@dudousxd/nestjs-agent-core 0.3.3 → 0.5.0

package/dist/index.d.ts

@@ -48,6 +48,13 @@ interface ToolCallRequest {
     id: string;
     name: string;
     input: unknown;
+    /**
+     * The tool's declared kind (`ToolSpec.kind`), stamped by the loop from the tool registry so
+     * thread-read consumers know a call's kind without hardcoding a tool-name allowlist. Undefined
+     * only for a call the loop couldn't resolve against the registry (defensively treated as `read`
+     * wherever a definite value is required).
+     */
+    kind?: ToolKind;
 }
 /** Result of running a tool. */
 interface ToolResult {
@@ -82,6 +89,13 @@ interface MessageUsage {
      * non-reasoning models or providers that don't report it.
      */
     reasoningTokens?: number;
+    /**
+     * This turn's USD cost: the provider's own reported figure when it has one, else an estimate from
+     * the bound `AgentPricingStore` (cached once per run — see `AgentLoopDeps.pricingStore`), else
+     * `null` when no pricing store is bound or the model has no price row. Never `0` for an unpriced
+     * model — a real $0 turn and "we don't know" must stay distinguishable.
+     */
+    costUsd?: number | null;
 }
 type UsagePurpose = 'chat' | 'follow_ups';
 interface QuotaState {
@@ -229,6 +243,18 @@ interface ThreadSummary {
     createdAt: string;
     updatedAt: string;
     lastMessagePreview?: string;
+    /**
+     * The agent a `chat()` call on this thread uses when the caller doesn't name one explicitly.
+     * Optional — undefined for a store that doesn't implement `AgentStore.updateThread` (the only
+     * way to set it). The REST/service read-model normalizes this to `null` when absent.
+     */
+    defaultAgent?: string | null;
+    /**
+     * The runId of a currently-running turn on this thread, or `null` if none is running. Optional —
+     * undefined for a store that doesn't implement `AgentStore.activeRunForThread`. The REST/service
+     * read-model normalizes this to `null` when absent, so a client can always do `?? null`.
+     */
+    activeRunId?: string | null;
 }
 interface StoredMessage {
     id: string;
@@ -249,6 +275,38 @@ interface ThreadDetail extends ThreadSummary {
     activeStreamId?: string;
 }
 type ToolCallStatus = 'auto_executed' | 'pending_approval' | 'executed' | 'rejected' | 'failed';
+/**
+ * Serializable input for a dispatched model-turn step. Carries only data — the serving worker
+ * re-resolves the model/sink/registry from its own DI via AGENT_DEPS_FACTORY.forAgent(agentName).
+ */
+interface LlmStepEnvelope {
+    /** Undefined = default agent (same semantics as {@link AgentRunInput.agentName}). */
+    agentName?: string;
+    system: string;
+    messages: ModelMessage[];
+    /** The turn's actor — the handler re-derives tool definitions from it (definitionsFor). */
+    actor: Actor;
+}
+/**
+ * The serializable subset of `AiToolCtx` — everything except `host` (re-attached handler-side
+ * from DI).
+ */
+interface ToolStepCtx {
+    actor: Actor;
+    threadId: string;
+    runId: string;
+    requestId: string;
+    agentName?: string;
+    pageContext?: PageContext;
+}
+/** Serializable input for a dispatched tool-execution step. */
+interface ToolStepEnvelope {
+    toolName: string;
+    input: unknown;
+    ctx: ToolStepCtx;
+    /** Applied INSIDE the handler (`withToolTimeout`) — never as a durable step `timeoutMs`. */
+    timeoutMs?: number;
+}
 
 /**
  * Public, cross-lib-discoverable DI tokens.
@@ -288,6 +346,17 @@ declare const AGENT_EMBEDDING_PROVIDER: unique symbol;
 declare const AGENT_DEPS_FACTORY: unique symbol;
 /** App-wide, ordered `@SystemPromptContributor()` functions the loop appends after the agent base. */
 declare const AGENT_PROMPT_CONTRIBUTORS: unique symbol;
+/**
+ * Resolves opaque store `actorRef`s to human display labels for governance/dashboard read surfaces.
+ * Optional — see {@link import('./spi/actor-directory.js').ActorDirectory}.
