@noetic-tools/core 5.3.0 → 5.4.1

package/dist/harness/agent-harness.d.ts

@@ -1,33 +1,21 @@
 import { ItemSchemaRegistry } from '@noetic-tools/types';
+import { AcpSessionStore } from '../runtime/acp-session-store.js';
 import type { LayerStateStore, RecallCache } from './deps/context.js';
 import type { DetachedSpawnOverrides } from './deps/interpreter.js';
 import type { CheckpointStore, EventBroadcaster, RestoreCheckpointOptions, StepLedgerRetention, StepLedgerStore } from './deps/runtime.js';
-import type { AgentConfig, AgentHarnessContract, AgentHooks, CallModelRequest, Channel, ChannelHandle, Context, ContextCacheConfig, ContextCacheStore, ContextData, ContextLayer, CwdState, DeliveryMode, DetachedHandle, ExecuteInput, ExecuteOptions, ExternalChannel, FsAdapter, HarnessResponse, HarnessStatus, Item, ItemSchemaExtensions, LLMResponse, LlmProviderConfig, ProjectionPolicy, RecallLayerOutput, SessionScope, ShellAdapter, Span, SteeringDecision, Step, StorageAdapter, StreamEvent, StreamingItem, SubHarnessSession, SubprocessAdapter, Tool, TraceExporter, ZodType } from './deps/types.js';
+import type { AcpLiveSession, AcpSessionInfo, AgentConfig, AgentHarnessContract, AgentHooks, CallModelRequest, Channel, ChannelHandle, Context, ContextCacheConfig, ContextCacheStore, ContextData, ContextLayer, CwdState, DeliveryMode, DetachedHandle, ExecuteInput, ExecuteOptions, ExternalChannel, FsAdapter, HarnessResponse, HarnessStatus, Item, ItemSchemaConfig, LLMResponse, LlmProviderConfig, ProjectionPolicy, RecallLayerOutput, SessionScope, SessionUsage, ShellAdapter, Span, SteeringDecision, Step, StorageAdapter, StreamEvent, StreamingItem, SubprocessAdapter, Tool, TraceExporter, ZodType } from './deps/types.js';
 export { createStreamIdleWatchdog } from './model-call.js';
-interface AgentHarnessOpts<TParams extends Record<string, unknown> = Record<string, unknown>> {
-    name: string;
-    initialStep?: Step<ContextData, string, string>;
-    /** Default context layers applied to every context created via `createContext()` / `execute()`. */
-    context?: ContextLayer[];
-    /** @deprecated Renamed to `context`. */
-    memory?: ContextLayer[];
-    storage?: StorageAdapter;
-    hooks?: AgentHooks;
+/**
+ * Storage-scoped environment configuration for `AgentHarness`.
+ *
+ * @public
+ */
+export interface StorageEnvironmentConfig {
     /**
-     * Harness-wide tool pool. Merged (identity-deduplicated) with tools
-     * collected from `initialStep` to form every context's `unifiedTools`.
-     * Use this when the workflow step tree is static and tools are supplied
-     * per harness instance rather than baked into individual `step.llm` calls.
+     * Key-value storage adapter backing context-layer persistence and the
+     * step-completion ledger. Defaults to an in-memory adapter.
      */
-    tools?: Tool[];
-    params: TParams;
-    paramsSchema?: ZodType<TParams>;
-    /** Filesystem adapter. Defaults to local node:fs when not provided. */
-    fs?: FsAdapter;
-    /** Shell adapter. Defaults to local sh when not provided. */
-    shell?: ShellAdapter;
-    /** Subprocess adapter. Defaults to an in-memory, same-process adapter. */
-    subprocess?: SubprocessAdapter;
+    adapter?: StorageAdapter;
     /**
      * Checkpoint store used by `harness.checkpoint(ctx)` / `harness.restore()`.
      * When absent, checkpoint/restore are no-ops — a zero-config harness keeps
@@ -35,6 +23,8 @@ interface AgentHarnessOpts<TParams extends Record<string, unknown> = Record<stri
      * to enable durable execution.
      */
     checkpointStore?: CheckpointStore;
+    /** Layer-state store override. Defaults to an in-memory store. */
+    layerStateStore?: LayerStateStore;
     /**
      * Bounds on the step-completion ledger that backs step-level resume. Only meaningful
      * alongside a `checkpointStore`. Defaults to `DEFAULT_STEP_LEDGER_RETENTION` — 128 KiB
@@ -43,11 +33,45 @@ interface AgentHarnessOpts<TParams extends Record<string, unknown> = Record<stri
      * never correctness.
