@tangle-network/agent-runtime 0.142.3 → 0.143.0

package/dist/{supervisor-B_0ISTFG.js → supervisor-s8SfZYbW.js}

@@ -2,10 +2,10 @@ import { f as RuntimeRunStateError, m as ValidationError, r as BackendTransportE
 import { t as detachedSnapshot } from "./snapshot-cK9K420n.js";
 import { S as contentAddress, a as aggregateProviderModelEvidence, c as pendingWaits, g as runtimeOwnedNestedDriverTreeRoot, h as nestedDriverTreeRoot, l as replaySpawnTree, m as driverRuntime, o as loadSpawnForest, p as attestNestedDriverTreeOwner, s as materializeTreeView, v as captureWorkerTraceEvidence } from "./spawn-journal-D2HKUImC.js";
 import { _ as zeroTokenUsage, a as hasCompleteCacheBreakdown, c as promptCacheTokenClasses, d as sleep, f as stringifySafe, g as withTimeout, h as usdEstimatedOf, i as deleteBoxSafe, l as randomSuffix, m as throwIfAborted, n as chargedTokens, p as throwAbort, r as cloneTokenUsage, s as mapWithConcurrency, t as addTokenUsage, u as randomUuid } from "./util-C1bv6yog.js";
-import { a as extractLlmCallEvent, c as notifySandboxEventObserver, n as assertSandboxServedModel, t as assertSandboxEventSucceeded } from "./sandbox-events-BK-g19wx.js";
+import { i as extractLlmCallEvent, o as mapSandboxToolEvent, r as createSandboxToolPartState, s as notifySandboxEventObserver, t as assertSandboxServedModel } from "./sandbox-events-COwCWUVB.js";
+import { C as unknownExecutionBindingReceipt, D as projectSandboxOutcome, E as executableAgentSpecSnapshot, S as runtimeOwnedScopeOwnerRuntime, T as executableAgentProfileSnapshot, a as authoredProfileDigest, b as runtimeOwnedExecutorProviderEvidence, c as knownExecutionBindingReceipt, f as recordRuntimeOwnedProviderAttemptStart, g as runtimeOwnedDeferredExecutorRuntime, h as recordRuntimeOwnedProviderModel, l as knownMaterializationReceipt, m as recordRuntimeOwnedProviderIdentityConflict, n as attestRuntimeOwnedExecutor, o as finalizeRuntimeOwnedPendingExecutor, p as recordRuntimeOwnedProviderDispatchNotStarted, r as attestRuntimeOwnedPendingExecutor, t as attestRuntimeOwnedDeferredExecutor, u as newExecutionAttemptId, v as runtimeOwnedExecutorExecutionBinding, w as unknownMaterializationReceipt, x as runtimeOwnedPendingExecutorMaterialization, y as runtimeOwnedExecutorMaterialization } from "./materialization-P86EciBw.js";
 import { a as concreteProfileModel, l as resolveRouterRetryPolicy, o as profileBridgeWireModel, s as profileModelExecutionSettings, t as assertExecutableAgentProfile, u as agentHarness } from "./model-policy-DgRJIJse.js";
-import { C as unknownExecutionBindingReceipt, E as executableAgentSpecSnapshot, S as runtimeOwnedScopeOwnerRuntime, T as executableAgentProfileSnapshot, a as authoredProfileDigest, b as runtimeOwnedExecutorProviderEvidence, c as knownExecutionBindingReceipt, f as recordRuntimeOwnedProviderAttemptStart, g as runtimeOwnedDeferredExecutorRuntime, h as recordRuntimeOwnedProviderModel, l as knownMaterializationReceipt, m as recordRuntimeOwnedProviderIdentityConflict, n as attestRuntimeOwnedExecutor, o as finalizeRuntimeOwnedPendingExecutor, p as recordRuntimeOwnedProviderDispatchNotStarted, r as attestRuntimeOwnedPendingExecutor, t as attestRuntimeOwnedDeferredExecutor, u as newExecutionAttemptId, v as runtimeOwnedExecutorExecutionBinding, w as unknownMaterializationReceipt, x as runtimeOwnedPendingExecutorMaterialization, y as runtimeOwnedExecutorMaterialization } from "./materialization-CrdcfWXp.js";
-import { i as resolveAgentEnvironmentProvider, n as providerAsExecutor, r as providerAsSandboxClient } from "./environment-provider-DCEACNt9.js";
+import { i as resolveAgentEnvironmentProvider, n as providerAsExecutor, r as providerAsSandboxClient } from "./environment-provider-DCBJ13VP.js";