+ */
+declare const AGENT_ACTOR_DIRECTORY: unique symbol;
+/**
+ * Persists an uploaded attachment somewhere the model can fetch it from and returns a
+ * `MessageAttachment`. Optional — see
+ * {@link import('./spi/attachment-staging.js').AttachmentStagingStore}.
+ */
+declare const AGENT_ATTACHMENT_STAGING: unique symbol;
 
 /**
  * Per-invocation context handed to a tool handler. Host-supplied bits are optional. Identity lives
@@ -406,20 +475,32 @@ interface ModelProvider {
  * Keeping this vocabulary neutral (not AI-SDK `UIMessageChunk`) means core never depends on `ai`:
  * the adapter owns model-parts → event, the transport owns event → UI-chunk.
  */
+
 type AgentStreamEvent = {
     kind: 'step-start';
-} | {
+}
+/**
+ * Closes the step opened by the matching `step-start`. Carries the model call's token usage and
+ * `costUsd` (an estimate from the bound pricing store, or `null` when unpriced/unbound — never a
+ * fabricated `0`) so a live client can render running cost without waiting for a thread re-fetch.
+ */
+ | {
     kind: 'step-finish';
+    usage?: MessageUsage;
+    costUsd?: number | null;
 } | {
     kind: 'text';
     text: string;
 } | {
     kind: 'reasoning';
     text: string;
-} | {
+}
+/** `toolKind` collapses `ToolKind`'s `'agent'` into `'read'` — delegation tools auto-execute like a read tool. */
+ | {
     kind: 'tool-input-start';
     id: string;
     name: string;
+    toolKind: 'read' | 'action';
 } | {
     kind: 'tool-input-delta';
     id: string;
@@ -429,6 +510,7 @@ type AgentStreamEvent = {
     id: string;
     name: string;
     input: unknown;
+    toolKind: 'read' | 'action';
 } | {
     kind: 'tool-output';
     id: string;
@@ -475,6 +557,12 @@ interface UpdateToolCallInput {
     executionMs?: number;
     executedByRef?: string;
 }
+/** Patch applied by {@link AgentStore.updateThread}. An omitted key leaves that field untouched. */
+interface UpdateThreadInput {
+    title?: string;
+    /** `null` clears the thread's default agent (falls back to the module default). */
+    defaultAgent?: string | null;
+}
 interface RecordUsageInput {
     threadId: string;
     actorRef: string;
@@ -485,6 +573,19 @@ interface RecordUsageInput {
     /** Provider-reported actual USD cost for this turn, when known (gateways report it). */
     costUsd?: number;
 }
+interface RecordRunStartInput {
+    runId: string;
+    threadId: string;
+    actorRef: string;
+    agentName?: string;
+}
+interface RecordRunEndInput {
+    runId: string;
+    status: 'completed' | 'failed';
+    durationMs?: number;
+    errorCode?: string;
+    errorMessage?: string;
+}
 /** ORM-agnostic persistence. Refs are string ids; adapters may add real relations. */
 interface AgentStore {
     createThread(input: CreateThreadInput): Promise<ThreadSummary>;
@@ -500,6 +601,29 @@ interface AgentStore {
      */
     promoteThread(threadId: string): Promise<void>;
     setActiveStream(threadId: string, runId: string | null): Promise<void>;
+    /**
+     * OPTIONAL: rename a thread and/or set its default agent in one write. Absent on a store that
+     * predates this — `setTitle` still covers title-only edits, so nothing else in the lib requires
+     * this method; the REST `PATCH /threads/:id` endpoint responds 501 for a `defaultAgent` change
+     * against a store that lacks it.
+     */
+    updateThread?(threadId: string, patch: UpdateThreadInput): Promise<void>;
+    /**
+     * OPTIONAL: the runId of a currently-running turn on this thread, or `null` if none is running.
+     * Lets a client that reconnects (page refresh) discover a run to reattach to via the existing
+     * `GET /chat/:runId/stream`, instead of only being told about a run right after starting it.
+     * Absent on a store that predates this — thread read/list payloads report `activeRunId: null`.
+     */
+    activeRunForThread?(threadId: string): Promise<string | null>;
+    /**
+     * OPTIONAL: persist the start of a run (turn). Replay-safe: called under a durable localStep.
+     * Absent on a store that predates run recording — reliability metrics degrade to zeros/empty.