      */
     stepLedgerRetention?: StepLedgerRetention;
-    llm?: LlmProviderConfig;
-    /** Harness-wide item schema extensions. */
-    itemSchemas?: ItemSchemaExtensions;
-    /** Whether unknown extension item types must match a registered schema. Defaults to true. */
-    strictItemSchemas?: boolean;
+}
+/**
+ * Execution-environment configuration for `AgentHarness`: the storage,
+ * filesystem, shell, and subprocess surfaces the agent runs against.
+ *
+ * @public
+ */
+export interface AgentEnvironmentConfig {
+    /** Storage configuration: adapter, durability stores, ledger retention. */
+    storage?: StorageEnvironmentConfig;
+    /** Filesystem adapter. Defaults to an in-memory adapter. */
+    fs?: FsAdapter;
+    /** Shell adapter. Defaults to an in-memory adapter. */
+    shell?: ShellAdapter;
+    /** Subprocess adapter. Defaults to an in-memory, same-process adapter. */
+    subprocess?: SubprocessAdapter;
+}
+interface AgentHarnessOpts<TParams extends Record<string, unknown> = Record<string, unknown>> {
+    name: string;
+    /** Root step tree executed for every turn submitted via `execute()`. */
+    agentGraph?: Step<ContextData, string, string>;
+    /** Default context layers applied to every context created via `createContext()` / `execute()`. */
+    contextLayers?: ContextLayer[];
+    hooks?: AgentHooks;
+    /**
+     * Harness-wide tool pool. Merged (identity-deduplicated) with tools
+     * collected from `agentGraph` to form every context's `unifiedTools`.
+     * Use this when the workflow step tree is static and tools are supplied
+     * per harness instance rather than baked into individual `step.callModel` calls.
+     */
+    tools?: Tool[];
+    params: TParams;
+    paramsSchema?: ZodType<TParams>;
+    /** Execution environment: storage, filesystem, shell, subprocess. */
+    environment?: AgentEnvironmentConfig;
+    /** Default provider configuration for model calls made through this harness. */
+    callModelDefaults?: LlmProviderConfig;
+    /** Harness-wide item schema extensions and validation strictness. */
+    itemSchemas?: ItemSchemaConfig;
     /** Default projection policy for all LLM steps. Individual steps override via `step.projection`. */
     projection?: ProjectionPolicy;
     /** When true, every layer is recalled atomically regardless of its `recallMode`. */
@@ -55,15 +79,8 @@ interface AgentHarnessOpts<TParams extends Record<string, unknown> = Record<stri
     /** Tuning for prompt-cache anchoring. See `ContextCacheConfig` for the defaults. */
     contextCache?: ContextCacheConfig;
     traceExporter?: TraceExporter;
-    layerStateStore?: LayerStateStore;
     /** Default delivery mode for messages that don't specify one. Defaults to `next-turn`. */
     defaultDeliveryMode?: DeliveryMode;
-    /**
-     * Abort the in-flight model call if the provider stream emits no events for this
-     * many milliseconds. Defaults to `DEFAULT_STREAM_IDLE_TIMEOUT_MS` (120s).
-     * Pass `0` or a negative number to disable the watchdog.
-     */
-    streamIdleTimeoutMs?: number;
     /**
      * Initial working directory for the harness. Used as the seed value for the
      * shared `cwdState.cwd` on every Context this harness creates, including
@@ -129,7 +146,7 @@ export declare class AgentHarness<TParams extends Record<string, unknown> = Reco
      * @internal
      */
     readonly stepLedgerRetention: Required<StepLedgerRetention>;
-    private readonly initialStep?;
+    private readonly agentGraph?;
     /** Harness-wide tool pool merged into every context's `unifiedTools`. */
     private readonly harnessTools;
     /** @internal Context layers configured for this harness. Exposed non-private
@@ -142,15 +159,42 @@ export declare class AgentHarness<TParams extends Record<string, unknown> = Reco
     private readonly rootRunDepth;
     private readonly callModelOverride?;
     private readonly defaultDeliveryMode;
-    private readonly streamIdleTimeoutMs;
     private readonly sessions;
     /**
-     * @internal Cross-step harness sessions keyed by `step.session.reuse`, kept
-     * alive for the life of this harness so reused coding-agent sessions span
-     * multiple steps. Reached by the interpreter's harness handler via
-     * `frameworkCast`; do not access from outside core.
+     * @internal Live ACP connections. Reached by the interpreter via
+     * `frameworkCast`; use {@link listAcpSessions}, {@link getAcpSession},
+     * {@link cancelAcpSession}, and {@link closeAcpSessions} from outside core.