 import { i as notifyRuntimeHookEvent } from "./runtime-hooks-tXpAarhW.js";
 import { AGENT_PROFILE_MATERIALIZATION_AXES, AgentTurnInputSchema, REASONING_EFFORTS, agentProfileSchema, canonicalAgentProfileDigest, canonicalCandidateDigest, profileMaterializationAxes, profileMaterializationAxes as profileMaterializationAxes$1, renderInputPartsAsText, sha256DigestSchema } from "@tangle-network/agent-interface";
 import { estimateCost, isModelPriced } from "@tangle-network/agent-eval";
@@ -17,11 +17,13 @@ import { appendFileSync, constants, createReadStream } from "node:fs";
 import { spawn } from "node:child_process";
 import { homedir, tmpdir } from "node:os";
 import { stripVTControlCharacters } from "node:util";
+import { createAgentRunOutcomeTracker } from "@tangle-network/sandbox/runtime";
 import { request } from "node:http";
 import { request as request$1 } from "node:https";
 import { Readable } from "node:stream";
 import { deriveHexId, isW3CSpanId, isW3CTraceId } from "@tangle-network/agent-trace-contract";
 import { SDKError, generateIdempotencyKey, withRetry } from "@tangle-network/agent-core";
+import { collectAgentFinalMessageText, collectAgentResponseText } from "@tangle-network/sandbox";
 //#region src/agent/profile-materialization.ts
 const KNOWN_AXIS_SET = new Set(AGENT_PROFILE_MATERIALIZATION_AXES);
 /**
@@ -3936,7 +3938,7 @@ async function acquireSandbox(client, options, acquire = {}) {
 	while (now() < deadline) {
 		throwIfAborted(acquire.signal);
 		try {
-			return await waitReadyOrDestroy(await client.create(createOpts), deadline, pollMs, acquire.signal, now, sleep$1);
+			return await waitReadyOrDestroy(await client.create(createOpts, acquire.signal === void 0 ? void 0 : { signal: acquire.signal }), deadline, pollMs, acquire.signal, now, sleep$1);
 		} catch (err) {
 			throwIfAborted(acquire.signal);
 			if (!isRetryable(err)) throw err;
@@ -4117,14 +4119,34 @@ async function* pollPromptEvents(box, prompt, sessionId, signal, promptOptions)
 	})).sessionId;
 	const result = await box.session(activeSessionId).result();
 	if (signal.aborted) throwAbort();
+	const resultData = {
+		finalText: result.response ?? "",
+		success: result.success,
+		...result.status ? { status: result.status } : {},
+		...result.error ? { error: result.error } : {},
+		...result.errorCode ? { errorCode: result.errorCode } : {},
+		...result.toolInvocations ? { toolInvocations: result.toolInvocations } : {},
+		...result.approval ? { approval: result.approval } : {},
+		...result.question ? { question: result.question } : {},
+		...result.interaction ? { interaction: result.interaction } : {},
+		...result.plan ? { plan: result.plan } : {},
+		...result.usage ? { usage: result.usage } : {},
+		...result.costUsd !== void 0 ? { costUsd: result.costUsd } : {}
+	};
 	yield {
 		type: "result",
 		id: activeSessionId,
+		data: resultData
+	};
+	yield {
+		type: "done",
+		id: activeSessionId,
 		data: {
-			finalText: result.response ?? "",
-			success: result.success,
-			...result.error ? { error: result.error } : {},
-			...result.usage ? { usage: result.usage } : {}
+			...resultData,
+			outcome: result.status === "awaiting_plan_decision" && result.plan ? {
+				type: "awaiting_plan_decision",
+				plan: result.plan
+			} : { type: "completed" }
 		}
 	};
 }
@@ -4702,14 +4724,14 @@ async function executeIteration(args) {
 			...requestedModel !== void 0 ? { model: requestedModel } : {}
 		};
 		let sawLlmCall = false;
+		const outcomeTracker = createAgentRunOutcomeTracker();
 		for await (const event of stream) {
 			events.push(event);
+			outcomeTracker.observe(event);
 			notifySandboxEventObserver(event, args.ctx.onSandboxEvent, {
