+                        current.recoveryReason !== "context_overflow"
                       ? { recoveryReason: current.recoveryReason }
                       : {}),
                   }),