@springbrand/agent-runtime 0.2.0-alpha.44 → 0.2.0-alpha.45

package/src/runtime-assembler.ts

@@ -58,7 +58,7 @@ import {
   codeExecutionPiToolCandidate,
 } from "./pi/tool/core";
 import { skillPiToolCandidates } from "./pi/tool/skill";
-import { createWebSearch } from "./pi/tool/web-search";
+import { createWebSearch, WEB_SEARCH_MODEL } from "./pi/tool/web-search";
 import { subagentPiToolCandidates } from "./pi/tool/subagent";
 import { resolvePiModel } from "./pi/runtime-adapter/models";
 import {
@@ -170,7 +170,7 @@ interface ToolSurfaceInput {
   readonly hostTools: readonly PiToolCandidate[];
   readonly skills: readonly RuntimeSkillSourceBinding[];
   readonly enabledSubagents: readonly string[];
-  readonly webSearch: Parameters<typeof basePiToolCandidates>[0];
+  readonly webSearch?: Parameters<typeof basePiToolCandidates>[0];
   readonly extensions: readonly RuntimeExtensionConfig[];
   readonly policy?: RuntimeToolSurfacePolicy;
 }
@@ -581,14 +581,20 @@ class RuntimeBuilder {
           : `Model is configured by multiple endpoints: ${profile.model}`,
       );
     }
-    const resolvedModel = resolvePiModel(provider, profile.model);
-    const endpoint = endpoints[0]!;
-    const webSearch = createWebSearch({
-      endpoint,
-      model: profile.model,
-      maxTokens: resolvedModel.maxTokens,
-      reasoning: resolvedModel.reasoning,
-    });
+    const webSearchEndpoint = provider.endpoints.find((endpoint) =>
+      endpoint.models.includes(WEB_SEARCH_MODEL)
+    );
+    const webSearch = webSearchEndpoint
+      ? (() => {
+          const webSearchModel = resolvePiModel(provider, WEB_SEARCH_MODEL);
+          return createWebSearch({
+            endpoint: webSearchEndpoint,
+            model: WEB_SEARCH_MODEL,
+            maxTokens: webSearchModel.maxTokens,
+            reasoning: webSearchModel.reasoning,
+          });
+        })()
+      : undefined;
     const skillSources = [...this.skillSources.values()];
     const toolSurface = await createToolSurface({
       ...(this.codeExecution ? { codeExecution: this.codeExecution } : {}),
@@ -606,7 +612,7 @@ class RuntimeBuilder {
       ],
       skills: skillSources,
       enabledSubagents: [...this.enabledSubagents],
-      webSearch,
+      ...(webSearch ? { webSearch } : {}),
       extensions: [...this.extensions.values()],
       ...(this.toolPolicy ? { policy: this.toolPolicy } : {}),
     });

package/src/runtime.ts

@@ -48,7 +48,6 @@ import type {
 } from "./kernel/extensions";
 import type {
   RuntimeActivity,
-  RuntimeActivityProjection,
   RuntimeState,
   RuntimeTurnState,
 } from "./kernel/state";
@@ -67,6 +66,7 @@ import {
 } from "./telemetry";
 import { connectConfiguredMcpServers } from "./lib/mcp";
 import {
+  ContextOverflowError,
   PiRuntimeAdapter,
   MODEL_STREAM_STALL_TIMEOUT_MS,
   readModelStreamStallDetails,
@@ -110,6 +110,7 @@ import {
   isTerminalSubmissionStatus,
   RuntimeDatabase,
   type StoredSubmissionAdmission,
+  type SubmissionRecoveryReason,
   type SubmissionStatus,
 } from "./db/index";
 import {
@@ -151,10 +152,11 @@ 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 }
-  | { readonly type: "activity"; readonly projection: RuntimeActivityProjection }
   | {
       readonly type: "tool-settlement";
       readonly event: RuntimeToolSettlementEvent;
@@ -174,7 +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;
@@ -534,6 +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),
     };
@@ -1001,6 +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 }
         : {}),
@@ -1107,12 +1134,12 @@ export abstract class AgentRuntimeKernel<
   async onStart(): Promise<void> {
     this.runtimeLoad.reset();
     this.telemetry.recover(() => this.ensureRuntimeReady());
-    await this.broadcastApprovals();
     if (this.db.runtimeEvents.hasPending()) {
       this.ctx.waitUntil(
         this.ensureRuntimeReady().then(() => this.drainRuntimeEvents()),
       );
     }
+    await this.broadcastApprovals();
     await this.approvals.dispatchPendingContinuations();
     const running = this.db.submissions.findRunning() as StoredSubmission | null;
     if (running) {
@@ -1136,13 +1163,6 @@ export abstract class AgentRuntimeKernel<
         const payload = JSON.parse(row.body) as RuntimeEventOutboxPayload;
         if (payload.type === "model-usage") {
           await turnEvents.onModelUsage?.(payload.event);
-        } else if (payload.type === "activity") {
-          if (
-            !this.state.activity ||
-            payload.projection.revision >= this.state.activity.revision
-          ) {
-            await turnEvents.onActivityChanged?.(payload.projection);
-          }
         } else if (payload.type === "tool-settlement") {
           await turnEvents.onToolSettled?.(payload.event);
         } else if (payload.type === "subagent-usage") {
@@ -1671,7 +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],
@@ -1680,13 +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) => {
@@ -1719,6 +1738,7 @@ export abstract class AgentRuntimeKernel<
         });
       },
     });
+    if (force) this.db.submissions.clearRecoveryReason(submissionId);
     await this.drainRuntimeEvents();
     return compacted;
   }
@@ -2849,7 +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();
@@ -3144,15 +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:
@@ -3172,10 +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(
             {
@@ -3216,7 +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,
           };
         }
       }
@@ -4114,6 +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;
@@ -4562,13 +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 }
                       : {}),
                   }),
@@ -4607,13 +4681,6 @@ export abstract class AgentRuntimeKernel<
             backgroundWork,
             revision: (currentActivity?.revision ?? 0) + 1,
           };
-      this.db.transaction(() => {
-        this.db.runtimeEvents.insert({
-          eventId: `session-activity:${nextActivity.revision}`,
-          body: json({ type: "activity", projection: nextActivity }),
-          createdAt: Date.now(),
-        });
-      });
       if (
         this.state?.approvals === undefined ||
         json(this.state.approvals) !== json(approvals) ||
@@ -4627,9 +4694,14 @@ export abstract class AgentRuntimeKernel<
           turn,
         });
       }
-      this.ctx.waitUntil(this.drainRuntimeEvents());
+      const projection = this.turnEventsPort()?.onActivityChanged?.(nextActivity);
+      if (projection) {
+        this.ctx.waitUntil(
+          projection.catch(() => undefined),
+        );
+      }
     } catch {
-      // Rebuilding UI state is best effort; queued activity delivery retries independently.
+      // Approval projection is best effort and must not block execution.
     }
   }