+     */
+    recordRunStart?(run: RecordRunStartInput): Promise<void>;
+    /** OPTIONAL: settle a run's outcome. `errorCode`/`errorMessage` only when status is 'failed'. */
+    recordRunEnd?(end: RecordRunEndInput): Promise<void>;
+    /** OPTIONAL: bump the run's llm-step retry counter (dispatched-step attempt > 1). */
+    bumpRunRetries?(runId: string): Promise<void>;
     /**
      * The `actorRef` that owns a thread, or `null` if no such thread exists. The authorization seam
      * for thread-scoped endpoints (detail / delete / fork): the service compares this against the
@@ -752,6 +876,52 @@ interface ThreadActivityRow {
     totalTokens: number;
     lastActivityAt: string;
 }
+/** Aggregated run reliability over a range. */
+interface RunMetrics {
+    runs: number;
+    completed: number;
+    failed: number;
+    /** completed / runs, 0 when runs = 0. */
+    successRate: number;
+    /** Total llm-step retries across the range's runs. */
+    retries: number;
+    durationP50Ms: number | null;
+    durationP95Ms: number | null;
+}
+/** Run/failure/retry rollup for one agent over a range. */
+interface RunAgentBreakdownRow {
+    /** '(default)' when the run had none. */
+    agentName: string;
+    runs: number;
+    failed: number;
+    retries: number;
+}
+/** Failed-run count for one error code over a range. */
+interface RunErrorBreakdownRow {
+    errorCode: string;
+    count: number;
+}
+/** One point on the daily run/failure trend. */
+interface RunTrendPoint {
+    day: string;
+    runs: number;
+    failed: number;
+}
+/** A recent run for the reliability feed. */
+interface RecentRunRow {
+    runId: string;
+    threadId: string;
+    actorRef: string;
+    agentName: string | null;
+    /** 'running' | 'completed' | 'failed'. */
+    status: string;
+    durationMs: number | null;
+    errorCode: string | null;
+    errorMessage: string | null;
+    retries: number;
+    /** ISO timestamp. */
+    startedAt: string;
+}
 /**
  * The governance read-model. Cost is `inputTokens/1e6 * inputPricePer1m + outputTokens/1e6 *
  * outputPricePer1m` against the current pricing row per model; an unpriced model contributes 0 cost
@@ -765,6 +935,50 @@ interface AgentGovernanceQueries {
     usageTrend(range: GovernanceRange): Promise<UsageTrendPoint[]>;
     recentToolCalls(limit: number): Promise<ToolCallActivityRow[]>;
     recentThreads(limit: number): Promise<ThreadActivityRow[]>;
+    runMetrics(range: GovernanceRange): Promise<RunMetrics>;
+    runsByAgent(range: GovernanceRange): Promise<RunAgentBreakdownRow[]>;
+    runErrors(range: GovernanceRange): Promise<RunErrorBreakdownRow[]>;
+    runTrend(range: GovernanceRange): Promise<RunTrendPoint[]>;
+    recentRuns(limit: number): Promise<RecentRunRow[]>;
+}
+
+/**
+ * Optional read-side lookup from opaque store `actorRef`s to human display labels.
+ *
+ * The store keeps `actorRef` opaque by design (no FK into the host's user table — hosts own their
+ * own identity schema). That's fine for enforcement and accounting, but a governance/dashboard read
+ * surface showing a raw ref ("u_8f21...") instead of a name is a bad experience. Binding an
+ * `ActorDirectory` lets those surfaces resolve refs to labels; leaving it unbound just means they
+ * render the raw ref. Consumers inject via `AGENT_ACTOR_DIRECTORY`.
+ */
+interface ActorDirectory {
+    /** Resolve opaque actor refs to display labels. Missing refs may be omitted. */
+    resolveDisplay(refs: readonly string[]): Promise<Record<string, string>>;
+}
+
+/** Input to {@link AttachmentStagingStore.stage} — the raw bytes plus who uploaded them. */
+interface StageAttachmentInput {
+    data: Buffer;
+    filename: string;
+    contentType: string;
+    sizeBytes: number;
+    actor: Actor;
+}
+/**
+ * Optional upload-side seam for message attachments (an image/PDF a user attaches to a chat
+ * message before the model ever sees it). The lib never fetches bytes itself — {@link MessageAttachment.url}
+ * must already be reachable by the model provider — so something has to turn an uploaded file into
+ * that URL first. A store adapter (or a thin wrapper over the host's own media pipeline) implements
+ * this; consumers inject via `AGENT_ATTACHMENT_STAGING`. Unbound, the optional `POST /agent/attachments`
+ * upload controller is never mounted.