 				iterationIndex: args.item.index,
 				agentRunName: slot.agentRunName
 			});
-			assertSandboxEventSucceeded(event);
-			assertSandboxServedModel(event, requested);
 			const llmCall = extractLlmCallEvent(event, slot.agentRunName);
 			if (llmCall) {
 				sawLlmCall = true;
@@ -4728,18 +4750,24 @@ async function executeIteration(args) {
 				});
 				args.ctx.runHandle?.observe(llmCall);
 			}
+			assertSandboxServedModel(event, requested);
 		}
+		const outcome = outcomeTracker.finish();
+		slot.sandboxOutcome = outcome;
 		if (!sawLlmCall) {
 			slot.tokenUsage.tokensKnown = false;
 			slot.costUsdKnown = false;
 		}
 		slot.output = args.output.parse(events);
-		if (args.validator) slot.verdict = await args.validator.validate(slot.output, {
+		const projection = projectSandboxOutcome(outcome);
+		if (projection.status === "failed") throw new Error(projection.reason);
+		if (projection.status === "completed" && args.validator) slot.verdict = await args.validator.validate(slot.output, {
 			iteration: args.item.index,
 			...box ? { box } : {},
 			signal: args.signal,
 			traceEmitter: args.ctx.traceEmitter
 		});
+		else if (projection.status !== "completed") slot.verdict = projection.verdict;
 	} catch (err) {
 		slot.error = err instanceof Error ? err : new Error(String(err));
 	} finally {
@@ -5110,6 +5138,26 @@ function hashJson(value) {
 	return (h >>> 0).toString(16).padStart(8, "0");
 }
 //#endregion
+//#region src/runtime/sandbox-executor-output.ts
+/** Project the complete Sandbox event stream into the standard executor result shape. */
+function sandboxLeafOutputFromEvents(events) {
+	const content = collectAgentFinalMessageText(events) ?? collectAgentResponseText(events) ?? "";
+	const toolCalls = sandboxToolCalls(events);
+	return {
+		events,
+		content,
+		...toolCalls.length === 0 ? {} : { toolCalls }
+	};
+}
+function sandboxToolCalls(events) {
+	const state = createSandboxToolPartState();
+	return events.flatMap((event) => mapSandboxToolEvent(event, state).flatMap((projected) => projected.type === "tool_call" ? [{
+		...projected.toolCallId === void 0 ? {} : { id: projected.toolCallId },
+		name: projected.toolName,
+		arguments: projected.args ?? {}
+	}] : []));
+}
+//#endregion
 //#region src/runtime/supervise/cost-estimate.ts
 /**
 * Pricing for work that arrived without a provider receipt.
@@ -5870,6 +5918,7 @@ function createSteerableSandboxSession(args) {
 		lastText: "",
 		seenToolCalls: /* @__PURE__ */ new Set(),
 		artifact: void 0,
+		latestOutcome: void 0,
 		teardown: void 0,
 		note: "starting"
 	};
@@ -5958,6 +6007,7 @@ function createSteerableSandboxSession(args) {
 				};
 				let events;
 				let sawLlmCall = false;
+				const outcomeTracker = createAgentRunOutcomeTracker();
 				try {
 					if (!handle) {
 						const opened = await lineage.start(spec, prompt, turnController.signal, promptOptions);
@@ -5965,36 +6015,38 @@ function createSteerableSandboxSession(args) {
 						events = opened.events;
 					} else events = await lineage.continue(handle, prompt, turnController.signal, promptOptions);
 					for await (const event of events) {
+						outcomeTracker.observe(event);
 						recordEvent(event);
-						assertSandboxEventSucceeded(event);
-						assertSandboxServedModel(event, requested);
 						const call = extractLlmCallEvent(event, spec.name ?? String(args.harness));
-						if (!call) continue;
-						sawLlmCall = true;
-						const callTokensKnown = call.tokensKnown !== false && typeof call.tokensIn === "number" && typeof call.tokensOut === "number";
-						if (!callTokensKnown) tokensKnown = false;
-						const input = call.tokensIn ?? 0;
-						const output = call.tokensOut ?? 0;