+     */
+    readonly acpSessions: AcpSessionStore;
+    /**
+     * Snapshot of every live ACP sub-agent, for a UI that wants to show what is
+     * running and let a user act on it. Read-only: a session's turns are driven
+     * by steps, so nothing here can start work behind the runtime's back.
+     * @public
+     */
+    listAcpSessions(): AcpSessionInfo[];
+    /**
+     * The live connection + session behind a handle from {@link listAcpSessions}
+     * (or a step's `session.reuse` key), or `undefined` when nothing matches.
+     * @public
+     */
+    getAcpSession(key: string): AcpLiveSession | undefined;
+    /**
+     * Interrupt whatever the sub-agent under `key` is doing, via `session/cancel`.
+     * The connection stays open and the turn still returns — with the `cancelled`
+     * stop reason, which the running step surfaces as a `cancelled` error.
+     * Returns false when nothing matches the key.
+     * @public
+     */
+    cancelAcpSession(key: string): Promise<boolean>;
+    /**
+     * Close every ACP connection this harness still holds, whatever its
+     * keep-alive scope. A connection owns a live agent — usually a child process
+     * whose stdio keeps the event loop alive — so a `keepAlive: 'harness'`
+     * session must be closed by its owner or the host will not exit. Idempotent.
+     * @public
      */
-    readonly subHarnessSessions: Map<string, SubHarnessSession>;
+    closeAcpSessions(): Promise<void>;
     readonly layerStateStore: LayerStateStore;
     /** Per-harness memoization cache for `recallMode: 'eventual'` layers. */
     readonly recallCache: RecallCache;
@@ -193,6 +237,7 @@ export declare class AgentHarness<TParams extends Record<string, unknown> = Reco
     }): Promise<void>;
     getStatus(scope?: SessionScope): HarnessStatus;
     getQueueSize(scope?: SessionScope): number;
+    getUsage(scope?: SessionScope): SessionUsage;
     seedSessionHistory(threadId: string, items: ReadonlyArray<Item>): void;
     private getOrCreateSession;
     private requireSession;
@@ -225,9 +270,7 @@ export declare class AgentHarness<TParams extends Record<string, unknown> = Reco
         state?: unknown;
         threadId?: string;
         resourceId?: string;
-        context?: ContextLayer[];
-        /** @deprecated Renamed to `context`. */
-        memory?: ContextLayer[];
+        contextLayers?: ContextLayer[];
         /**
          * Initial cwd for the new context. When set, takes precedence over both
          * the parent snapshot and the harness root cwd — used by worktree
@@ -332,7 +375,7 @@ export declare class AgentHarness<TParams extends Record<string, unknown> = Reco
      *
      * Aborting a context rejects everything blocked on it (channel `recv`
      * waiters, parked senders) with `{ kind: 'cancelled' }`, stops its in-flight
-     * model call and sub-harness turn, and makes the next step boundary throw
+     * model call and ACP agent turn, and makes the next step boundary throw
      * `cancelled`.
      *
      * Cancellation is cooperative: a layer hook already in flight is allowed to

package/dist/harness/agent-harness.js

@@ -1,6 +1,7 @@
 import { frameworkCast, ItemSchemaRegistry, NoeticConfigError } from '@noetic-tools/types';
 import { SpanImpl } from '../observability/span-impl.js';
 import { NoopExporter } from '../observability/trace-exporter.js';
+import { AcpSessionStore } from '../runtime/acp-session-store.js';
 import { createInMemoryFsAdapter, createInMemoryShellAdapter, createInMemorySubprocessAdapter, OpenRouter, } from './deps/adapters.js';
 import { afterModelCallLayers, allocateBudgets, assembleView, beforeToolCallLayers, completeLayers, contextToExecCtx, createContextCacheStore, createLayerStateStore, createRecallCache, DEFAULT_PROJECTION, disposeLayers, executeRerender, initLayers, projectHistoryLayers, recallLayers, recallLayersAtomic, recallLayersEventual, resolveLayerTools, runAppendPipeline, storeLayers, } from './deps/context.js';
 import { collectAllTools, deduplicateTools, dispatchStepThroughAdapter, execute, prepareBandedView, } from './deps/interpreter.js';
@@ -8,7 +9,6 @@ import { buildItemStream, ChannelStore, ContextImpl, captureCheckpoint, clearChe
 import { SteeringAction } from './deps/types.js';
 import { AgentHarnessModelCaller } from './model-call.js';
 export { createStreamIdleWatchdog } from './model-call.js';
-import { resolveContextOption } from '../builders/context-option.js';
 import { buildItemSchemaRegistry } from './model-schema.js';
 //#endregion
 const DEFAULT_THREAD_ID = '__default__';
@@ -178,7 +178,7 @@ export class AgentHarness {
      * @internal
      */
     stepLedgerRetention;
-    initialStep;
+    agentGraph;
     /** Harness-wide tool pool merged into every context's `unifiedTools`. */
     harnessTools;
     /** @internal Context layers configured for this harness. Exposed non-private
@@ -191,15 +191,55 @@ export class AgentHarness {
     rootRunDepth = new Map();
     callModelOverride;
     defaultDeliveryMode;
-    streamIdleTimeoutMs;
     sessions = new Map();
     /**
-     * @internal Cross-step harness sessions keyed by `step.session.reuse`, kept
-     * alive for the life of this harness so reused coding-agent sessions span
-     * multiple steps. Reached by the interpreter's harness handler via
-     * `frameworkCast`; do not access from outside core.
