@tangle-network/agent-runtime 0.201.0 → 0.203.0

package/dist/{runtime-DeqdBVeC.js → runtime-B4HhRqE8.js}

@@ -1,15 +1,15 @@
-import { Bt as insideCursorNamespace, C as settledToIteration, Ci as notifySandboxEventObserver, Cn as hasCompleteCacheBreakdown, Dn as sleep, En as randomSuffix, Hr as parseCommittedJsonLines, Mn as zeroSpend, On as stringifySafe, Rt as InMemoryResultBlobStore, Ur as prepareJsonlAppend, Vr as isNoEntError, Wr as writeAllBytes, Xt as contentAddress, Y as rollingDispatch, _t as createPushTraceSource, a as createSupervisor, an as createSandboxLineage, bn as chargedTokens, cn as assertBoxlessPromptOptions, di as detachedFrozen, dn as runBrainLoop, fn as canonicalObservedModelParts, in as runAgentRounds, li as executableAgentProfileSnapshot, ln as readPromptOptions, lt as createWorktreeCliExecutor, m as withDriverExecutor, nn as defaultSelectWinner, on as probeSandboxCapabilities, ot as createExecutor, qt as isTraceAnalysisStore, st as createExecutorRegistry, ui as executableAgentSpecSnapshot, un as routerBrain, wn as isAbortError, xn as cloneSpend, yn as addSpend, zt as InMemorySpawnJournal } from "./redact-h-oaw11Q.js";
+import { Bt as insideCursorNamespace, C as settledToIteration, Ci as notifySandboxEventObserver, Cn as hasCompleteCacheBreakdown, Dn as sleep, En as randomSuffix, Hr as parseCommittedJsonLines, Mn as zeroSpend, On as stringifySafe, Rt as InMemoryResultBlobStore, Ur as prepareJsonlAppend, Vr as isNoEntError, Wr as writeAllBytes, Xt as contentAddress, Y as rollingDispatch, _t as createPushTraceSource, a as createSupervisor, an as createSandboxLineage, bn as chargedTokens, cn as assertBoxlessPromptOptions, di as detachedFrozen, dn as runBrainLoop, fn as canonicalObservedModelParts, in as runAgentRounds, li as executableAgentProfileSnapshot, ln as readPromptOptions, lt as createWorktreeCliExecutor, m as withDriverExecutor, nn as defaultSelectWinner, on as probeSandboxCapabilities, ot as createExecutor, qt as isTraceAnalysisStore, st as createExecutorRegistry, ui as executableAgentSpecSnapshot, un as routerBrain, wn as isAbortError, xn as cloneSpend, yn as addSpend, zt as InMemorySpawnJournal } from "./redact-B5g0jY2n.js";
 import { m as ValidationError, n as AnalystError, s as PlannerError } from "./errors-DodWX-cb.js";
-import { C as observe, O as strategyAuthorMethod, T as profileChatClient, b as sample, v as refine, x as sampleThenRefine, y as runAgentic } from "./structural-rollout-BcMmrpv4.js";
+import { E as profileChatClient, S as ObservationError, T as renderReport, b as sample, k as strategyAuthorMethod, v as refine, w as observe, x as sampleThenRefine, y as runAgentic } from "./structural-rollout-DPwqGLUp.js";
 import { c as profileModelExecutionSettings, i as concreteModelId, l as profileProviderModel, o as enforceTokenLimits, t as assertExecutableAgentProfile } from "./model-policy-DKDyr-fc.js";
 import { i as notifyRuntimeHookEvent } from "./runtime-hooks-tXpAarhW.js";
 import { f as sha256Bytes, i as redactProtectedValue, o as canonicalCandidateDigest$1, r as redactProtectedReason, u as immutableCandidateValue } from "./protected-redaction-wGo44k2K.js";
-import { b as mapExecutorResult, n as supervise, y as gateOnDeliverable } from "./supervise-DmFO50N6.js";
-import "./coordination-driver-qdAwriPV.js";
+import { b as mapExecutorResult, n as supervise, y as gateOnDeliverable } from "./supervise-CskGiJMk.js";
+import "./coordination-driver-B5TLsu_f.js";
 import "./run-layout-B8I_LXN-.js";
-import "./provision-supervisor-BA2-GPth.js";
-import "./delegate-BWHG-zZW.js";
-import "./graph-cCqKhLqz.js";
+import "./provision-supervisor-hc_SOkP5.js";
+import "./delegate-B8y9EvTY.js";
+import "./graph-C3WsOkdi.js";
 import { agentProfileSchema, canonicalAgentProfileDigest, canonicalCandidateDigest, validateAgentProfileSecurity } from "@tangle-network/agent-interface";
 import { CODING_HARNESSES, DEFAULT_TRACE_ANALYST_KINDS, InMemoryTraceStore, OUTPUT_VALUE, benjaminiHochberg, buildTrajectory, computeFindingId as computeFindingId$1, confidenceInterval, createTraceAnalyst, expandProfileAxes, makeFinding as makeFinding$1, pairedBootstrap, paretoFrontier, wilcoxonSignedRank, wilson } from "@tangle-network/agent-eval";
 import { createAgentRunOutcomeTracker } from "@tangle-network/sandbox/runtime";
@@ -2213,9 +2213,34 @@ function defineLeaderboard(spec) {
 }
 //#endregion
 //#region src/runtime/harvest-corpus.ts
-/** Batch the firewalled `observe()` analyst over completed runs and accrete the trace-derived facts into the durable corpus — the production-traces→corpus write side of the flywheel. */
+/**
+* Analyze completed runs through the selected observer and retain findings in a corpus.
