@tangle-network/agent-runtime 0.273.5 → 0.277.0

package/dist/{substrate-VKjmrksz.d.ts → substrate-CfE3oUB9.d.ts}

@@ -1,4 +1,4 @@
-import { Ca as OutputAdapter, Oa as Validator, ea as AgentRunSpec, na as Driver } from "./types-DMRc3jru.js";
+import { Ca as OutputAdapter, Oa as Validator, ea as AgentRunSpec, na as Driver } from "./types-CI7ve2sA.js";
 import { AgentProfile } from "@tangle-network/agent-interface";
 //#region src/profiles/coder.d.ts
 /**
@@ -281,4 +281,4 @@ interface UiFinding {
 }
 //#endregion
 export { CoderTask as _, UiFindingSeverity as a, KnowledgeUpdate as c, ResearchSource as d, ResearchTask as f, researcherProfile as g, multiHarnessResearcherFanout as h, UiFindingScreenshot as i, MultiHarnessResearcherFanoutOptions as l, createResearcherValidator as m, UI_LENSES as n, UiLens as o, ResearcherProfileOptions as p, UiFinding as r, KnowledgeItem as s, UI_FINDING_SEVERITIES as t, ResearchOutput as u, coderTaskToPrompt as v };
-//# sourceMappingURL=substrate-VKjmrksz.d.ts.map
+//# sourceMappingURL=substrate-CfE3oUB9.d.ts.map

package/dist/{supervise-Dej5aSDa.js → supervise-BoCOUAmr.js}

@@ -1,15 +1,16 @@
-import { $ as bindScopeRetainedOwnerProvider, $i as detachedSnapshot, A as driverExecutorFactory, Aa as resourceTelemetry, Ai as readCommittedJsonLines, At as snapshotExecutorConfig, B as recordScopeOwnerMaterialization, Bi as providerAttemptEvidence, C as createContinuationKeeper, Da as withTimeout, Di as isNoEntError, Dn as executorFailure, Dr as persistHarnessTranscript, Ea as unmeteredSpend, Ei as worktreeCliProfileMaterialization, F as beginScopeOwnerAttempt, Fa as DEFAULT_SUCCESSFUL_SHUTDOWN_MS, Fn as bridgeStopSignalKey, Gi as runtimeOwnedExecutorProviderEvidence, H as restoreScopeOwnerAcceptedExecution, Hi as runtimeOwnedDriveHarnessProviderEvidence, Hr as generateSpanId, Ii as inheritRuntimeOwnedExecutorAttestation, Ki as runtimeOwnedPendingExecutorMaterialization, Kn as routerBrain, Kr as toOtelAttributes, L as deriveNodeExecutionIdentity, La as teardownExecutor, Ni as attestRuntimeOwnedScopeOwner, Oa as zeroSpend, On as bridgeAdmissionRefusal, Or as readHarnessTranscript, Ot as createExecutor, P as assertRecursiveReservationPolicy, Pn as bridgeRuntimeAttachmentsKey, Pt as createRouterTranscript, Q as bindScopeRetainedOwnerEnvironmentId, R as meterRuntimeOwnedAccounting, Ra as teardownSurfaces, Sa as sleep, Si as renderUnsupported, Tt as captureReusableExecutorConfig, U as scopeOwnerExecutorNodeContext, Ui as runtimeOwnedExecutorExecutionBinding, V as recordScopeOwnerPause, Vi as recordRuntimeOwnedDriveHarnessProviderEvidence, Wi as runtimeOwnedExecutorMaterialization, _ as CheckUnavailableError, _i as promptControlProfileMaterialization, at as scopeRetainedOwnerResourceReader, b as bestComposite, br as linkAbort, c as runTree, cr as registerRetainedExecutorPreparation, et as consumeScopeRetainedOwnerResult, f as runDriverWithRetry, fi as assertProfileMaterialization, gi as profileMaterializationAxes$1, gn as createInbox, h as CONTINUATIONS_DIR, hi as fullProfileMaterialization, i as bestDelivered, in as newUsageTotals, it as scopeRetainedOwnerPriorSpend, j as isDriverSpec, ja as withBudgetResources, k as driverChild, ka as addResourceSpend, ki as prepareJsonlAppend, kn as bridgeModelRouteRefusal, kr as contentAddress, lr as