+     * @internal Live ACP connections. Reached by the interpreter via
+     * `frameworkCast`; use {@link listAcpSessions}, {@link getAcpSession},
+     * {@link cancelAcpSession}, and {@link closeAcpSessions} from outside core.
      */
-    subHarnessSessions = new Map();
+    acpSessions = new AcpSessionStore();
+    /**
+     * Snapshot of every live ACP sub-agent, for a UI that wants to show what is
+     * running and let a user act on it. Read-only: a session's turns are driven
+     * by steps, so nothing here can start work behind the runtime's back.
+     * @public
+     */
+    listAcpSessions() {
+        return this.acpSessions.list();
+    }
+    /**
+     * The live connection + session behind a handle from {@link listAcpSessions}
+     * (or a step's `session.reuse` key), or `undefined` when nothing matches.
+     * @public
+     */
+    getAcpSession(key) {
+        return this.acpSessions.get(key);
+    }
+    /**
+     * Interrupt whatever the sub-agent under `key` is doing, via `session/cancel`.
+     * The connection stays open and the turn still returns — with the `cancelled`
+     * stop reason, which the running step surfaces as a `cancelled` error.
+     * Returns false when nothing matches the key.
+     * @public
+     */
+    async cancelAcpSession(key) {
+        const entry = this.acpSessions.get(key);
+        if (!entry) {
+            return false;
+        }
+        await entry.session.cancel();
+        return true;
+    }
+    /**
+     * Close every ACP connection this harness still holds, whatever its
+     * keep-alive scope. A connection owns a live agent — usually a child process
+     * whose stdio keeps the event loop alive — so a `keepAlive: 'harness'`
+     * session must be closed by its owner or the host will not exit. Idempotent.
+     * @public
+     */
+    async closeAcpSessions() {
+        await this.acpSessions.closeAll();
+    }
     layerStateStore;
     /** Per-harness memoization cache for `recallMode: 'eventual'` layers. */
     recallCache;
@@ -225,45 +265,47 @@ export class AgentHarness {
     itemSchemas;
     constructor(opts) {
         const validatedParams = opts.paramsSchema ? opts.paramsSchema.parse(opts.params) : opts.params;
+        const environment = opts.environment;
+        const storageEnv = environment?.storage;
         this.config = {
             name: opts.name,
-            storage: opts.storage ?? createInMemoryStorage(),
+            storage: storageEnv?.adapter ?? createInMemoryStorage(),
             hooks: opts.hooks,
             params: validatedParams,
             itemSchemas: opts.itemSchemas,
-            strictItemSchemas: opts.strictItemSchemas ?? true,
             projection: opts.projection,
             forceAtomicRecall: opts.forceAtomicRecall,
             contextCache: opts.contextCache,
         };
-        this.fs = opts.fs ?? createInMemoryFsAdapter();
-        this.shell = opts.shell ?? createInMemoryShellAdapter();
-        this.subprocess = opts.subprocess ?? createInMemorySubprocessAdapter();
-        this.checkpointStore = opts.checkpointStore;
-        this.stepLedgerStore = opts.checkpointStore
+        this.fs = environment?.fs ?? createInMemoryFsAdapter();
+        this.shell = environment?.shell ?? createInMemoryShellAdapter();
+        this.subprocess = environment?.subprocess ?? createInMemorySubprocessAdapter();
+        this.checkpointStore = storageEnv?.checkpointStore;
+        this.stepLedgerStore = storageEnv?.checkpointStore
             ? createStepLedgerStore({
                 storage: this.config.storage ?? createInMemoryStorage(),
             })
             : undefined;
-        this.stepLedgerRetention = resolveStepLedgerRetention(opts.stepLedgerRetention);
-        this.initialStep = opts.initialStep;
+        this.stepLedgerRetention = resolveStepLedgerRetention(storageEnv?.stepLedgerRetention);
+        this.agentGraph = opts.agentGraph;
         this.harnessTools = opts.tools ?? [];
-        this._contextLayers = resolveContextOption(opts);
+        this._contextLayers = opts.contextLayers;
         this.callModelOverride = opts._testCallModel;
-        this.client = opts._testCallModel ? undefined : createClient(opts.llm, opts.contextCache);
+        this.client = opts._testCallModel
+            ? undefined
+            : createClient(opts.callModelDefaults, opts.contextCache);
         this.channelStore = new ChannelStore();
         this.traceExporter = opts.traceExporter ?? new NoopExporter();
         this.layerStateStore =
-            opts.layerStateStore ??
