@tangle-network/agent-runtime 0.115.1 → 0.117.0

package/dist/{runtime-BatQajPB.js → runtime-Ce39zfqn.js}

@@ -1,14 +1,14 @@
 import { n as AnalystError, r as BackendTransportError, s as PlannerError, u as ValidationError } from "./errors-DEAvWQPy.js";
 import { i as normalizeBackendStreamEvent, o as newRuntimeSession, s as nowIso } from "./backends-CiOCyRHb.js";
+import { i as InMemorySpawnJournal, r as InMemoryResultBlobStore, v as isTraceAnalysisStore } from "./spawn-journal-B_mu_oH2.js";
 import { c as sleep, i as isAbortError, l as stringifySafe, n as addTokenUsage, o as randomSuffix, p as zeroTokenUsage } from "./util-Cc9g9Y-o.js";
-import { i as InMemorySpawnJournal, r as InMemoryResultBlobStore } from "./spawn-journal-DCPbicXB.js";
 import { i as redactProtectedValue, r as redactProtectedReason } from "./protected-redaction--F3v1oo8.js";
-import { C as routerChatWithUsage, E as runBrainLoop, d as settledToIteration, l as withDriverExecutor, t as createSupervisor, x as routerBrain } from "./supervisor-DzQu5Ydu.js";
-import { C as observe, O as strategyAuthorMethod, b as sample, v as refine, x as sampleThenRefine, y as runAgentic } from "./structural-rollout-DEf37yQy.js";
+import { Y as rollingDispatch, ct as routerBrain, l as withDriverExecutor, m as settledToIteration, n as createSupervisor, pt as runBrainLoop, ut as routerChatWithUsage } from "./supervisor-CAXiCKnX.js";
+import { C as observe, O as strategyAuthorMethod, b as sample, v as refine, x as sampleThenRefine, y as runAgentic } from "./structural-rollout-DLwW69R4.js";
 import { i as notifyRuntimeHookEvent } from "./runtime-hooks-C7iJOWm3.js";
-import { a as notifySandboxEventObserver, i as mapSandboxToolEvent, r as mapSandboxEvent, t as createSandboxToolPartState } from "./sandbox-events-DeI5xX8P.js";
-import { At as createWorktreeCliExecutor, Jt as runAgentRounds, Ot as createExecutor, St as rollingDispatch, Xt as createSandboxLineage, Zt as probeSandboxCapabilities, kt as createExecutorRegistry, nn as gateOnDeliverable, qt as defaultSelectWinner, t as supervise } from "./supervise-B7TIJR3D.js";
-import { CODING_HARNESSES, InMemoryTraceStore, benjaminiHochberg, buildTrajectory, computeFindingId as computeFindingId$1, confidenceInterval, expandProfileAxes, harnessAxisOf, makeFinding as makeFinding$1, pairedBootstrap, paretoFrontier, scoreKnowledgeReadiness, wilcoxonSignedRank, wilson } from "@tangle-network/agent-eval";
+import { a as notifySandboxEventObserver, i as mapSandboxToolEvent, r as mapSandboxEvent, t as createSandboxToolPartState } from "./sandbox-events-Yhd1GYWl.js";
+import { Dt as createPushTraceSource, Ht as probeSandboxCapabilities, Rt as defaultSelectWinner, Vt as createSandboxLineage, bt as createExecutorRegistry, n as supervise, tn as gateOnDeliverable, xt as createWorktreeCliExecutor, yt as createExecutor, zt as runAgentRounds } from "./supervise-U7_K5yuV.js";
+import { CODING_HARNESSES, InMemoryTraceStore, OUTPUT_VALUE, benjaminiHochberg, buildTrajectory, computeFindingId as computeFindingId$1, confidenceInterval, expandProfileAxes, harnessAxisOf, makeFinding as makeFinding$1, pairedBootstrap, paretoFrontier, scoreKnowledgeReadiness, wilcoxonSignedRank, wilson } from "@tangle-network/agent-eval";
 import { heldoutSignificance, runProfileMatrix } from "@tangle-network/agent-eval/campaign";
 import { canonicalCandidateDigest, validateAgentProfileSecurity } from "@tangle-network/agent-interface";
 import { randomUUID } from "node:crypto";
@@ -2905,7 +2905,7 @@ async function trajectoryReport(journal, blobs, root, options = {}) {
 	const events = await journal.loadTree(root);
 	if (events === void 0) throw new Error(`trajectoryReport: no journaled tree for root '${root}'`);