retainedExecutorSeamKey, ma as addSpend, mi as defineProfileMaterializationContract, n as createRootHandle, nt as reconcileScopeRetainedOwnerEnvironment, on as spendFromUsageTotals, ot as scopeRetainedOwnerResult, p as summarizeDriverAttempts, pi as controlProfileMaterialization, qi as runtimeOwnedScopeOwnerRuntime, r as createSupervisor, rn as meterUsageEvent, rt as scopeRetainedOwnerContext, s as runFinalizer, sn as WORKER_TRACE_PROPAGATION, tn as assertValidBudget, tr as resolveAgentEnvironmentProvider, tt as prepareScopeRetainedOwnerTask, u as HarnessTurnFailedError, ur as errMessage, v as admitContinuationPolicy, vi as promptModelProfileMaterialization, vr as RunCancellationReason, wi as unsupportedProfileDimensions, wt as bindReusableExecutorExecutionId, x as checkVerdictOf, xr as runAbortable, z as meterRuntimeOwnedProviderAttempt, zr as createOtelExporter } from "./supervisor-DdKPGluv.js";
+import { $ as bindScopeRetainedOwnerProvider, $i as detachedSnapshot, A as driverExecutorFactory, Aa as resourceTelemetry, Ai as readCommittedJsonLines, At as snapshotExecutorConfig, B as recordScopeOwnerMaterialization, Bi as providerAttemptEvidence, C as createContinuationKeeper, Da as withTimeout, Di as isNoEntError, Dn as executorFailure, Dr as persistHarnessTranscript, Ea as unmeteredSpend, Ei as worktreeCliProfileMaterialization, F as beginScopeOwnerAttempt, Fa as DEFAULT_SUCCESSFUL_SHUTDOWN_MS, Fn as bridgeStopSignalKey, Gi as runtimeOwnedExecutorProviderEvidence, H as restoreScopeOwnerAcceptedExecution, Hi as runtimeOwnedDriveHarnessProviderEvidence, Hr as generateSpanId, Ii as inheritRuntimeOwnedExecutorAttestation, Ki as runtimeOwnedPendingExecutorMaterialization, Kn as routerBrain, Kr as toOtelAttributes, L as deriveNodeExecutionIdentity, La as teardownExecutor, Ni as attestRuntimeOwnedScopeOwner, Oa as zeroSpend, On as bridgeAdmissionRefusal, Or as readHarnessTranscript, Ot as createExecutor, P as assertRecursiveReservationPolicy, Pn as bridgeRuntimeAttachmentsKey, Pt as createRouterTranscript, Q as bindScopeRetainedOwnerEnvironmentId, R as meterRuntimeOwnedAccounting, Ra as teardownSurfaces, Sa as sleep, Si as renderUnsupported, Tt as captureReusableExecutorConfig, U as scopeOwnerExecutorNodeContext, Ui as runtimeOwnedExecutorExecutionBinding, V as recordScopeOwnerPause, Vi as recordRuntimeOwnedDriveHarnessProviderEvidence, Wi as runtimeOwnedExecutorMaterialization, Zt as parseWorkerToolTraceArtifact, _ as CheckUnavailableError, _i as promptControlProfileMaterialization, at as scopeRetainedOwnerResourceReader, b as bestComposite, br as linkAbort, c as runTree, cr as registerRetainedExecutorPreparation, et as consumeScopeRetainedOwnerResult, f as runDriverWithRetry, fi as assertProfileMaterialization, gi as profileMaterializationAxes$1, gn as createInbox, h as CONTINUATIONS_DIR, hi as fullProfileMaterialization, i as bestDelivered, in as newUsageTotals, it as scopeRetainedOwnerPriorSpend, j as isDriverSpec, ja as withBudgetResources, k as driverChild, ka as addResourceSpend, ki as prepareJsonlAppend, kn as bridgeModelRouteRefusal, kr as contentAddress, lr as retainedExecutorSeamKey, ma as addSpend, mi as defineProfileMaterializationContract, n as createRootHandle, nt as reconcileScopeRetainedOwnerEnvironment, on as spendFromUsageTotals, ot as scopeRetainedOwnerResult, p