+*
+* Store-agnostic by design: the caller maps its trace store's rows (a
+* `ProductionTraceSink` ndjson, OTLP spans, RunRecords) to `ObserveInput` — task text,
+* final output, the event trace, terminal outcome.
+* The default observer reads execution evidence; an injected analysis owns its admitted sources.
+*
+* The nightly product job is then three lines:
+*   const runs = mapSinkRowsToObserveInputs(await readSink(yesterday))
+*   const report = await harvestCorpus({ runs, profile, executor, corpus })
+*   log(report)   // runsObserved / findings / learned / failures
+*
+* Consumers choose how retained findings inform later work and how to assess their value.
+*/
+/** The completed batch evidence remains available even when every analysis failed. */
+var HarvestError = class extends Error {
+	report;
+	constructor(report) {
+		super(`harvestCorpus: every run failed analysis (${report.failures.length}) — first: ${report.failures[0]?.error}`);
+		this.report = report;
+		this.name = "HarvestError";
+	}
+};
+/** Batch the selected observation implementation over completed runs and retain its findings. */
 async function harvestCorpus(opts) {
-	const concurrency = Math.max(1, opts.concurrency ?? 4);
+	const concurrency = opts.concurrency ?? 4;
+	if (!Number.isSafeInteger(concurrency) || concurrency < 1 || opts.maxRuns !== void 0 && (!Number.isSafeInteger(opts.maxRuns) || opts.maxRuns < 0)) throw new TypeError("harvest limits require positive integer concurrency and nonnegative integer maxRuns");
 	const report = {
 		runsObserved: 0,
 		findings: 0,
@@ -2233,26 +2258,34 @@ async function harvestCorpus(opts) {
 	let consumed = 0;
 	let done = false;
 	const next = async () => {
-		if (done || opts.maxRuns !== void 0 && consumed >= opts.maxRuns) return null;
-		const r = await iterator.next();
+		if (done || opts.signal?.aborted || opts.maxRuns !== void 0 && consumed >= opts.maxRuns) return null;
+		const sequence = ++consumed;
+		let r;
+		try {
+			r = await iterator.next();
+		} catch (error) {
+			done = true;
+			report.failures.push({
+				runId: `source-${sequence}`,
+				error: `trace source: ${error instanceof Error ? error.message.slice(0, 300) : String(error)}`
+			});
+			return null;
+		}
 		if (r.done) {
 			done = true;
 			return null;
 		}
-		consumed += 1;
-		return r.value;
+		return {
+			input: r.value,
+			sequence
+		};
 	};
 	const workers = Array.from({ length: concurrency }, async () => {
-		for (let input = await next(); input !== null; input = await next()) {
+		for (let run = await next(); run !== null; run = await next()) {
+			const { input, sequence } = run;
 			if (opts.signal?.aborted) return;
 			try {
-				const obs = await observe(input, {
-					profile: opts.profile,
-					executor: opts.executor,
-					corpus: opts.corpus,
-					tags: opts.tags ?? [],
-					...opts.signal ? { signal: opts.signal } : {}
-				});
+				const obs = await observe(input, opts);
 				report.runsObserved += 1;
 				report.findings += obs.findings.length;
 				report.learned += obs.learned.length;
@@ -2260,16 +2293,20 @@ async function harvestCorpus(opts) {
 				report.usage.output += obs.usage.output;
 				report.usage.known &&= obs.usage.known;
 			} catch (e) {
-				report.usage.known = false;
+				report.usage.known &&= e instanceof ObservationError && e.usage.known;
+				if (e instanceof ObservationError) {
+					report.usage.input += e.usage.input;
+					report.usage.output += e.usage.output;
+				}
 				report.failures.push({
-					runId: input.runId ?? `run-${consumed}`,
+					runId: input.runId ?? `run-${sequence}`,
 					error: e instanceof Error ? e.message.slice(0, 300) : String(e)
 				});
 			}
 		}
 	});
 	await Promise.all(workers);
