@dudousxd/nestjs-agent-core 0.4.0 → 0.6.0

package/dist/index.d.cts

@@ -1,4 +1,5 @@
 import { StandardSchemaV1 } from '@standard-schema/spec';
+import { ChannelRegistry } from '@dudousxd/nestjs-diagnostics';
 
 /** Who is driving the turn. Roles + tenant come from the host app (nestjs-context/authz). */
 interface Actor {
@@ -119,6 +120,11 @@ interface QuotaView {
 interface Decision {
     approved: boolean;
     reason?: string;
+    /**
+     * Opaque ref of WHO decided (e.g. a console admin). When absent, the run's own actor decided
+     * (the chat flow).
+     */
+    executedByRef?: string;
 }
 type MessageRole = 'user' | 'assistant' | 'system';
 /**
@@ -275,6 +281,38 @@ 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.
@@ -325,6 +363,11 @@ declare const AGENT_ACTOR_DIRECTORY: unique symbol;
  * {@link import('./spi/attachment-staging.js').AttachmentStagingStore}.
  */
 declare const AGENT_ATTACHMENT_STAGING: unique symbol;
+/**
+ * Console-side HITL approve/reject for the cross-thread approvals inbox. Optional — see
+ * {@link import('./spi/approval-port.js').AgentApprovalPort}.
+ */
+declare const AGENT_APPROVAL_PORT: unique symbol;
 
 /**
  * Per-invocation context handed to a tool handler. Host-supplied bits are optional. Identity lives
@@ -541,6 +584,21 @@ 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;
+    /** sha256 hex of the run's resolved (pre-RAG) system prompt — identifies the prompt VERSION. */
+    promptHash?: 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>;
@@ -570,6 +628,15 @@ interface AgentStore {
      * 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
@@ -822,6 +889,77 @@ 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;
+    /** sha256 hex of the run's resolved (pre-RAG) system prompt; `null` for a run recorded before this shipped. */
+    promptHash: string | null;
+}
+/** One tool call awaiting a HITL decision, for the cross-thread approvals inbox. */
+interface PendingApprovalRow {
+    toolCallId: string;
+    toolName: string;
+    input: unknown;
+    threadId: string;
+    threadTitle: string;
+    /** Who asked — the run's actor. */
+    actorRef: string;
+    agentName: string | null;
+    /** ISO timestamp. */
+    requestedAt: string;
+}
+/** Governance rollup for one tool over a range. */
+interface ToolStatRow {
+    toolName: string;
+    toolType: string;
+    calls: number;
+    failed: number;
+    rejected: number;
+    /** p95 of executionMs across executed calls; null when none carry it. */
+    p95ExecutionMs: number | null;
+}
 /**
  * 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
@@ -835,6 +973,15 @@ 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[]>;
+    /** Tool calls sitting `pending_approval`, oldest first — an inbox drains from the back. Capped at `limit`. */
+    pendingApprovals(limit: number): Promise<PendingApprovalRow[]>;
+    /** Per-tool call/failure/rejection/latency rollup over the range, highest call count first. */
+    toolStats(range: GovernanceRange): Promise<ToolStatRow[]>;
 }
 
 /**
@@ -876,6 +1023,21 @@ interface AttachmentStagingStore {
     stage(input: StageAttachmentInput): Promise<MessageAttachment>;
 }
 
+/**
+ * Console-side HITL decisions. Implemented by the agent runtime (the nestjs package binds it to
+ * the same signal path chat approvals use); the dashboard injects it OPTIONALLY — absent = the
+ * approvals inbox renders read-only.
+ */
+interface AgentApprovalPort {
+    approve(toolCallId: string, opts?: {
+        executedByRef?: string;
+    }): Promise<void>;
+    reject(toolCallId: string, opts?: {
+        executedByRef?: string;
+        reason?: string;
+    }): Promise<void>;
+}
+
 /**
  * The pure aggregation core shared by every {@link import('../spi/governance-queries.js').AgentGovernanceQueries}
  * adapter (MikroORM, Drizzle, in-memory). An adapter's only job is the DB-specific row fetch; the
@@ -1066,15 +1228,36 @@ 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`
@@ -1157,5 +1340,28 @@ declare function publishAgentRunFinished(payload: AgentRunFinished): void;
 declare function publishAgentRunFailed(payload: AgentRunFailed): void;
 declare function publishAgentDelegated(payload: AgentDelegated): void;
 declare function publishAgentRetrieved(payload: AgentRetrieved): void;
+/** Every event key declared on `ChannelRegistry['agent']` above — derived, not hand-copied. */
+type AgentDiagnosticEvent = keyof ChannelRegistry['agent'];
+/**
+ * All 8 events on `ChannelRegistry['agent']`, in a stable order — handy for wiring subscribers
+ * (mirrors nestjs-media's `MEDIA_DIAGNOSTIC_EVENTS`). A drift between this list and the registry is
+ * a compile error in both directions: an extra/misspelled entry fails this array's own
+ * `readonly AgentDiagnosticEvent[]` annotation immediately; a missing entry fails the
+ * {@link AgentDiagnosticEventsCoverAllKeys} check below.
+ */
+declare const AGENT_DIAGNOSTIC_EVENTS: readonly AgentDiagnosticEvent[];
+/**
+ * The telescope key for an agent diagnostics channel — `agent:<event>`. This is the key the
+ * `@dudousxd/nestjs-diagnostics-telescope` generic bridge matches its `exclude` option against,
+ * and the label its "Busiest events" panel shows. Distinct from the `aviary:agent:<event>` channel
+ * name used on the wire. Mirrors `mediaDiagnosticKey`.
+ */
+type AgentDiagnosticKey = `agent:${AgentDiagnosticEvent}`;
+/**
+ * Compose the telescope key for an agent event, typed against {@link AgentDiagnosticEvent} so a
+ * misspelled event is a compile error. Feed the result to `nestjsDiagnosticsTelescope({ exclude:
+ * [...] })` to mute a noisy channel, e.g. `agentDiagnosticKey('message')`.
+ */
+declare function agentDiagnosticKey(event: AgentDiagnosticEvent): AgentDiagnosticKey;
 