+            storageEnv?.layerStateStore ??
                 createLayerStateStore((layerId, hook, error) => {
                     console.warn(`[noetic] context layer '${layerId}' ${hook} error:`, error);
                 });
         this.recallCache = createRecallCache();
         this.contextCache = createContextCacheStore();
         this.defaultDeliveryMode = opts.defaultDeliveryMode ?? 'next-turn';
-        this.streamIdleTimeoutMs = opts.streamIdleTimeoutMs ?? DEFAULT_STREAM_IDLE_TIMEOUT_MS;
-        this.itemSchemas = new ItemSchemaRegistry(opts.itemSchemas, {
-            strictUnknownExtensions: opts.strictItemSchemas ?? true,
+        this.itemSchemas = new ItemSchemaRegistry(opts.itemSchemas?.schemas, {
+            strictUnknownExtensions: opts.itemSchemas?.strict ?? true,
         });
         this.rootCwdState = {
             cwd: opts.initialCwd ?? '/',
@@ -284,11 +326,11 @@ export class AgentHarness {
     }
     //#region Session Accessors
     execute(input, options) {
-        if (!this.initialStep) {
+        if (!this.agentGraph) {
             return Promise.reject(new NoeticConfigError({
                 code: 'NO_STEP_CONFIGURED',
-                message: 'No initialStep configured on this harness.',
-                hint: 'Pass `initialStep` in constructor options, or use run() directly.',
+                message: 'No agentGraph configured on this harness.',
+                hint: 'Pass `agentGraph` in constructor options, or use run() directly.',
             }));
         }
         const threadId = options?.threadId ?? DEFAULT_THREAD_ID;
@@ -353,6 +395,17 @@ export class AgentHarness {
         const session = this.sessions.get(threadId);
         return session ? session.runner.queue.size : 0;
     }
+    getUsage(scope) {
+        const threadId = scope?.threadId ?? DEFAULT_THREAD_ID;
+        const session = this.sessions.get(threadId);
+        if (!session) {
+            return {
+                inputTokens: 0,
+                outputTokens: 0,
+            };
+        }
+        return session.runner.getUsage();
+    }
     seedSessionHistory(threadId, items) {
         const session = this.getOrCreateSession(threadId);
         session.accumulatedItems = [
@@ -385,15 +438,15 @@ export class AgentHarness {
                         threadId,
                         resourceId: perTurnOptions.resourceId,
                         state: perTurnOptions.state,
-                        context: resolveContextOption(perTurnOptions),
+                        contextLayers: perTurnOptions.contextLayers,
                         _broadcaster: session.runner.broadcaster,
                     });
                     const ext = frameworkCast(ctx);
                     ext._sessionQueue = session.runner.queue;
                     ext._sessionBetweenRounds = true;
                     ext._sessionRunnerAgentName = this.config.name;
-                    if (this.initialStep || this.harnessTools.length > 0) {
-                        const stepTools = this.initialStep ? collectAllTools(this.initialStep) : [];
+                    if (this.agentGraph || this.harnessTools.length > 0) {
+                        const stepTools = this.agentGraph ? collectAllTools(this.agentGraph) : [];
                         this.setUnifiedTools(ctx, [
                             ...stepTools,
                             ...this.harnessTools,
@@ -402,11 +455,11 @@ export class AgentHarness {
                     return ctx;
                 },
                 runTurn: async (ctx, _turn, signal) => {
-                    if (!this.initialStep) {
+                    if (!this.agentGraph) {
                         throw new NoeticConfigError({
                             code: 'NO_STEP_CONFIGURED',
-                            message: 'No initialStep configured on this harness.',
-                            hint: 'Pass `initialStep` in constructor options.',
+                            message: 'No agentGraph configured on this harness.',
+                            hint: 'Pass `agentGraph` in constructor options.',
                         });
                     }
                     // Wire signal-abort to context-abort so the interpreter bails cleanly.
