@dudousxd/nestjs-agent 1.37.0 → 1.38.0

package/dist/{index-4caNqEQ-.d.cts → index-BJFM1tG8.d.cts}

@@ -1,5 +1,5 @@
 import { A2uiStreamOptions, a2uiThreadReplay } from '@dudousxd/nestjs-agent-core/a2ui';
-import { AgentAttachmentConfig, AttachmentStagingStore, AgentStore, Actor, ListActionProposals, ActionProposalView, ActionProposalMutationView, TextActionProposalVocabulary, AgentRunner, ModelCatalog, QuotaProvider, ActionApprovalMode, MessageAttachment, QuotaReport, ModelCatalogView, AgentHostContext, AttachmentRef, PageContext, ChatQueueState, SubscribeOptions, HumanReply, ThreadSummary, ThreadDetail, UpdateThreadInput, MessageFeedbackValue, MessageFeedback, StagedAttachment, ActorResolver } from '@dudousxd/nestjs-agent-core';
+import { AgentAttachmentConfig, AttachmentStagingStore, AgentStore, Actor, ListActionProposals, ActionProposalView, ActionProposalMutationView, TextActionProposalVocabulary, AgentRunner, ModelCatalog, QuotaProvider, ActionApprovalMode, MessageAttachment, QuotaReport, ModelCatalogView, AgentHostContext, AttachmentRef, PageContext, ChatQueueState, SubscribeOptions, HumanReply, ThreadSummary, ThreadDetail, UpdateThreadInput, MessageFeedbackValue, MessageFeedback, StagedAttachment, AiUsageInput, AiUsageOptions, AiUsageResult, ActorResolver } from '@dudousxd/nestjs-agent-core';
 import { Request, Response } from 'express';
 import { UiCapabilities } from '@dudousxd/nestjs-agent-core/genui';
 import { b as AgentModuleOptions, A as AgentDepsFactory, C as ChatQueueService, d as AgentProtocolAdapter } from './chat-queue.service-fopUN313.cjs';
@@ -588,6 +588,22 @@ declare class AgentService {
     private resolveRunForToolCall;
     /** Authorization seam for `cancel`: the caller must own the thread currently streaming this run. */
     private assertOwnsActiveStream;
+    /**
+     * Record AI spend that did not come from an agent turn — a transcription, a TTS render, a batch of
+     * embeddings, a direct OpenAI/OpenRouter SDK call — on the same ledger the agent's turns write, so
+     * cost, quota and the dashboard see one figure.
+     *
+     * Never throws into the caller's flow by default: a failed write or an invalid input is logged and
+     * answered with `{ recorded: false, error }`. Pass `{ throwOnError: true }` where losing the row
+     * must fail the caller.
+     *
+     * Counted against `actorRef` like an agent turn: the ledger is the budget (`LedgerQuotaProvider`),
+     * so its tokens count in the token windows (a row with no tokens adds 0 there) and its `costUsd` in
+     * the USD windows. A `QuotaStore` counter is bumped only where loop deps carry one (`AgentModule`
+     * wires none). `actorRef: null` is platform cost, recorded under `PLATFORM_ACTOR_REF` and counted
+     * against nobody.
+     */
+    recordUsage(input: AiUsageInput, options?: AiUsageOptions): Promise<AiUsageResult>;
 }
 