+ */
+interface AttachmentStagingStore {
+    /**
+     * Persist an uploaded file somewhere the model can later fetch (presigned URL etc.) and return the
+     * {@link MessageAttachment} to send with the next chat message. The lib never fetches bytes; the
+     * returned url must be reachable by the model provider.
+     */
+    stage(input: StageAttachmentInput): Promise<MessageAttachment>;
 }
 
 /**
@@ -936,6 +1150,12 @@ interface AgentLoopDeps {
     retriever?: Retriever;
     /** How many passages inject-mode retrieval requests. Undefined → 5. */
     retrievalTopK?: number;
+    /**
+     * Prices each step's token usage into `costUsd` (on the `step-finish` stream frame and the
+     * persisted assistant message's `usage`). The current price list is fetched ONCE per run (not per
+     * message/step) and reused for every step's estimate. Undefined → `costUsd` is always `null`.
+     */
+    pricingStore?: AgentPricingStore;
 }
 interface AgentLoopHooks {
     runId: string;
@@ -951,15 +1171,36 @@ interface AgentLoopHooks {
         text: string;
     }>;
     /**
-     * Checkpoint wrapper. Inline = call fn directly; durable = ctx.step(name, fn).
+     * Checkpoint wrapper. Inline = call fn directly; durable = ctx.localStep(name, fn) — the
+     * in-process checkpointed primitive (`name` is a checkpoint identity, not a worker group).
      * EVERY side-effect and control-flow read goes through this so durable replay returns
      * cached results (stable ids, no double-write, no re-streaming).
      */
     step<T>(name: string, fn: () => Promise<T>): Promise<T>;
+    /**
+     * Dispatch the model turn as a routed remote step. When present the loop uses it INSTEAD of
+     * running the model inline under hooks.step. Must resolve to exactly what deps.model.runTurn
+     * resolves. The durable runner enriches the envelope with sink routing (sinkRunId/childSink)
+     * on its side — core stays sink-topology-agnostic.
+     */
+    dispatchLlm?(index: number, envelope: LlmStepEnvelope): Promise<ModelTurnResult>;
+    /**
+     * Dispatch one tool execution as a routed remote step. Replaces ONLY the registry.invoke call
+     * (+ its timeout, applied handler-side); all persist steps around it stay local.
+     */
+    dispatchTool?(call: ToolCallRequest, envelope: ToolStepEnvelope): Promise<unknown>;
+    /**
+     * Recognizes the runner's control-flow signals (durable suspend / continue-as-new) so the loop
+     * rethrows them untouched instead of recording a failure. Inline runners omit it — they have no
+     * control-flow exceptions.
+     */
+    isControlFlowError?(error: unknown): boolean;
 }
 declare class QuotaExceededError extends Error {
     constructor();
 }
+/** Reject if `work` doesn't settle within `ms`. The underlying work is left to finish on its own. */
+declare function withToolTimeout<T>(work: Promise<T>, ms: number, toolName: string): Promise<T>;
 /**
  * The provider-agnostic agent turn, reused by both the inline and durable runners.