 	const spawns = events.filter(isNodeCreation).sort(bySeq);
-	const closes = events.filter((ev) => ev.kind !== "spawned" && ev.kind !== "waiting" && ev.kind !== "metered").sort(bySeq);
+	const closes = events.filter((ev) => ev.kind !== "spawned" && ev.kind !== "waiting" && ev.kind !== "metered" && ev.kind !== "materialized" && ev.kind !== "execution-bound").sort(bySeq);
 	const nodes = /* @__PURE__ */ new Map();
 	for (const ev of spawns) nodes.set(ev.id, {
 		id: ev.id,
@@ -3068,6 +3068,7 @@ function addNodeSpend(a, b) {
 			input: a.tokens.input + b.tokens.input,
 			output: a.tokens.output + b.tokens.output
 		},
+		...a.tokensKnown === false || b.tokensKnown === false ? { tokensKnown: false } : {},
 		usd: a.usd + b.usd,
 		...a.tokensKnown === false || b.tokensKnown === false ? { tokensKnown: false } : {},
 		...a.usdKnown === false || b.usdKnown === false ? { usdKnown: false } : {},
@@ -3081,6 +3082,7 @@ function cloneSpend(spend) {
 			input: spend.tokens.input,
 			output: spend.tokens.output
 		},
+		...spend.tokensKnown === false ? { tokensKnown: false } : {},
 		usd: spend.usd,
 		...spend.tokensKnown === false ? { tokensKnown: false } : {},
 		...spend.usdKnown === false ? { usdKnown: false } : {},
@@ -3091,6 +3093,7 @@ function cloneSpend(spend) {
 function addSpend(acc, delta) {
 	acc.iterations += delta.iterations;
 	addTokenUsage(acc.tokens, delta.tokens);
+	if (delta.tokensKnown === false) acc.tokensKnown = false;
 	acc.usd += delta.usd;
 	if (delta.tokensKnown === false) acc.tokensKnown = false;
 	if (delta.usdKnown === false) acc.usdKnown = false;
@@ -5213,24 +5216,31 @@ function worktreeFanout(options) {
 		...options.require !== void 0 ? { require: options.require } : {}
 	});
 	const itemSpec = (item) => {
-		const executor = gateOnDeliverable(createWorktreeCliExecutor({
-			repoRoot: options.repoRoot,
-			profile: item.profile,
-			harness: item.harness,
-			taskPrompt: options.taskPrompt,
-			...item.runId ? { runId: item.runId } : {},
-			...item.baseRef ? { baseRef: item.baseRef } : {},
-			...options.testCmd !== void 0 ? { testCmd: options.testCmd } : {},
-			...options.typecheckCmd !== void 0 ? { typecheckCmd: options.typecheckCmd } : {},
-			...options.harnessTimeoutMs !== void 0 ? { harnessTimeoutMs: options.harnessTimeoutMs } : {},
-			...options.runGit ? { runGit: options.runGit } : {},
-			...options.runHarness ? { runHarness: options.runHarness } : {},
-			...options.runCommand ? { runCommand: options.runCommand } : {}
-		}), deliverable);
+		const executorFactory = (_spec, ctx) => {
+			if (!ctx.node) throw new Error("worktreeFanout: supervised node context required");
+			return gateOnDeliverable(createWorktreeCliExecutor({
+				repoRoot: options.repoRoot,
+				profile: item.profile,
+				harness: item.harness,
+				taskPrompt: options.taskPrompt,
+				executionAttemptId: ctx.node.attemptId,
+				...item.budgetExempt !== void 0 ? { budgetExempt: item.budgetExempt } : {},
+				...item.codexReproducible !== void 0 ? { codexReproducible: item.codexReproducible } : {},
+				...item.codexReadDeniedPaths !== void 0 ? { codexReadDeniedPaths: item.codexReadDeniedPaths } : {},
+				...item.runId ? { runId: item.runId } : {},
+				...item.baseRef ? { baseRef: item.baseRef } : {},
+				...options.testCmd !== void 0 ? { testCmd: options.testCmd } : {},
+				...options.typecheckCmd !== void 0 ? { typecheckCmd: options.typecheckCmd } : {},
+				...options.harnessTimeoutMs !== void 0 ? { harnessTimeoutMs: options.harnessTimeoutMs } : {},
+				...options.runGit ? { runGit: options.runGit } : {},
+				...options.runHarness ? { runHarness: options.runHarness } : {},
+				...options.runCommand ? { runCommand: options.runCommand } : {}
+			}), deliverable);
+		};
 		return {