-	if (report.runsObserved === 0 && report.failures.length > 0) throw new Error(`harvestCorpus: every run failed analysis (${report.failures.length}) — first: ${report.failures[0]?.error}`);
+	if (report.runsObserved === 0 && report.failures.length > 0) throw new HarvestError(report);
 	return report;
 }
 //#endregion
@@ -2397,6 +2434,46 @@ function inProcessSandboxClient(options) {
 	} };
 }
 //#endregion
+//#region src/runtime/observation-registry.ts
+/** Adapt any Eval analyst registry, including recursive engines, to observation and harvesting. */
+function observationFromRegistry(registry, options) {
+	if (!["production", "search"].includes(options.proposalOrigin)) throw new TypeError("registry observation requires an explicit production or search origin");
+	return async (input, context) => {
+		const signals = [context.signal, options.runOptions?.signal].filter((signal) => signal !== void 0);
+		const result = await registry.run(input.runId ?? randomUUID(), typeof options.inputs === "function" ? await options.inputs(input) : options.inputs, {
+			...options.runOptions,
+			...signals.length ? { signal: AbortSignal.any(signals) } : {}
+		});
+		const usage = {
+			input: 0,
+			output: 0,
+			known: true
+		};
+		for (const analyst of result.per_analyst) {
+			const receipt = analyst.usage;
+			usage.input += receipt.tokens?.input ?? receipt.partialTokens?.input ?? 0;
+			usage.output += receipt.tokens?.output ?? receipt.partialTokens?.output ?? 0;
+			usage.known &&= receipt.tokens !== null && receipt.tokens.tokensKnown !== false && receipt.tokensEstimated !== true;
+		}
+		try {
+			await options.record?.(result, input);
+			const incomplete = result.per_analyst.filter((analyst) => analyst.status !== "ok");
+			if (incomplete.length) throw new Error(incomplete.map((analyst) => `${analyst.analyst_id}: ${analyst.error?.message ?? analyst.reason ?? analyst.status}`).join("; "));
+			const findings = result.findings.map((finding) => ({
+				...finding,
+				proposal_origin: options.proposalOrigin
+			}));
+			return {
+				findings,
+				report: renderReport(findings),
+				usage
+			};
+		} catch (cause) {
+			throw new ObservationError(cause instanceof Error ? cause.message : String(cause), usage, { cause });
+		}
+	};
+}
+//#endregion
 //#region src/runtime/personify/analyst.ts
 /**
 * The diagnosis a combinator steers from must be TRACE-derived, never JUDGE-derived. A finding
@@ -6837,6 +6914,6 @@ function tail(s) {
 	return s.slice(-400);
 }
 //#endregion
-export { InMemoryCorpus as $, chatWorkerSeam as A, renderLeaderboardSvg as At, SandboxRunAbortError as B, mcpSecretEnvMetadataKey as Bt, withUntrackedArtifacts as C, deterministicCompletion as Ct, codeModeSupervisorTools as D, pairwiseSignificance as Dt, runCoderChecks as E, leaderboard as Et, selectChampion as F, connectStdioMcp as Ft, equalKOnCost as G, printBenchmarkReport as H, resolveSecretEnv as Ht, assertStrategyContract as I, materializeLocalMcp as It, runPersonified as J, trajectoryReport as K, authorStrategy as L, createMcpEnvironment as Lt, discriminatingMeans as M, auditIntent as Mt, pickChampion as N, defaultAuditorInstruction as Nt, unsafeInProcessRunner as O, renderLeaderboardHtml as Ot, runStrategyEvolution as P, McpSpawnFault as Pt, FileCorpus as Q, strategyAuthorContract as R, sanitizeMcpToolSchema as Rt, copyUntrackedIntoClone as S, completionAuthorizes as St, patchDelivered as T, stopSentinel as Tt, runBenchmark as U, secretEnvOfMcpServer as Ut, openSandboxRun as V, resolveMcpServerLaunch as Vt, promotionGate as W, createTangleSandboxExactProcessProvider as Wt, createShapeRegistry as X, builtinShapes as Y, registerShape as Z, NOTE_MAX_CHARS as _, localSandboxClient as _t, localShell as