@@ -420,7 +473,7 @@ export class AgentHarness {
                             once: true,
                         });
                     }
-                    const result = await this.initAndRun(this.initialStep, '', ctx);
+                    const result = await this.initAndRun(this.agentGraph, '', ctx);
                     // Snapshot final items into session history for the next turn.
                     session.accumulatedItems = [
                         ...ctx.itemLog.items,
@@ -450,7 +503,7 @@ export class AgentHarness {
             itemSchemas: this.itemSchemas,
             client: this.client,
             callModelOverride: this.callModelOverride,
-            streamIdleTimeoutMs: this.streamIdleTimeoutMs,
+            streamIdleTimeoutMs: DEFAULT_STREAM_IDLE_TIMEOUT_MS,
             harness: this,
             traceExporter: this.traceExporter,
         }).callModel(request);
@@ -509,6 +562,7 @@ export class AgentHarness {
             else {
                 this.rootRunDepth.delete(ctx.id);
                 this.channelStore.closeExecution(ctx.id);
+                await this.acpSessions.closeOwnedBy(ctx.id);
             }
         }
     }
@@ -540,10 +594,10 @@ export class AgentHarness {
     }
     createContext(opts) {
         const resolved = opts ?? {};
-        // `ContextImpl` takes the layers as `layers`, so both option spellings are
+        // `ContextImpl` takes the layers as `layers`, so the option spelling is
         // dropped from `rest` rather than spread through under the wrong name.
-        const { context: _context, memory: _memory, cwdInit, ...rest } = resolved;
-        const effectiveLayers = resolveContextOption(resolved) ?? this._contextLayers;
+        const { contextLayers, cwdInit, ...rest } = resolved;
+        const effectiveLayers = contextLayers ?? this._contextLayers;
         const itemSchemas = buildItemSchemaRegistry({
             base: this.itemSchemas,
             layers: effectiveLayers,
@@ -696,7 +750,7 @@ export class AgentHarness {
         const ctx = this.createContext({
             items: historyItems,
             threadId,
-            context: this._contextLayers,
+            contextLayers: this._contextLayers,
         });
         const layers = ctx.layers ?? [];
         if (layers.length === 0) {
@@ -837,7 +891,7 @@ export class AgentHarness {
      *
      * Aborting a context rejects everything blocked on it (channel `recv`
      * waiters, parked senders) with `{ kind: 'cancelled' }`, stops its in-flight
-     * model call and sub-harness turn, and makes the next step boundary throw
+     * model call and ACP agent turn, and makes the next step boundary throw
      * `cancelled`.
      *
      * Cancellation is cooperative: a layer hook already in flight is allowed to

package/dist/harness/deps/types.d.ts

@@ -1,4 +1,4 @@
 export type { ContextCacheConfig, ContextCacheStore, ContextData, ContextLayer, ExecutionContext, ProjectionPolicy, StorageAdapter, } from '@noetic-tools/context';
-export type { AgentConfig, AgentHarnessContract, AgentHooks, CallModelRequest, Channel, ChannelHandle, Context, CwdState, DeliveryMode, DetachedHandle, ExecuteInput, ExecuteOptions, ExternalChannel, FsAdapter, HarnessResponse, HarnessStatus, Item, ItemSchemaExtensions, LLMResponse, LlmProviderConfig, RecallLayerOutput, SessionScope, ShellAdapter, Span, SteeringDecision, Step, StreamEvent, StreamingItem, SubHarnessSession, SubprocessAdapter, Tool, TraceExporter, } from '@noetic-tools/types';
+export type { AcpKeepAlive, AcpLiveSession, AcpSessionInfo, AgentConfig, AgentHarnessContract, AgentHooks, CallModelRequest, Channel, ChannelHandle, Context, CwdState, DeliveryMode, DetachedHandle, ExecuteInput, ExecuteOptions, ExternalChannel, FsAdapter, HarnessResponse, HarnessStatus, Item, ItemSchemaConfig, ItemSchemaExtensions, LLMResponse, LlmProviderConfig, RecallLayerOutput, SessionScope, SessionUsage, ShellAdapter, Span, SteeringDecision, Step, StreamEvent, StreamingItem, SubprocessAdapter, Tool, TraceExporter, } from '@noetic-tools/types';
 export { SteeringAction } from '@noetic-tools/types';
 export type { ZodType } from 'zod';

package/dist/harness/model-call.d.ts

@@ -10,7 +10,7 @@ interface StreamIdleWatchdog {
 /** @internal Create a watchdog that aborts `controller` when no
  *  {@link StreamIdleWatchdog.reset reset} is called within `timeoutMs`. When