 			profile: item.profile,
 			harness: null,
-			executor
+			executorFactory
 		};
 	};
 	const selectWinner = selectValidWinner({
@@ -5246,29 +5256,63 @@ function worktreeFanout(options) {
 }
 //#endregion
 //#region src/runtime/supervise-surface.ts
-/** Remember the worker's LAST `run_tests` output so the analyst can name the still-failing tests — a
-*  transparent passthrough for every other surface call. Local to this module (no surface-zoo concept). */
-function captureFailures(base) {
+/**
+* superviseSurface — drive a team of agents to solve a graded `AgenticSurface` task. ONE capability that
+* replaces the worker-seam + "self-improving supervisor" wrapper pair: the driver (`profile`) spawns
+* workers that each run `runAgentic` over the surface (`refine` by default), settle on the surface's OWN
+* check (settled ⟺ resolved — a worker that ran but didn't pass settles invalid, so a keep-best driver
+* never counts it done), and feed the driver a self-improvement lens (the still-FAILING tests, by default)
+* so the next spawn targets the persistently-hard cases. Returns the deployable outcome + the full
+* conserved spend.
+*
+* WHY this lives here and not as a `supervise()` backend: `runAgentic` depends on the supervise core
+* (`strategy.ts` → `supervise/`), so a surface-solving worker cannot be a supervise built-in without an
+* import cycle. It is therefore a COMPOSITION of `supervise()` + `runAgentic` at the layer above both —
+* the right home for "supervise over a graded surface". The within-run self-improvement is the analyst
+* (authored content, swap `analysts`); the across-run kind wraps this call in `improve()`.
+*/
+/** Instrument every real surface call with the shared push trace source. The last test report remains
+*  available on `SurfaceWorkerOut` for compatibility, but analysts read only the persisted spans. */
+function traceSurfaceCalls(base) {
 	let lastReport = "";
-	const surface = {
-		name: base.name,
-		open: (t) => base.open(t),
-		tools: (t, h) => base.tools(t, h),
-		async call(h, name, args) {
-			const out = await base.call(h, name, args);
-			if (name === "run_tests") lastReport = out;
-			return out;
-		},
-		score: (t, h) => base.score(t, h),
-		close: (h) => base.close(h)
-	};
-	const failing = () => {
-		const body = /FAILING:\s*(.+)/i.exec(lastReport)?.[1];
-		return body ? body.split(",").map((s) => s.trim()).filter(Boolean) : [];
-	};
+	const trace = createPushTraceSource();
 	return {
-		surface,
-		failing
+		surface: {
+			name: base.name,
+			open: (t) => base.open(t),
+			tools: (t, h) => base.tools(t, h),
+			async call(h, name, args) {
+				const startedAt = Date.now();
+				const recordedArgs = structuredClone(args);
+				try {
+					const out = await base.call(h, name, args);
+					trace.record({
+						toolName: name,
+						args: recordedArgs,
+						result: out,
+						status: out.startsWith("ERROR:") ? "error" : "ok",
+						startedAt,
+						endedAt: Date.now()
+					});
+					if (name === "run_tests") lastReport = out;
+					return out;
+				} catch (error) {
+					trace.record({
+						toolName: name,
+						args: recordedArgs,
+						result: `ERROR: ${error instanceof Error ? error.message : String(error)}`,
+						status: "error",
+						startedAt,
+						endedAt: Date.now()
+					});
+					throw error;
+				}
+			},
+			score: (t, h) => base.score(t, h),
+			close: (h) => base.close(h)
+		},
+		failing: () => failingTestNames(lastReport),
+		traceSource: trace.source
 	};
 }
 /** The default self-improvement LENS — authored content, not a code path. On each settled worker it hands
@@ -5282,20 +5326,68 @@ function failuresAnalyst() {
 			area: "progress"
 		}],
 		run: async (_kindId, trace) => {