-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 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 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 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 };
+export { AGENT_ACTOR_DIRECTORY, AGENT_ACTOR_RESOLVER, AGENT_APPROVAL_PORT, AGENT_ATTACHMENT_STAGING, AGENT_DEPS_FACTORY, AGENT_DIAGNOSTIC_EVENTS, 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 AgentApprovalPort, type AgentCatalogEntry, type AgentDefinition, type AgentDelegated, type AgentDiagnosticEvent, type AgentDiagnosticKey, 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 PendingApprovalRow, 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 ToolStatRow, type ToolStepCtx, type ToolStepEnvelope, type UpdateThreadInput, type UpdateToolCallInput, type UsagePurpose, type UsageTrendPoint, agentDiagnosticKey, bucketByActor, bucketByModel, bucketByThread, bucketUsageTrend, dayBoundsUtc, encodeStreamEvent, estimateCost, filterToolsByAllowList, filterToolsByRole, publishAgentDelegated, publishAgentMessage, publishAgentQuotaExceeded, publishAgentRetrieved, publishAgentRunFailed, publishAgentRunFinished, publishAgentRunStarted, publishAgentToolCall, runAgentLoop, seedModelPrices, withToolTimeout };

package/dist/index.d.ts

@@ -1,4 +1,5 @@
 import { StandardSchemaV1 } from '@standard-schema/spec';
+import { ChannelRegistry } from '@dudousxd/nestjs-diagnostics';
 
 /** Who is driving the turn. Roles + tenant come from the host app (nestjs-context/authz). */
 interface Actor {
@@ -119,6 +120,11 @@ interface QuotaView {
 interface Decision {
     approved: boolean;
     reason?: string;
+    /**
+     * Opaque ref of WHO decided (e.g. a console admin). When absent, the run's own actor decided
+     * (the chat flow).
+     */
+    executedByRef?: string;
 }
 type MessageRole = 'user' | 'assistant' | 'system';
 /**
@@ -275,6 +281,38 @@ 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.
@@ -325,6 +363,11 @@ declare const AGENT_ACTOR_DIRECTORY: unique symbol;
  * {@link import('./spi/attachment-staging.js').AttachmentStagingStore}.
  */
 declare const AGENT_ATTACHMENT_STAGING: unique symbol;
+/**
+ * Console-side HITL approve/reject for the cross-thread approvals inbox. Optional — see
+ * {@link import('./spi/approval-port.js').AgentApprovalPort}.
+ */
+declare const AGENT_APPROVAL_PORT: unique symbol;
 
 /**
  * Per-invocation context handed to a tool handler. Host-supplied bits are optional. Identity lives
@@ -541,6 +584,21 @@ 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;
+    /** sha256 hex of the run's resolved (pre-RAG) system prompt — identifies the prompt VERSION. */
+    promptHash?: 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>;
@@ -570,6 +628,15 @@ interface AgentStore {
      * 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
@@ -822,6 +889,77 @@ 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;
+    /** sha256 hex of the run's resolved (pre-RAG) system prompt; `null` for a run recorded before this shipped. */
+    promptHash: string | null;
+}
+/** One tool call awaiting a HITL decision, for the cross-thread approvals inbox. */
+interface PendingApprovalRow {
+    toolCallId: string;
+    toolName: string;
+    input: unknown;
+    threadId: string;
+    threadTitle: string;
+    /** Who asked — the run's actor. */
+    actorRef: string;
+    agentName: string | null;
+    /** ISO timestamp. */
+    requestedAt: string;
+}
+/** Governance rollup for one tool over a range. */
+interface ToolStatRow {
+    toolName: string;
+    toolType: string;
+    calls: number;
+    failed: number;
+    rejected: number;
+    /** p95 of executionMs across executed calls; null when none carry it. */
+    p95ExecutionMs: number | null;
+}
 /**
  * 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
@@ -835,6 +973,15 @@ 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[]>;
+    /** Tool calls sitting `pending_approval`, oldest first — an inbox drains from the back. Capped at `limit`. */
+    pendingApprovals(limit: number): Promise<PendingApprovalRow[]>;
+    /** Per-tool call/failure/rejection/latency rollup over the range, highest call count first. */
+    toolStats(range: GovernanceRange): Promise<ToolStatRow[]>;
 }
 