  * It drives the model→tools→model iteration; the runner supplies the `step`/`awaitApproval`
@@ -1043,4 +1284,4 @@ declare function publishAgentRunFailed(payload: AgentRunFailed): void;
 declare function publishAgentDelegated(payload: AgentDelegated): void;
 declare function publishAgentRetrieved(payload: AgentRetrieved): void;
 
-export { AGENT_ACTOR_RESOLVER, AGENT_DEPS_FACTORY, AGENT_DURABLE_RUNNER, AGENT_EMBEDDING_PROVIDER, AGENT_GOVERNANCE_QUERIES, AGENT_MODEL, AGENT_OPTIONS, AGENT_PRICING_STORE, AGENT_PROMPT_CONTRIBUTORS, AGENT_QUOTA_STORE, AGENT_REGISTRY, AGENT_RETRIEVER, AGENT_ROLES_POLICY, AGENT_RUNNER, AGENT_SINK, AGENT_STORE, AGENT_TOOL_REGISTRY, type Actor, type ActorResolver, type ActorSpendRow, type AgentCatalogEntry, type AgentDefinition, type AgentDelegated, type AgentGovernanceQueries, type AgentLoopDeps, type AgentLoopHooks, type AgentMessageEvent, type AgentPricingStore, type AgentQuotaExceeded, AgentRegistry, type AgentRetrieved, type AgentRunFailed, type AgentRunFinished, type AgentRunInput, type AgentRunStarted, type AgentRunner, type AgentStore, AgentStreamError, type AgentStreamEvent, type AgentToolCallEvent, type AiToolCtx, type AppendMessageInput, type CostUsage, type CreateThreadInput, type CurrentModelPrice, type Decision, DefaultRolesPolicy, type EmbeddingProvider, type GovernanceRange, type GovernanceUsageInput, type MessageAttachment, type MessageRole, type MessageUsage, type ModelMessage, type ModelPrice, type ModelPriceInput, type ModelProvider, type ModelSpendRow, type ModelTurnArgs, type ModelTurnResult, type PageContext, type Passage, type PromptBuilder, type PromptContext, type PromptContributor, QuotaExceededError, type QuotaState, type QuotaStore, type QuotaView, type RecordToolCallInput, type RecordUsageInput, type RerankOptions, type Reranker, type RetrieveOptions, type Retriever, type RolesPolicy, type SinkWriter, type StoredMessage, type StreamError, type ThreadActivityRow, type ThreadDetail, type ThreadMeta, type ThreadSpendRow, type ThreadSummary, type TokenStreamSink, type ToolCallActivityRow, type ToolCallRequest, type ToolCallStatus, type ToolDefinition, ToolForbiddenError, type ToolHandler, ToolInputInvalidError, type ToolKind, ToolNotFoundError, ToolRegistry, type ToolResult, type ToolSpec, type UpdateToolCallInput, type UsagePurpose, type UsageTrendPoint, bucketByActor, bucketByModel, bucketByThread, bucketUsageTrend, dayBoundsUtc, encodeStreamEvent, estimateCost, filterToolsByAllowList, filterToolsByRole, publishAgentDelegated, publishAgentMessage, publishAgentQuotaExceeded, publishAgentRetrieved, publishAgentRunFailed, publishAgentRunFinished, publishAgentRunStarted, publishAgentToolCall, runAgentLoop, seedModelPrices };
+export { AGENT_ACTOR_DIRECTORY, AGENT_ACTOR_RESOLVER, AGENT_ATTACHMENT_STAGING, AGENT_DEPS_FACTORY, AGENT_DURABLE_RUNNER, AGENT_EMBEDDING_PROVIDER, AGENT_GOVERNANCE_QUERIES, AGENT_MODEL, AGENT_OPTIONS, AGENT_PRICING_STORE, AGENT_PROMPT_CONTRIBUTORS, AGENT_QUOTA_STORE, AGENT_REGISTRY, AGENT_RETRIEVER, AGENT_ROLES_POLICY, AGENT_RUNNER, AGENT_SINK, AGENT_STORE, AGENT_TOOL_REGISTRY, type Actor, type ActorDirectory, type ActorResolver, type ActorSpendRow, type AgentCatalogEntry, type AgentDefinition, type AgentDelegated, type AgentGovernanceQueries, type AgentLoopDeps, type AgentLoopHooks, type AgentMessageEvent, type AgentPricingStore, type AgentQuotaExceeded, AgentRegistry, type AgentRetrieved, type AgentRunFailed, type AgentRunFinished, type AgentRunInput, type AgentRunStarted, type AgentRunner, type AgentStore, AgentStreamError, type AgentStreamEvent, type AgentToolCallEvent, type AiToolCtx, type AppendMessageInput, type AttachmentStagingStore, type CostUsage, type CreateThreadInput, type CurrentModelPrice, type Decision, DefaultRolesPolicy, type EmbeddingProvider, type GovernanceRange, type