 interface A2uiAdapterOptions extends A2uiStreamOptions {

package/dist/{index-yC_Ra3CE.d.ts → index-De98CGXa.d.ts}

@@ -1,5 +1,5 @@
 import { A2uiStreamOptions, a2uiThreadReplay } from '@dudousxd/nestjs-agent-core/a2ui';
-import { AgentAttachmentConfig, AttachmentStagingStore, AgentStore, Actor, ListActionProposals, ActionProposalView, ActionProposalMutationView, TextActionProposalVocabulary, AgentRunner, ModelCatalog, QuotaProvider, ActionApprovalMode, MessageAttachment, QuotaReport, ModelCatalogView, AgentHostContext, AttachmentRef, PageContext, ChatQueueState, SubscribeOptions, HumanReply, ThreadSummary, ThreadDetail, UpdateThreadInput, MessageFeedbackValue, MessageFeedback, StagedAttachment, ActorResolver } from '@dudousxd/nestjs-agent-core';
+import { AgentAttachmentConfig, AttachmentStagingStore, AgentStore, Actor, ListActionProposals, ActionProposalView, ActionProposalMutationView, TextActionProposalVocabulary, AgentRunner, ModelCatalog, QuotaProvider, ActionApprovalMode, MessageAttachment, QuotaReport, ModelCatalogView, AgentHostContext, AttachmentRef, PageContext, ChatQueueState, SubscribeOptions, HumanReply, ThreadSummary, ThreadDetail, UpdateThreadInput, MessageFeedbackValue, MessageFeedback, StagedAttachment, AiUsageInput, AiUsageOptions, AiUsageResult, ActorResolver } from '@dudousxd/nestjs-agent-core';
 import { Request, Response } from 'express';
 import { UiCapabilities } from '@dudousxd/nestjs-agent-core/genui';
 import { b as AgentModuleOptions, A as AgentDepsFactory, C as ChatQueueService, d as AgentProtocolAdapter } from './chat-queue.service-YjiKVTKE.js';
@@ -588,6 +588,22 @@ declare class AgentService {
     private resolveRunForToolCall;
     /** Authorization seam for `cancel`: the caller must own the thread currently streaming this run. */
     private assertOwnsActiveStream;
+    /**
+     * Record AI spend that did not come from an agent turn — a transcription, a TTS render, a batch of
+     * embeddings, a direct OpenAI/OpenRouter SDK call — on the same ledger the agent's turns write, so
+     * cost, quota and the dashboard see one figure.
+     *
+     * Never throws into the caller's flow by default: a failed write or an invalid input is logged and
+     * answered with `{ recorded: false, error }`. Pass `{ throwOnError: true }` where losing the row
+     * must fail the caller.
+     *
+     * Counted against `actorRef` like an agent turn: the ledger is the budget (`LedgerQuotaProvider`),
+     * so its tokens count in the token windows (a row with no tokens adds 0 there) and its `costUsd` in
+     * the USD windows. A `QuotaStore` counter is bumped only where loop deps carry one (`AgentModule`
+     * wires none). `actorRef: null` is platform cost, recorded under `PLATFORM_ACTOR_REF` and counted
+     * against nobody.
+     */
+    recordUsage(input: AiUsageInput, options?: AiUsageOptions): Promise<AiUsageResult>;
 }
 