as summarizeDriverAttempts, pi as controlProfileMaterialization, qi as runtimeOwnedScopeOwnerRuntime, r as createSupervisor, rn as meterUsageEvent, rt as scopeRetainedOwnerContext, s as runFinalizer, sn as WORKER_TRACE_PROPAGATION, tn as assertValidBudget, tr as resolveAgentEnvironmentProvider, tt as prepareScopeRetainedOwnerTask, u as HarnessTurnFailedError, ur as errMessage, v as admitContinuationPolicy, vi as promptModelProfileMaterialization, vr as RunCancellationReason, wi as unsupportedProfileDimensions, wt as bindReusableExecutorExecutionId, x as checkVerdictOf, xr as runAbortable, z as meterRuntimeOwnedProviderAttempt, zr as createOtelExporter } from "./supervisor-CjNbOcLO.js";
 import { f as RuntimeRunStateError, i as ConfigError, m as ValidationError } from "./errors-DodWX-cb.js";
 import { a as concreteProfileModel, c as profileModelExecutionSettings, d as agentHarness, f as harnessRunsAgent, n as assertModelAllowed, o as enforceTokenLimits, r as assertProfileModelsAllowed, s as profileBridgeWireModel, t as assertExecutableAgentProfile } from "./model-policy-BbSCSak0.js";
 import { t as composeRuntimeHooks } from "./runtime-hooks-tXpAarhW.js";
 import { n as assertNoSymlinkDescendant } from "./durable-file-DWQA4ooo.js";
-import { C as applyRunCancellation, E as DEFAULT_AWAIT_EVENT_TIMEOUT_MS, b as withRunContext, c as createProgressTracker, d as progressStop, j as createCoordinationToolsForManager, k as coordinationVerbNames, n as createSteerAcknowledger, r as driverAgent, t as createCancelAcknowledger, u as plateau, v as createFileRunContext, w as watchRunCancellation, y as createInMemoryRunContext } from "./coordination-driver-CkcW4dxU.js";
+import { C as applyRunCancellation, E as DEFAULT_AWAIT_EVENT_TIMEOUT_MS, b as withRunContext, c as createProgressTracker, d as progressStop, j as createCoordinationToolsForManager, k as coordinationVerbNames, n as createSteerAcknowledger, r as driverAgent, t as createCancelAcknowledger, u as plateau, v as createFileRunContext, w as watchRunCancellation, y as createInMemoryRunContext } from "./coordination-driver-ZfUf-H4H.js";
 import { r as createStdioToolServer, t as PROTOCOL_VERSION } from "./tool-server-Dqer2M4_.js";
 import { k as writeRunCancellation, o as readRunCancelRequest, s as readRunCancellation } from "./run-layout-vwd4SwX_.js";
 import { agentProfileSchema, canonicalAgentProfileDigest, canonicalCandidateDigest, canonicalCandidateJson, validateAgentProfileSecurity } from "@tangle-network/agent-interface";
+import { toolSpansToTraceAnalysisStore } from "@tangle-network/agent-eval";
 import { createHash, createHmac, randomBytes, randomUUID, timingSafeEqual } from "node:crypto";
 import { closeSync, constants, fsyncSync, openSync, writeSync } from "node:fs";
-import { resolve } from "node:path";
+import { dirname, resolve } from "node:path";
 import { mkdir } from "node:fs/promises";
 import { isMaterializerHarness } from "@tangle-network/agent-profile-materialize";
 import { ATTR } from "@tangle-network/agent-trace-contract";
@@ -1516,6 +1517,139 @@ function errorMessage(error) {
 	return error instanceof Error ? error.message : String(error);
 }
 //#endregion
+//#region src/runtime/supervise/run-traces.ts
+/**
+* A manager's own run as one trace store: what the continuation's question panel reads.