  *  `timeoutMs <= 0`, returns an inert no-op so callers can always call `.reset()`
- *  / `.stop()` without a branch. Starts armed: the caller is responsible for
+ *  / `.stop()` without branching. Starts armed: the caller is responsible for
  *  `.stop()` in a `finally`. `onTimeout` runs before the abort so observers can
  *  emit a framework event with the original cause. Exported only for unit tests. */
 export declare function createStreamIdleWatchdog(timeoutMs: number, controller: AbortController, onTimeout?: () => void): StreamIdleWatchdog;

package/dist/harness/model-call.js

@@ -1,7 +1,6 @@
 import { frameworkCast, NoeticConfigError } from '@noetic-tools/types';
 import { serverTool } from '@openrouter/agent';
-import { z } from 'zod';
-import { convertTools, executeToolCall, extractOutputItems, extractSystemInstruction, extractUsage, itemsToInput, sanitizeToolNameForWire, } from '../adapters/openrouter.js';
+import { convertTools, executeToolCall, extractOutputItems, extractSystemInstruction, extractUsage, itemsToInput, resolveWireJsonSchema, sanitizeToolNameForWire, } from '../adapters/openrouter.js';
 import { isFunctionCall } from '../interpreter/typeguards.js';
 import { GenAI, NoeticAttr, ToolAttr } from '../observability/genai-attributes.js';
 import { emitFrameworkEvent, getBroadcaster, shouldEmit } from '../runtime/broadcaster-utils.js';
@@ -9,8 +8,12 @@ import { buildItemSchemaRegistry, createToolResultItem } from './model-schema.js
 const MAX_TOOL_ROUNDS = 32;
 const MAX_RECOVERY_CONTINUATIONS = 3;
 const EPHEMERAL_CONTINUE_INPUT = 'continue';
-function buildTextFormat(schema) {
-    const jsonSchema = z.toJSONSchema(schema);
+function buildTextFormat(schema, explicitJsonSchema) {
+    const jsonSchema = resolveWireJsonSchema({
+        schema,
+        explicitJsonSchema,
+        what: 'structured output schema',
+    });
     return {
         format: {
             type: 'json_schema',
@@ -133,7 +136,7 @@ async function pipeStreamEventsToBroadcaster(opts) {
 /** @internal Create a watchdog that aborts `controller` when no
  *  {@link StreamIdleWatchdog.reset reset} is called within `timeoutMs`. When
  *  `timeoutMs <= 0`, returns an inert no-op so callers can always call `.reset()`
- *  / `.stop()` without a branch. Starts armed: the caller is responsible for
+ *  / `.stop()` without branching. Starts armed: the caller is responsible for
  *  `.stop()` in a `finally`. `onTimeout` runs before the abort so observers can
  *  emit a framework event with the original cause. Exported only for unit tests. */
 export function createStreamIdleWatchdog(timeoutMs, controller, onTimeout) {
@@ -150,7 +153,7 @@ export function createStreamIdleWatchdog(timeoutMs, controller, onTimeout) {
             if (stopped) {
                 return;
             }
-            const reason = new Error(`llm stream idle timeout after ${timeoutMs}ms`);
+            const reason = new Error(`model stream idle timeout after ${timeoutMs}ms`);
             onTimeout?.();
             controller.abort(reason);
         }, timeoutMs);
@@ -354,7 +357,7 @@ export class AgentHarnessModelCaller {
         throw new NoeticConfigError({
             code: 'NO_LLM_PROVIDER',
             message: 'No LLM provider configured on this harness.',
-            hint: 'Pass `llm: { apiKey: "..." }` (defaults to the Noetic platform) or set NOETIC_API_KEY. For direct OpenRouter, use `llm: { provider: "openrouter", apiKey: "..." }` or set OPENROUTER_API_KEY.',
+            hint: 'Pass `callModelDefaults: { apiKey: "..." }` (defaults to the Noetic platform) or set NOETIC_API_KEY. For direct OpenRouter, use `callModelDefaults: { provider: "openrouter", apiKey: "..." }` or set OPENROUTER_API_KEY.',
         });
     }
     async callModel(request) {
@@ -384,7 +387,9 @@ export class AgentHarnessModelCaller {
         const rounds = [];
         let totalCost = 0;
         const conversationInput = itemsToInput(prepared.remaining);
-        const textFormat = request.outputSchema ? buildTextFormat(request.outputSchema) : undefined;
+        const textFormat = request.outputSchema
+            ? buildTextFormat(request.outputSchema, request.outputJsonSchema)
+            : undefined;
         let round = 0;
         let invalidRecoveryContinuations = 0;
         let toolLimitRecoveryContinuations = 0;