GovernanceUsageInput, type LlmStepEnvelope, type MessageAttachment, type MessageRole, type MessageUsage, type ModelMessage, type ModelPrice, type ModelPriceInput, type ModelProvider, type ModelSpendRow, type ModelTurnArgs, type ModelTurnResult, type PageContext, type Passage, type PromptBuilder, type PromptContext, type PromptContributor, QuotaExceededError, type QuotaState, type QuotaStore, type QuotaView, type RecentRunRow, type RecordRunEndInput, type RecordRunStartInput, type RecordToolCallInput, type RecordUsageInput, type RerankOptions, type Reranker, type RetrieveOptions, type Retriever, type RolesPolicy, type RunAgentBreakdownRow, type RunErrorBreakdownRow, type RunMetrics, type RunTrendPoint, type SinkWriter, type StageAttachmentInput, type StoredMessage, type StreamError, type ThreadActivityRow, type ThreadDetail, type ThreadMeta, type ThreadSpendRow, type ThreadSummary, type TokenStreamSink, type ToolCallActivityRow, type ToolCallRequest, type ToolCallStatus, type ToolDefinition, ToolForbiddenError, type ToolHandler, ToolInputInvalidError, type ToolKind, ToolNotFoundError, ToolRegistry, type ToolResult, type ToolSpec, type ToolStepCtx, type ToolStepEnvelope, type UpdateThreadInput, type UpdateToolCallInput, type UsagePurpose, type UsageTrendPoint, bucketByActor, bucketByModel, bucketByThread, bucketUsageTrend, dayBoundsUtc, encodeStreamEvent, estimateCost, filterToolsByAllowList, filterToolsByRole, publishAgentDelegated, publishAgentMessage, publishAgentQuotaExceeded, publishAgentRetrieved, publishAgentRunFailed, publishAgentRunFinished, publishAgentRunStarted, publishAgentToolCall, runAgentLoop, seedModelPrices, withToolTimeout };

package/dist/index.js

@@ -19,6 +19,8 @@ var AGENT_RETRIEVER = Symbol.for("@dudousxd/nestjs-agent:retriever");
 var AGENT_EMBEDDING_PROVIDER = Symbol.for("@dudousxd/nestjs-agent:embedding-provider");
 var AGENT_DEPS_FACTORY = Symbol.for("@dudousxd/nestjs-agent:deps-factory");
 var AGENT_PROMPT_CONTRIBUTORS = Symbol.for("@dudousxd/nestjs-agent:prompt-contributors");
+var AGENT_ACTOR_DIRECTORY = Symbol.for("@dudousxd/nestjs-agent:actor-directory");
+var AGENT_ATTACHMENT_STAGING = Symbol.for("@dudousxd/nestjs-agent:attachment-staging");
 
 // src/spi/token-stream-sink.ts
 var AgentStreamError = class extends Error {
@@ -358,6 +360,13 @@ function publishAgentRetrieved(payload) {
 __name(publishAgentRetrieved, "publishAgentRetrieved");
 
 // src/agent-loop.ts
+function resolveCostUsd(usage, reportedCostUsd, price) {
+  if (reportedCostUsd !== void 0) {
+    return reportedCostUsd;
+  }
+  return price === void 0 ? null : estimateCost(usage, price);
+}
+__name(resolveCostUsd, "resolveCostUsd");
 function buildContextBlock(passages) {
   const items = passages.map((passage, index) => {
     const label = passage.source !== void 0 ? ` (${passage.source})` : "";
@@ -427,7 +436,7 @@ var ToolTimeoutError = class ToolTimeoutError2 extends Error {
     this.name = "ToolTimeoutError";
   }
 };
-function withTimeout(work, ms, toolName) {
+function withToolTimeout(work, ms, toolName) {
   return new Promise((resolve, reject) => {
     const timer = setTimeout(() => reject(new ToolTimeoutError(toolName, ms)), ms);
     work.then((value) => {
@@ -439,7 +448,7 @@ function withTimeout(work, ms, toolName) {
     });
   });
 }
-__name(withTimeout, "withTimeout");
+__name(withToolTimeout, "withToolTimeout");
 function parseFollowUps(text, count) {
   const source = text.match(/\[[\s\S]*\]/)?.[0] ?? text;
   try {
@@ -544,6 +553,17 @@ async function runAgentLoop(deps, input, hooks) {
       agentName: input.agentName
     } : {}
   });
+  const startedAt = await hooks.step("run:started-at", () => Promise.resolve(Date.now()));
+  await hooks.step("persist:run:start", async () => {
+    await deps.store.recordRunStart?.({
+      runId: hooks.runId,