 interface A2uiAdapterOptions extends A2uiStreamOptions {

package/dist/index.cjs

@@ -1723,6 +1723,7 @@ function _ts_param4(paramIndex, decorator) {
   };
 }
 __name(_ts_param4, "_ts_param");
+var usageLogger = new import_common6.Logger("AgentService");
 var NO_USER_MESSAGE_CODE = "no_user_message";
 function sendableParams(params) {
   const message = params.message;
@@ -2966,6 +2967,31 @@ var AgentService2 = class {
       throw new import_common6.ForbiddenException("run belongs to another actor");
     }
   }
+  /**
+  * Record AI spend that did not come from an agent turn — a transcription, a TTS render, a batch of
+  * embeddings, a direct OpenAI/OpenRouter SDK call — on the same ledger the agent's turns write, so
+  * cost, quota and the dashboard see one figure.
+  *
+  * Never throws into the caller's flow by default: a failed write or an invalid input is logged and
+  * answered with `{ recorded: false, error }`. Pass `{ throwOnError: true }` where losing the row
+  * must fail the caller.
+  *
+  * Counted against `actorRef` like an agent turn: the ledger is the budget (`LedgerQuotaProvider`),
+  * so its tokens count in the token windows (a row with no tokens adds 0 there) and its `costUsd` in
+  * the USD windows. A `QuotaStore` counter is bumped only where loop deps carry one (`AgentModule`
+  * wires none). `actorRef: null` is platform cost, recorded under `PLATFORM_ACTOR_REF` and counted
+  * against nobody.
+  */
+  recordUsage(input, options = {}) {
+    return (0, import_nestjs_agent_core9.recordAiUsage)({
+      store: this.store,
+      quota: this.deps.forAgent().quota,
+      day: utcDay(),
+      input,
+      options,
+      warn: /* @__PURE__ */ __name((message) => usageLogger.warn(message), "warn")
+    });
+  }
 };
 AgentService2 = _ts_decorate5([
   (0, import_common6.Injectable)(),
@@ -6176,10 +6202,12 @@ var InlineAgentRunner = class _InlineAgentRunner {
   }
   async start(input, options = {}) {
     const runId = options.runId ?? crypto.randomUUID();
+    const acceptedAt = Date.now();
     const day = input.day ?? utcDay();
     const deps = this.factory.forAgent(input.agentName);
     this.live.add(runId);
     const hooks = this.topLevelHooks({
+      acceptedAt,
       runId,
       deps,
       actor: input.actor,
@@ -6201,8 +6229,9 @@ var InlineAgentRunner = class _InlineAgentRunner {
       const timed = durationMs !== void 0 ? {
         durationMs
       } : {};
+      const measured = (0, import_nestjs_agent_core32.runEndTimings)(error);
       if (error instanceof import_nestjs_agent_core32.RunCancelledError) {
-        await this.settleCancelled(runId, input, deps, durationMs);
+        await this.settleCancelled(runId, input, deps, durationMs, measured);
         return;
       }
       const message = error instanceof Error ? error.message : String(error);
@@ -6214,6 +6243,7 @@ var InlineAgentRunner = class _InlineAgentRunner {
         code,
         message,
         ...timed,
+        ...measured,
         context: (0, import_nestjs_agent_core32.agentRunContext)(input)
       });
       await this.store.recordRunEnd?.({
@@ -6291,7 +6321,7 @@ var InlineAgentRunner = class _InlineAgentRunner {
   * the stream ENDED after a `cancelled` frame rather than failed, so a client that retries failed
   * streams does not retry this one.
   */
-  async settleCancelled(runId, input, deps, durationMs) {
+  async settleCancelled(runId, input, deps, durationMs, measured = {}) {
     this.logger.log(`agent run ${runId} cancelled`);
     const timed = durationMs !== void 0 ? {
       durationMs
@@ -6300,6 +6330,7 @@ var InlineAgentRunner = class _InlineAgentRunner {
       runId,
       threadId: input.threadId,
       ...timed,
+      ...measured,
       context: (0, import_nestjs_agent_core32.agentRunContext)(input)
     });
     const writer = await deps.sink.open(runId);
@@ -6392,9 +6423,10 @@ var InlineAgentRunner = class _InlineAgentRunner {
     });
   }
   topLevelHooks(args) {
-    const { runId, deps, actor, day, threadId, chainBelow, input } = args;
+    const { runId, deps, actor, day, threadId, chainBelow, input, acceptedAt } = args;
     return {
       runId,
+      queueMs: /* @__PURE__ */ __name(() => Date.now() - acceptedAt, "queueMs"),
       // The loop ends the stream when the turn completes. Just before, the thread passes to the next
       // queued message (or is released), and the reader is told which — so by the time a client
       // sees the run end, the thread is already free or already running what it queued.
@@ -6519,10 +6551,12 @@ var InlineAgentRunner = class _InlineAgentRunner {
       const timed = durationMs !== void 0 ? {
         durationMs
       } : {};
+      const measured = (0, import_nestjs_agent_core32.runEndTimings)(error);
       const settled = {
         runId,
         threadId: subThread.id,
         ...timed,
+        ...measured,
         context: (0, import_nestjs_agent_core32.agentRunContext)(childInput)
       };
       if (cancelled) {