a, pipeline as at, composeWorkerEvidence as b, loopDispatch as bt, createVerifierEnvironment as c, widen as ct, harvestSurfaceDiffs as d, createScopeAnalyst as dt, renderCorpusToInstructions as et, analystsFromRegistry as f, registryScopeAnalyst as ft, EVIDENCE_MAX_CHARS as g, resolveSandboxClient as gt, worktreeFanout as h, defineLeaderboard as ht, jjWorkspace as i, panel as it, createChatSessionStore as j, renderPairwiseMarkdown as jt, chatTransportExecutor as k, renderLeaderboardMarkdown as kt, boxSurfaceReader as l, assertTraceDerivedFindings as lt, superviseSurface as m, harvestCorpus as mt, makeFinding$1 as n, flatWidenGate as nt, runInWorkspace as o, selectValidWinner as ot, failuresAnalyst as p, inProcessSandboxClient as pt, definePersona as q, gitWorkspace as r, loopUntil as rt, createWaterfallCollector as s, verify as st, computeFindingId$1 as t, fanout as tt, fsSurfaceReader as u, buildSteerContext as ut, VERIFY_TAIL_CHARS as v, inlineSandboxClient as vt, analyzeTrace as w, sentinelCompletion as wt, settledWorkerOut as x, superviseDispatch as xt, closingWorkerNote as y, loopCampaignDispatch as yt, strategyAuthorSystemPrompt as z, envKeyProvider as zt };
+export { InMemoryCorpus as $, chatWorkerSeam as A, renderLeaderboardHtml as At, SandboxRunAbortError as B, sanitizeMcpToolSchema as Bt, withUntrackedArtifacts as C, superviseDispatch as Ct, codeModeSupervisorTools as D, stopSentinel as Dt, runCoderChecks as E, sentinelCompletion as Et, selectChampion as F, defaultAuditorInstruction as Ft, equalKOnCost as G, secretEnvOfMcpServer as Gt, printBenchmarkReport as H, mcpSecretEnvMetadataKey as Ht, assertStrategyContract as I, McpSpawnFault as It, runPersonified as J, trajectoryReport as K, createTangleSandboxExactProcessProvider as Kt, authorStrategy as L, connectStdioMcp as Lt, discriminatingMeans as M, renderLeaderboardSvg as Mt, pickChampion as N, renderPairwiseMarkdown as Nt, unsafeInProcessRunner as O, leaderboard as Ot, runStrategyEvolution as P, auditIntent as Pt, FileCorpus as Q, strategyAuthorContract as R, materializeLocalMcp as Rt, copyUntrackedIntoClone as S, loopDispatch as St, patchDelivered as T, deterministicCompletion as Tt, runBenchmark as U, resolveMcpServerLaunch as Ut, openSandboxRun as V, envKeyProvider as Vt, promotionGate as W, resolveSecretEnv as Wt, createShapeRegistry as X, builtinShapes as Y, registerShape as Z, NOTE_MAX_CHARS as _, defineLeaderboard as _t, localShell as a, pipeline as at, composeWorkerEvidence as b, inlineSandboxClient as bt, createVerifierEnvironment as c, widen as ct, harvestSurfaceDiffs as d, createScopeAnalyst as dt, renderCorpusToInstructions as et, analystsFromRegistry as f, registryScopeAnalyst as ft, EVIDENCE_MAX_CHARS as g, harvestCorpus as gt, worktreeFanout as h, HarvestError as ht, jjWorkspace as i, panel as it, createChatSessionStore as j, renderLeaderboardMarkdown as jt, chatTransportExecutor as k, pairwiseSignificance as kt, boxSurfaceReader as l, assertTraceDerivedFindings as lt, superviseSurface as m, inProcessSandboxClient as mt, makeFinding$1 as n, flatWidenGate as nt, runInWorkspace as o, selectValidWinner as ot, failuresAnalyst as p, observationFromRegistry as pt, definePersona as q, gitWorkspace as r, loopUntil as rt, createWaterfallCollector as s, verify as st, computeFindingId$1 as t, fanout as tt, fsSurfaceReader as u, buildSteerContext as ut, VERIFY_TAIL_CHARS as v, resolveSandboxClient as vt, analyzeTrace as w, completionAuthorizes as wt, settledWorkerOut as x, loopCampaignDispatch as xt, closingWorkerNote as y, localSandboxClient as yt, strategyAuthorSystemPrompt as z, createMcpEnvironment as zt };
 
-//# sourceMappingURL=runtime-DeqdBVeC.js.map
+//# sourceMappingURL=runtime-B4HhRqE8.js.map