+      threadId: input.threadId,
+      actorRef: input.actor.id,
+      ...input.agentName !== void 0 ? {
+        agentName: input.agentName
+      } : {}
+    });
+  });
   let injectedPassages;
   if (deps.retriever !== void 0) {
     const retriever = deps.retriever;
@@ -562,24 +582,50 @@ ${buildContextBlock(passages)}`;
       count: passages.length
     });
   }
+  let prices = [];
+  if (deps.pricingStore !== void 0) {
+    const pricingStore = deps.pricingStore;
+    prices = await hooks.step("pricing:list", () => pricingStore.listCurrentPrices());
+  }
+  const priceByModel = new Map(prices.map((price) => [
+    price.modelId,
+    price
+  ]));
   for (let i = 0; i < maxSteps; i += 1) {
     await hooks.step(`stream:step-start:${i}`, async () => {
       await writer.write(encodeStreamEvent({
         kind: "step-start"
       }));
     });
-    const tools = await deps.registry.definitionsFor(input.actor, deps.rolesPolicy, deps.toolAllowList);
-    const turn = await hooks.step(`llm:${i}`, () => deps.model.runTurn({
-      system,
-      messages: modelMessages,
-      tools,
-      sink: writer
+    let turn;
+    if (hooks.dispatchLlm) {
+      turn = await hooks.dispatchLlm(i, {
+        ...input.agentName !== void 0 ? {
+          agentName: input.agentName
+        } : {},
+        system,
+        messages: modelMessages,
+        actor: input.actor
+      });
+    } else {
+      const tools = await deps.registry.definitionsFor(input.actor, deps.rolesPolicy, deps.toolAllowList);
+      turn = await hooks.step(`llm:${i}`, () => deps.model.runTurn({
+        system,
+        messages: modelMessages,
+        tools,
+        sink: writer
+      }));
+    }
+    const resolvedModelId = turn.modelId ?? deps.modelId ?? "unknown";
+    const costUsd = resolveCostUsd(turn.usage, turn.costUsd, priceByModel.get(resolvedModelId));
+    const toolCallsWithKind = turn.toolCalls.map((call) => ({
+      ...call,
+      kind: deps.registry.spec(call.name)?.kind ?? "read"
     }));
     await hooks.step(`persist:usage:${i}`, () => deps.store.recordUsage({
       threadId: input.threadId,
       actorRef: input.actor.id,
-      // provider-reported model wins over the configured fallback, so cost can't misattribute
-      modelId: turn.modelId ?? deps.modelId ?? "unknown",
+      modelId: resolvedModelId,
       purpose: "chat",
       usage: turn.usage,
       // persist the provider's actual cost when reported; the read-model prefers it over pricing
@@ -627,12 +673,15 @@ ${buildContextBlock(passages)}`;
       threadId: input.threadId,
       role: "assistant",
       content: turn.text,
-      usage: turn.usage,
+      usage: {
+        ...turn.usage,
+        costUsd
+      },
       ...input.agentName !== void 0 ? {
         agentName: input.agentName
       } : {},
-      ...turn.toolCalls.length > 0 ? {
-        toolCalls: turn.toolCalls
+      ...toolCallsWithKind.length > 0 ? {
+        toolCalls: toolCallsWithKind
       } : {},
       ...followUps !== void 0 ? {
         followUps
@@ -641,8 +690,8 @@ ${buildContextBlock(passages)}`;
     const assistantMessage = {
       role: "assistant",
       content: turn.text,
-      ...turn.toolCalls.length > 0 ? {
-        toolCalls: turn.toolCalls
+      ...toolCallsWithKind.length > 0 ? {
+        toolCalls: toolCallsWithKind
       } : {}
     };
     modelMessages.push(assistantMessage);
@@ -672,15 +721,17 @@ ${buildContextBlock(passages)}`;
     if (isFinalTurn) {
       await hooks.step(`stream:step-finish:${i}`, async () => {
         await writer.write(encodeStreamEvent({
-          kind: "step-finish"
+          kind: "step-finish",
+          usage: turn.usage,
+          costUsd
         }));
       });
       break;
     }
     const results = [];
-    for (const call of turn.toolCalls) {