+*
+* The panel asks about the director's own work and about its workers' (discovery
+* `docs/38-one-loop-and-continuation.md`, section 5), so it reads both: the director's root stream
+* (`root-stream.jsonl`) and each settled worker's structured tool trace. Each agent is one trace,
+* named by its node id, so a `{worker}` question and a `trace://<worker id>/<span id>` citation
+* name the same worker.
+*
+* The root stream is progress events, not spans. Each tool call becomes one span with its result,
+* and each run of the director's text, and of its reasoning, between two other events becomes one
+* span named for what it is (`assistant_text`, `assistant_reasoning`): the director's claims live in
+* its text, and a panel that saw only its tool calls could not check them. Span ids are the
+* record's `seq`, which is unique in the file and stable across reads.
+*/
+/** The span a run of the director's text becomes. */
+const ASSISTANT_TEXT_SPAN = "assistant_text";
+/** The span a run of the director's reasoning becomes. */
+const ASSISTANT_REASONING_SPAN = "assistant_reasoning";
+/** The director's root stream as tool spans in one trace named `traceId`. */
+function rootStreamToolSpans(records, traceId) {
+	const spans = [];
+	const open = /* @__PURE__ */ new Map();
+	let prose;
+	const at = (record) => {
+		const ms = Date.parse(record.at);
+		if (!Number.isFinite(ms)) throw new ValidationError(`root stream line ${record.seq} has no valid time`);
+		return ms;
+	};
+	const base = (record, name) => ({
+		kind: "tool",
+		spanId: `s${record.seq}`,
+		runId: traceId,
+		name,
+		toolName: name,
+		startedAt: at(record),
+		attributes: {
+			attempt: record.attempt,
+			seq: record.seq
+		}
+	});
+	for (const record of records) {
+		if (!("event" in record)) {
+			prose = void 0;
+			continue;
+		}
+		const { event } = record;
+		if (event.kind === "text_delta" || event.kind === "reasoning_delta") {
+			if (prose !== void 0 && prose.kind === event.kind && prose.attempt === record.attempt) {
+				prose.text += event.text;
+				prose.span.result = prose.text;
+				prose.span.endedAt = at(record);
+				continue;
+			}
+			const span = {
+				...base(record, event.kind === "text_delta" ? ASSISTANT_TEXT_SPAN : ASSISTANT_REASONING_SPAN),
+				args: {},
+				result: event.text,
+				endedAt: at(record)
+			};
+			spans.push(span);
+			prose = {
+				span,
+				text: event.text,
+				kind: event.kind,
+				attempt: record.attempt
+			};
+			continue;
+		}
+		prose = void 0;
+		if (event.kind === "tool_call") {
+			const span = {
+				...base(record, event.toolName),
+				args: event.args,
+				...event.args === void 0 ? { argsCaptured: false } : {},
+				endedAt: at(record)
+			};
+			if (event.toolCallId !== void 0) span.attributes = {
+				...span.attributes,
+				toolCallId: event.toolCallId
+			};
+			spans.push(span);
+			open.set(`${record.attempt}\u0000${event.toolCallId ?? `name:${event.toolName}`}`, span);
+			continue;
+		}
+		if (event.kind === "tool_result") {
+			const key = `${record.attempt}\u0000${event.toolCallId ?? `name:${event.toolName}`}`;
+			const call = open.get(key);
+			if (call !== void 0) {
+				open.delete(key);
+				call.result = event.result;
+				call.endedAt = at(record);
+				continue;
+			}
+			spans.push({
+				...base(record, event.toolName),