-			const w = trace ?? {};
-			if (!(typeof w === "object" && w !== null && "resolved" in w)) return { summary: `worker produced: ${JSON.stringify(trace).slice(0, 300)}` };
-			if (w.resolved) return { summary: "worker RESOLVED — every check passed; stop." };
-			const failing = w.failing ?? [];
-			const head = `worker did NOT resolve — score ${(100 * (w.score ?? 0)).toFixed(0)}%, ${w.shots ?? "?"} shot(s)`;
-			return { summary: failing.length ? `${head}. STILL FAILING (${failing.length}): ${failing.slice(0, 12).join(", ")}. Spawn the next worker to fix exactly these; if a test keeps failing across workers, give it concrete guidance about that case.` : `${head}. (no failing-test list available this round)` };
+			if (!isTraceAnalysisStore(trace)) return missingRunTestsEvidence();
+			const report = await latestRunTestsReport(trace);
+			if (report === void 0) return missingRunTestsEvidence();
+			const failing = failingTestNames(report);
+			return { summary: failing.length ? `Latest structured run_tests evidence reports STILL FAILING (${failing.length}): ${failing.join(", ")}. Spawn the next worker to fix exactly these; if a test keeps failing across workers, give it concrete guidance about that case.` : allTestsPassed(report) ? "Latest structured run_tests evidence reports every test passed; stop." : `Latest structured run_tests evidence contains no parseable failing-test names. Refusing to infer them from worker prose. run_tests output: ${report.slice(0, 300)}` };
 		}
 	};
 }
+async function latestRunTestsReport(store) {
+	const overview = await store.getOverview({ tool_names: ["run_tests"] });
+	const candidates = [];
+	let ordinal = 0;
+	for (const traceId of overview.sample_trace_ids) {
+		let spans = (await store.viewTrace({
+			trace_id: traceId,
+			per_attribute_byte_cap: 16384
+		})).spans;
+		if (spans === void 0) {
+			const matches = await store.searchTrace({
+				trace_id: traceId,
+				regex_pattern: "run_tests",
+				max_matches: 100
+			});
+			const spanIds = [...new Set(matches.hits.filter((hit) => hit.span_name === "run_tests").map((hit) => hit.span_id))];
+			spans = spanIds.length ? (await store.viewSpans({
+				trace_id: traceId,
+				span_ids: spanIds,
+				per_attribute_byte_cap: 16384
+			})).spans : [];
+		}
+		for (const span of spans) {
+			if (span.tool_name !== "run_tests") continue;
+			const output = span.attributes[OUTPUT_VALUE];
+			if (typeof output !== "string") continue;
+			candidates.push({
+				output,
+				endedAt: span.end_time,
+				ordinal: ordinal++
+			});
+		}
+	}
+	candidates.sort((left, right) => Date.parse(left.endedAt) - Date.parse(right.endedAt) || left.ordinal - right.ordinal);
+	return candidates.at(-1)?.output;
+}
+function failingTestNames(report) {
+	const body = /FAILING:\s*([^\n]+)/iu.exec(report)?.[1];
+	if (body === void 0) return [];
+	return body.replace(/\.\s+COLLECTION-BLOCKED:.*$/iu, "").replace(/\s*\(\+\d+\s+more\)\s*$/iu, "").split(",").map((name) => name.trim()).filter(Boolean);
+}
+function allTestsPassed(report) {
+	const fraction = /(\d+)\s*\/\s*(\d+)\s+tests?\s+passed/iu.exec(report);
+	return fraction !== null && Number(fraction[1]) === Number(fraction[2]);
+}
+function missingRunTestsEvidence() {
+	return { summary: "Missing structured run_tests span evidence. Refusing to infer failing-test names from worker prose." };
+}
 /** One spawned worker = one `runAgentic` attempt over the surface task. The driver's brief is threaded
 *  into the attempt (so a re-spawn can take a targeted angle, not an identical retry); `runAgentic` stamps
 *  real tokens/usd/ms, forwarded as `Spend`; the still-failing tests are captured for the analyst. */
 function surfaceWorkerExecutor(surface, task, worker, strategy) {