-						if (input || output || !callTokensKnown) {
-							const usage = {
-								kind: "tokens",
-								input,
-								output,
-								...callTokensKnown ? {} : { tokensKnown: false },
-								...promptCacheTokenClasses(call.tokensIn, call.promptCache)
-							};
-							addTokenUsage(tokens, usage);
-							yield usage;
-						}
-						if (typeof call.costUsd === "number" && call.costUsd > 0) {
-							usd += call.costUsd;
-							yield {
-								kind: "cost",
-								usd: call.costUsd,
-								usdKnown: false
-							};
+						if (call) {
+							sawLlmCall = true;
+							const callTokensKnown = call.tokensKnown !== false && typeof call.tokensIn === "number" && typeof call.tokensOut === "number";
+							if (!callTokensKnown) tokensKnown = false;
+							const input = call.tokensIn ?? 0;
+							const output = call.tokensOut ?? 0;
+							if (input || output || !callTokensKnown) {
+								const usage = {
+									kind: "tokens",
+									input,
+									output,
+									...callTokensKnown ? {} : { tokensKnown: false },
+									...promptCacheTokenClasses(call.tokensIn, call.promptCache)
+								};
+								addTokenUsage(tokens, usage);
+								yield usage;
+							}
+							if (typeof call.costUsd === "number" && call.costUsd > 0) {
+								usd += call.costUsd;
+								yield {
+									kind: "cost",
+									usd: call.costUsd,
+									usdKnown: false
+								};
+							}
 						}
+						assertSandboxServedModel(event, requested);
 					}
+					state.latestOutcome = outcomeTracker.finish();
 				} catch (e) {
 					cleanup();
 					if (isInterruptAbort(e, interruptSig, signal, args.controller.signal)) {
@@ -6048,14 +6100,17 @@ function createSteerableSandboxSession(args) {
 		const out = {
 			content: state.lastText,
 			turns: state.turns,
-			toolCalls: activity.read().filter((n) => n.kind === "tool").map((n) => n.label)
+			toolCalls: activity.read().filter((n) => n.kind === "tool").map((n) => n.label),
+			...state.latestOutcome ? { outcome: state.latestOutcome } : {}
 		};
+		const verdict = state.latestOutcome ? projectSandboxOutcome(state.latestOutcome).verdict : void 0;
 		state.artifact = {
 			outRef: args.contentRef("sandbox-steerable", {
 				harness: args.harness,
 				out
 			}),
 			out,
+			...verdict ? { verdict } : {},
 			spent
 		};
 	}
@@ -7217,7 +7272,7 @@ const sandboxExecutor = (spec, ctx) => {
 		}, sandboxMaterialization, sandboxBinding);
 	}
 	const output = { parse(events) {
-		return { events };
+		return sandboxLeafOutputFromEvents(events);
 	} };
 	const driver = singleShotDriver(maxIterations);
 	return attestRuntimeOwnedExecutor({
@@ -7289,12 +7344,21 @@ async function* streamSandboxLeaf(args) {
 		const result = await runAgentRounds(loopOptions);
 		const failure = failedRound(result);
 		if (failure) throw failure;
-		const out = result.winner?.output ?? { events: [] };
-		const verdict = result.winner?.verdict ?? leafVerdict(result);
+		const winningIteration = result.winner ? result.iterations.find((iteration) => iteration.index === result.winner?.iterationIndex) : result.iterations.at(-1);
+		const out = winningIteration?.output ?? sandboxLeafOutputFromEvents(winningIteration?.events ?? []);
+		const sandboxOutcome = winningIteration?.sandboxOutcome;
+		const outWithOutcome = {
+			...out,
+			...sandboxOutcome ? { outcome: sandboxOutcome } : {}
+		};
+		const verdict = sandboxOutcome && !sandboxOutcome.success ? {
+			valid: false,
+			score: 0
+		} : result.winner?.verdict ?? leafVerdict(result);
 		const tokensKnown = result.tokenUsage.tokensKnown !== false;
 		const usdKnown = result.costUsdKnown !== false;
 		const outWithUsage = {
-			...out,
+			...outWithOutcome,