+				args: void 0,
+				argsCaptured: false,
+				result: event.result,
+				endedAt: at(record)
+			});
+		}
+	}
+	return spans;
+}
+/** One worker's persisted tool spans, re-keyed to the worker's node id. Unavailable: none. */
+async function workerSpans(worker, blobs) {
+	if (worker.trace.status !== "available") return [];
+	const raw = await blobs.get(worker.trace.traceRef);
+	if (raw === void 0) throw new ValidationError(`worker ${worker.id}'s trace blob '${worker.trace.traceRef}' is missing; its settlement evidence is incomplete`);
+	const { spans } = parseWorkerToolTraceArtifact(raw, worker.trace.traceRef);
+	return spans.map((span) => ({
+		...span,
+		runId: worker.id,
+		...span.runId === worker.id ? {} : { attributes: {
+			...span.attributes,
+			sourceRunId: span.runId
+		} }
+	}));
+}
+/**
+* The store over a manager's run: its own root stream when it has one, and every settled worker's
+* tool trace. `undefined` when the run has recorded no span yet, which is not a store a panel can
+* read: agent-eval refuses an empty trace, because it cannot tell a tool-free run from a failed
+* capture.
+*/
+async function runTraceAnalysisStore(input) {
+	const records = input.root === void 0 ? void 0 : await input.root.read();
+	const spans = [...records === void 0 || input.root === void 0 ? [] : rootStreamToolSpans(records, input.root.id), ...(await Promise.all(input.workers.map((worker) => workerSpans(worker, input.blobs)))).flat()];
+	return spans.length === 0 ? void 0 : toolSpansToTraceAnalysisStore(spans);
+}
+//#endregion
 //#region src/runtime/supervise/worker-control-observer.ts
 /** Observe durable worker controls while the native harness owns the manager's turn loop. */
 function observeWorkerControls(options) {
@@ -1999,7 +2133,17 @@ function buildSupervisorAgent(profile, deps, testBrain) {
 						}));
 					},
 					...deps.continuationDir === void 0 ? {} : { dir: deps.continuationDir },
-					...deps.rootStreamPath === void 0 ? {} : { rootStreamPath: deps.rootStreamPath },
+					traces: () => {
+						const rootStreamPath = deps.rootStreamPath;
+						return runTraceAnalysisStore({
+							...rootStreamPath === void 0 ? {} : { root: {
+								id: scope.view.root,
+								read: () => readRootStream(dirname(rootStreamPath))
+							} },
+							workers: mcp.settled(),
+							blobs: deps.blobs
+						});
+					},
 					canReadMore: runtimeToolNames.includes("read_continuation")
 				});
 				let candidate;
@@ -4358,4 +4502,4 @@ function rootProviderModelEvidenceFromExecution(evidence) {
 //#endregion
 export { gateOnDeliverable as C, serveCoordinationMcp as S, readRootStream as _, superviseWithTestBrain as a, composeReentryTask as b, resolveWorkerSpawnRetry as c, assertCoordinationBinding as d, coordinationProfileToolPrefix as f, ROOT_STREAM_FILE as g, supervisorAgentWithTestBrain as h, superviseRootProfile as i, retryPreSpawnRefusals as l, supervisorAgent as m, profileGuidanceComposer as n, workerFromBackend as o, resolveSupervisorProfile as p, supervise as r, isPreSpawnExecutorFailure as s, DEFAULT_AUTHORED_PROFILE_SECURITY_POLICY as t, withWorkerSpawnRetry as u, readRootStreamReceipt as v, mapExecutorResult as w, createSupervisorSpanRecorder as x, UNPROVEN_CONTINUITY as y };
 
-//# sourceMappingURL=supervise-Dej5aSDa.js.map
+//# sourceMappingURL=supervise-BoCOUAmr.js.map