+    for (const call of toolCallsWithKind) {
       const spec = deps.registry.spec(call.name);
-      const toolType = spec?.kind ?? "read";
+      const toolType = call.kind ?? "read";
       const ctx = {
         actor: input.actor,
         threadId: input.threadId,
@@ -784,11 +835,36 @@ ${buildContextBlock(passages)}`;
           status: "auto_executed"
         }));
       }
-      const startedAt = Date.now();
+      const startedAt2 = Date.now();
       try {
-        const invocation = hooks.step(`tool:${call.id}`, () => deps.registry.invoke(call.name, call.input, ctx, deps.rolesPolicy));
-        const output = deps.toolTimeoutMs !== void 0 ? await withTimeout(invocation, deps.toolTimeoutMs, call.name) : await invocation;
-        const executionMs = Date.now() - startedAt;
+        let output;
+        if (hooks.dispatchTool) {
+          const stepCtx = {
+            actor: input.actor,
+            threadId: input.threadId,
+            runId: hooks.runId,
+            requestId: hooks.runId,
+            ...input.agentName !== void 0 ? {
+              agentName: input.agentName
+            } : {},
+            ...input.pageContext !== void 0 ? {
+              pageContext: input.pageContext
+            } : {}
+          };
+          const envelope = {
+            toolName: call.name,
+            input: call.input,
+            ctx: stepCtx,
+            ...deps.toolTimeoutMs !== void 0 ? {
+              timeoutMs: deps.toolTimeoutMs
+            } : {}
+          };
+          output = await hooks.dispatchTool(call, envelope);
+        } else {
+          const invocation = hooks.step(`tool:${call.id}`, () => deps.registry.invoke(call.name, call.input, ctx, deps.rolesPolicy));
+          output = deps.toolTimeoutMs !== void 0 ? await withToolTimeout(invocation, deps.toolTimeoutMs, call.name) : await invocation;
+        }
+        const executionMs = Date.now() - startedAt2;
         await hooks.step(`persist:toolexec:${call.id}`, () => deps.store.updateToolCall({
           toolCallId: call.id,
           status: "executed",
@@ -811,7 +887,10 @@ ${buildContextBlock(passages)}`;
           durationMs: executionMs
         });
       } catch (error) {
-        const executionMs = Date.now() - startedAt;
+        if (hooks.isControlFlowError?.(error) === true) {
+          throw error;
+        }
+        const executionMs = Date.now() - startedAt2;
         const message = error instanceof Error ? error.message : String(error);
         await hooks.step(`persist:toolfail:${call.id}`, () => deps.store.updateToolCall({
           toolCallId: call.id,
@@ -850,13 +929,22 @@ ${buildContextBlock(passages)}`;
     });
     await hooks.step(`stream:step-finish:${i}`, async () => {
       await writer.write(encodeStreamEvent({
-        kind: "step-finish"
+        kind: "step-finish",
+        usage: turn.usage,
+        costUsd
       }));
     });
   }
   if (thread !== null && (thread.title === "" || thread.title === "New chat")) {
     await hooks.step("persist:title", () => deps.store.setTitle(input.threadId, deriveTitle(input.userText)));
   }
+  await hooks.step("persist:run:end", async () => {
+    await deps.store.recordRunEnd?.({
+      runId: hooks.runId,
+      status: "completed",
+      durationMs: Date.now() - startedAt
+    });
+  });
   await writer.end();
   publishAgentRunFinished({
     runId: hooks.runId,
@@ -871,7 +959,9 @@ ${buildContextBlock(passages)}`;
 }
 __name(runAgentLoop, "runAgentLoop");
 export {
+  AGENT_ACTOR_DIRECTORY,
   AGENT_ACTOR_RESOLVER,
+  AGENT_ATTACHMENT_STAGING,
   AGENT_DEPS_FACTORY,
   AGENT_DURABLE_RUNNER,
   AGENT_EMBEDDING_PROVIDER,
@@ -914,6 +1004,7 @@ export {
   publishAgentRunStarted,
   publishAgentToolCall,
   runAgentLoop,
-  seedModelPrices
+  seedModelPrices,
+  withToolTimeout
 };
 //# sourceMappingURL=index.js.map