@@ -406,7 +411,7 @@ export class AgentHarnessModelCaller {
             let idleStalled = false;
             const watchdog = createStreamIdleWatchdog(this.opts.streamIdleTimeoutMs, roundController, () => {
                 idleStalled = true;
-                prepared.emitIfAllowed('llm_call_stalled', {
+                prepared.emitIfAllowed('model_call_stalled', {
                     round,
                     idleTimeoutMs: this.opts.streamIdleTimeoutMs,
                 });
@@ -414,7 +419,7 @@ export class AgentHarnessModelCaller {
             const modelInput = recoveryContinuation
                 ? withEphemeralContinueInput(conversationInput)
                 : frameworkCast(conversationInput);
-            prepared.emitIfAllowed('llm_call_started', {
+            prepared.emitIfAllowed('model_call_started', {
                 round,
                 messageCount: modelInput.length,
                 toolCount: prepared.sdkTools?.length ?? 0,
@@ -446,7 +451,7 @@ export class AgentHarnessModelCaller {
                     watchdog.reset();
                     if (!firstEventSeen) {
                         firstEventSeen = true;
-                        prepared.emitIfAllowed('llm_call_first_event', {
+                        prepared.emitIfAllowed('model_call_first_event', {
                             round,
                         });
                     }
@@ -462,7 +467,7 @@ export class AgentHarnessModelCaller {
                 if (idleStalled) {
                     throw roundSignal.reason instanceof Error
                         ? roundSignal.reason
-                        : new Error(`llm stream idle timeout after ${this.opts.streamIdleTimeoutMs}ms`);
+                        : new Error(`model stream idle timeout after ${this.opts.streamIdleTimeoutMs}ms`);
                 }
                 if (request.signal?.aborted) {
                     break;
@@ -507,7 +512,7 @@ export class AgentHarnessModelCaller {
                 continue;
             }
             invalidRecoveryContinuations = 0;
-            prepared.emitIfAllowed('llm_call_completed', {
+            prepared.emitIfAllowed('model_call_completed', {
                 round,
                 itemCount: sdkResponse.output?.length ?? 0,
             });
@@ -605,7 +610,7 @@ export class AgentHarnessModelCaller {
                 kind: 'ok',
             };
         }
-        params.emitIfAllowed('llm_call_failed', {
+        params.emitIfAllowed('model_call_failed', {
             round: params.round,
             status: terminalError.status,
             error: terminalError.message,
@@ -615,7 +620,7 @@ export class AgentHarnessModelCaller {
             throw new Error(terminalError.message);
         }
         const next = params.invalidRecoveryContinuations + 1;
-        params.emitIfAllowed('llm_call_recovery_continue', {
+        params.emitIfAllowed('model_call_recovery_continue', {
             round: params.round,
             status: terminalError.status,
             attempt: next,
@@ -633,7 +638,7 @@ export class AgentHarnessModelCaller {
             };
         }
         const message = 'LLM response completed with no output items';
-        params.emitIfAllowed('llm_call_failed', {
+        params.emitIfAllowed('model_call_failed', {
             round: params.round,
             status: 'completed',
             error: message,
@@ -643,7 +648,7 @@ export class AgentHarnessModelCaller {
             throw new Error(message);
         }
         const next = params.invalidRecoveryContinuations + 1;
-        params.emitIfAllowed('llm_call_recovery_continue', {
+        params.emitIfAllowed('model_call_recovery_continue', {
             round: params.round,
             status: 'completed',
             attempt: next,
@@ -750,7 +755,7 @@ export class AgentHarnessModelCaller {
     recordMalformedToolCall(params) {
         const { fc, request, allItems, conversationInput } = params;
         const errorOutput = `Error: malformed JSON in tool arguments: ${fc.arguments}`;
-        const toolForCall = request.tools?.find((tool) => tool.name === fc.name);
+        const toolForCall = request.tools?.find((tool) => tool.name === fc.name || sanitizeToolNameForWire(tool.name) === fc.name);
         // Owner-scoped result validation (see executeFunctionCall).
         const outputItem = createToolResultItem({
             output: errorOutput,