 	let artifact;
+	const traced = traceSurfaceCalls(surface);
 	return {
 		runtime: "surface-worker",
 		async execute(brief) {
@@ -5304,9 +5396,8 @@ function surfaceWorkerExecutor(surface, task, worker, strategy) {
 				...task,
 				systemPrompt: `${task.systemPrompt ?? ""}\n\n— Supervisor guidance for THIS attempt (incorporate it; do not just repeat a prior approach) —\n${guidance}`
 			} : task;
-			const cap = captureFailures(surface);
 			const r = await runAgentic({
-				surface: cap.surface,
+				surface: traced.surface,
 				task: attemptTask,
 				strategy,
 				budget: worker.budget ?? 1,
@@ -5321,7 +5412,7 @@ function surfaceWorkerExecutor(surface, task, worker, strategy) {
 				score: r.score,
 				shots: r.shots,
 				summary: `${strategy.name} ${r.shots} shot(s) → ${(100 * r.score).toFixed(0)}% (${r.resolved ? "resolved" : "unresolved"})`,
-				failing: r.resolved ? [] : cap.failing()
+				failing: r.resolved ? [] : traced.failing()
 			};
 			const spent = {
 				iterations: r.completions,
@@ -5340,6 +5431,7 @@ function surfaceWorkerExecutor(surface, task, worker, strategy) {
 			};
 			return artifact;
 		},
+		traceSource: () => traced.traceSource,
 		teardown: () => Promise.resolve({ destroyed: true }),
 		resultArtifact() {
 			if (!artifact) throw new Error("surfaceWorkerExecutor: resultArtifact before execute");
@@ -5822,4 +5914,4 @@ function tail(s) {
 //#endregion
 export { InMemoryCorpus as $, runCoderChecks as A, renderLeaderboardMarkdown as At, SandboxRunAbortError as B, envKeyProvider as Bt, safeWorkerFile as C, completionAuthorizes as Ct, workerInboxFileFromEventDir as D, leaderboard as Dt, workerInboxFile as E, stopSentinel as Et, runStrategyEvolution as F, McpSpawnFault as Ft, equalKOnCost as G, printBenchmarkReport as H, resolveMcpServerLaunch as Ht, selectChampion as I, connectStdioMcp as It, runPersonified as J, trajectoryReport as K, assertStrategyContract as L, materializeLocalMcp as Lt, streamAgentTurn as M, renderPairwiseMarkdown as Mt, discriminatingMeans as N, auditIntent as Nt, writeWorkerSteer as O, pairwiseSignificance as Ot, pickChampion as P, defaultAuditorInstruction as Pt, FileCorpus as Q, authorStrategy as R, createMcpEnvironment as Rt, readWorkerSteerRequests as S, loopDispatch as St, supervisorRunsRoot as T, sentinelCompletion as Tt, runBenchmark as U, resolveSecretEnv as Ut, openSandboxRun as V, mcpSecretEnvMetadataKey as Vt, promotionGate as W, secretEnvOfMcpServer as Wt, createShapeRegistry as X, builtinShapes as Y, registerShape as Z, settledWorkerOut as _, naiveDriver as _t, localShell as a, pipeline as at, analyzeTrace as b, inlineSandboxClient as bt, createVerifierEnvironment as c, widen as ct, worktreeFanout as d, createScopeAnalyst as dt, renderCorpusToInstructions as et, EVIDENCE_MAX_CHARS as f, registryScopeAnalyst as ft, composeWorkerEvidence as g, dumbDriver as gt, closingWorkerNote as h, defineLeaderboard as ht, jjWorkspace as i, panel as it, collectAgentTurn as j, renderLeaderboardSvg as jt, patchDelivered as k, renderLeaderboardHtml as kt, failuresAnalyst as l, assertTraceDerivedFindings as lt, VERIFY_TAIL_CHARS as m, harvestCorpus as mt, makeFinding$1 as n, flatWidenGate as nt, runInWorkspace as o, selectValidWinner as ot, NOTE_MAX_CHARS 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, superviseSurface as u, buildSteerContext as ut, copyUntrackedIntoClone as v, resolveSandboxClient as vt, supervisorRunDir as w, deterministicCompletion as wt, legacySupervisorRunDir as x, loopCampaignDispatch as xt, withUntrackedArtifacts as y, localSandboxClient as yt, strategyAuthorContract as z, sanitizeMcpToolSchema as zt };
 
-//# sourceMappingURL=runtime-BatQajPB.js.map
+//# sourceMappingURL=runtime-Ce39zfqn.js.map