 /**
@@ -876,6 +1023,21 @@ interface AttachmentStagingStore {
     stage(input: StageAttachmentInput): Promise<MessageAttachment>;
 }
 
+/**
+ * Console-side HITL decisions. Implemented by the agent runtime (the nestjs package binds it to
+ * the same signal path chat approvals use); the dashboard injects it OPTIONALLY — absent = the
+ * approvals inbox renders read-only.
+ */
+interface AgentApprovalPort {
+    approve(toolCallId: string, opts?: {
+        executedByRef?: string;
+    }): Promise<void>;
+    reject(toolCallId: string, opts?: {
+        executedByRef?: string;
+        reason?: string;
+    }): Promise<void>;
+}
+
 /**
  * The pure aggregation core shared by every {@link import('../spi/governance-queries.js').AgentGovernanceQueries}
  * adapter (MikroORM, Drizzle, in-memory). An adapter's only job is the DB-specific row fetch; the
@@ -1066,15 +1228,36 @@ 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`
@@ -1157,5 +1340,28 @@ declare function publishAgentRunFinished(payload: AgentRunFinished): void;
 declare function publishAgentRunFailed(payload: AgentRunFailed): void;
 declare function publishAgentDelegated(payload: AgentDelegated): void;
 declare function publishAgentRetrieved(payload: AgentRetrieved): void;
+/** Every event key declared on `ChannelRegistry['agent']` above — derived, not hand-copied. */
+type AgentDiagnosticEvent = keyof ChannelRegistry['agent'];
+/**
+ * All 8 events on `ChannelRegistry['agent']`, in a stable order — handy for wiring subscribers
+ * (mirrors nestjs-media's `MEDIA_DIAGNOSTIC_EVENTS`). A drift between this list and the registry is
+ * a compile error in both directions: an extra/misspelled entry fails this array's own
+ * `readonly AgentDiagnosticEvent[]` annotation immediately; a missing entry fails the
+ * {@link AgentDiagnosticEventsCoverAllKeys} check below.
+ */
+declare const AGENT_DIAGNOSTIC_EVENTS: readonly AgentDiagnosticEvent[];
+/**
+ * The telescope key for an agent diagnostics channel — `agent:<event>`. This is the key the
+ * `@dudousxd/nestjs-diagnostics-telescope` generic bridge matches its `exclude` option against,
+ * and the label its "Busiest events" panel shows. Distinct from the `aviary:agent:<event>` channel
+ * name used on the wire. Mirrors `mediaDiagnosticKey`.
+ */
+type AgentDiagnosticKey = `agent:${AgentDiagnosticEvent}`;
+/**
+ * Compose the telescope key for an agent event, typed against {@link AgentDiagnosticEvent} so a
+ * misspelled event is a compile error. Feed the result to `nestjsDiagnosticsTelescope({ exclude:
+ * [...] })` to mute a noisy channel, e.g. `agentDiagnosticKey('message')`.
+ */
+declare function agentDiagnosticKey(event: AgentDiagnosticEvent): AgentDiagnosticKey;
 
-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 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 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 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 };
+export { AGENT_ACTOR_DIRECTORY, AGENT_ACTOR_RESOLVER, AGENT_APPROVAL_PORT, AGENT_ATTACHMENT_STAGING, AGENT_DEPS_FACTORY, AGENT_DIAGNOSTIC_EVENTS, 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 AgentApprovalPort, type AgentCatalogEntry, type AgentDefinition, type AgentDelegated, type AgentDiagnosticEvent, type AgentDiagnosticKey, 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 PendingApprovalRow, 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 ToolStatRow, type ToolStepCtx, type ToolStepEnvelope, type UpdateThreadInput, type UpdateToolCallInput, type UsagePurpose, type UsageTrendPoint, agentDiagnosticKey, bucketByActor, bucketByModel, bucketByThread, bucketUsageTrend, dayBoundsUtc, encodeStreamEvent, estimateCost, filterToolsByAllowList, filterToolsByRole, publishAgentDelegated, publishAgentMessage, publishAgentQuotaExceeded, publishAgentRetrieved, publishAgentRunFailed, publishAgentRunFinished, publishAgentRunStarted, publishAgentToolCall, runAgentLoop, seedModelPrices, withToolTimeout };