 			...result.estimatedCostUsd !== void 0 ? { estimatedCostUsd: result.estimatedCostUsd } : {},
 			...result.promptCache ? { promptCache: result.promptCache } : {}
 		};
@@ -13045,4 +13109,4 @@ function isNonEmptySpend(s) {
 //#endregion
 export { createInbox as $, AGENT_PROFILE_MATERIALIZATION_AXES as $t, createBudgetPool as A, exportEvalRuns as At, WORKER_TRACE_PROPAGATION as B, sanitizeKnowledgeReadinessReport as Bt, effectiveConcurrency as C, traceContextToEnv as Ct, DEFAULT_SUCCESSFUL_SHUTDOWN_MS as D, buildRuntimeEventOtelSpans as Dt, rollingDispatch as E, buildLoopSpanNodes as Et, cliWorktreeExecutor as F, padTraceId as Ft, DEFAULT_SANDBOX_STEERING_MAX_TURNS as G, createWorktree as Gt, workerTraceEnv as H, runSettledCommand as Ht, createExecutor as I, toOtelAttributes as It, decodeToolPart as J, LOCAL_HARNESSES as Jt, createSteerableSandboxSession as K, removeWorktree as Kt, createExecutorRegistry as L, createRuntimeEventCollector as Lt, bindReusableExecutorExecutionId as M, generateSpanId as Mt, bridgeStopSignalKey as N, loopEventToOtelSpan as Nt, teardownExecutor as O, createOpenInferenceFileExporter as Ot, captureReusableExecutorConfig as P, padSpanId as Pt, readWorkerProgress as Q, CodexExecutionDiagnosticError as Qt, snapshotExecutorConfig as R, createRuntimeStreamEventCollector as Rt, waitUntil as S, readTraceContextFromEnv as St, queueOf as T, buildLoopOtelSpans as Tt, workerTraceHeaders as U, runWorktreeHarness as Ut, readWorkerTraceContext as V, sanitizeRuntimeStreamEvent as Vt, workerTraceSeamKey as W, captureWorktreeDiff as Wt, DEFAULT_STALL_AFTER_MS as X, localHarnessExecutable as Xt, sandboxSessionTraceSource as Y, harnessSupportsReasoningEffort as Yt, createActivityLog as Z, parseCodexTokenUsage as Zt, createWaitProbes as _, runBrainLoop as _t, pickBestDelivered as a, promptControlProfileMaterialization as an, isPeerMailEnvelope as at, timerAt as b, createPropagatingTraceEmitter as bt, driverChild as c, promptResourceProfileMaterialization as cn, TERMINAL_DECISIONS as ct, deriveNodeExecutionIdentity as d, validateProfileMaterialization as dn, isTerminalDecision as dt, assertProfileMaterialization as en, AUTHORITY_MARKERS as et, meterRuntimeOwnedAccounting as f, worktreeCliProfileMaterialization as fn, runAgentRounds as ft, settledToIteration as g, routerBrain as gt, scopeOwnerExecutorNodeContext as h, acquireSandbox as ht, collectDelivered as i, profileMaterializationAxes$1 as in, createPeerMailbox as it, spendFromUsageEvents as j, flatOtelSpan as jt, assertValidBudget as k, createOtelExporter as kt, withDriverExecutor as l, renderProfileMaterializationIssues as ln, createSandboxForSpec as lt, recordScopeOwnerMaterialization as m, probeSandboxCapabilities as mt, createSupervisor as n, defineProfileMaterializationContract as nn, PEER_MAIL_WIRE_KEY as nt, runFinalizer as o, promptModelProfileMaterialization as on, peerMailTools as ot, meterRuntimeOwnedProviderAttempt as p, createSandboxLineage as pt, createPushTraceSource as q, DEFAULT_LOCAL_HARNESS as qt, bestDelivered as r, fullProfileMaterialization as rn, claimsAuthority as rt, runTree as s, promptOnlyProfileMaterialization as sn, peerMailVerbNames as st, createRootHandle as t, controlProfileMaterialization as tn, DEFAULT_PEER_MAIL_LIMITS as tt, createScope as u, sandboxActProfileMaterialization as un, defaultSelectWinner as ut, isWaitOutcome as v, canonicalObservedModel as vt, freeSlots as w, INTELLIGENCE_WIRE_VERSION as wt, validateWaitSpec as x, mergeTraceEnv as xt, pollFor as y, observedModelMatchesDeclared as yt, createWorktreeCliExecutor as z, sanitizeAgentRuntimeEvent as zt };
 
-//# sourceMappingURL=supervisor-B_0ISTFG.js.map
+//# sourceMappingURL=supervisor-s8SfZYbW.js.map