@tangle-network/agent-runtime 0.194.0 → 0.195.0

package/dist/{runtime-BVrqSu0-.js → runtime-CuNJ0opC.js}

@@ -1,13 +1,14 @@
+import { E as detachedFrozen } from "./materialization-vZssF9Nx.js";
 import { m as ValidationError, n as AnalystError, s as PlannerError } from "./errors-CDZ8XsVj.js";
 import "./stream-agent-turn-DOtEsvCV.js";
-import { i as redactProtectedValue, r as redactProtectedReason } from "./protected-redaction--F3v1oo8.js";
+import { f as sha256Bytes, i as redactProtectedValue, o as canonicalCandidateDigest$1, r as redactProtectedReason, u as immutableCandidateValue } from "./protected-redaction-wGo44k2K.js";
 import { l as notifySandboxEventObserver } from "./sandbox-events-DWeQRWxg.js";
 import { a as executableAgentProfileSnapshot, i as writeAllBytes, n as parseCommittedJsonLines, o as executableAgentSpecSnapshot, r as prepareJsonlAppend, t as isNoEntError } from "./jsonl-file-Bh1Q9nt0.js";
 import { c as profileModelExecutionSettings, i as concreteModelId, l as profileProviderModel, o as enforceTokenLimits, t as assertExecutableAgentProfile } from "./model-policy-Wf2MxXJ9.js";
 import { C as settledToIteration, Ht as InMemorySpawnJournal, Vt as InMemoryResultBlobStore, Wt as insideCursorNamespace, Y as rollingDispatch, Zt as isTraceAnalysisStore, a as createSupervisor, an as runAgentRounds, bt as createPushTraceSource, dn as routerBrain, en as contentAddress, fn as runBrainLoop, ft as createWorktreeCliExecutor, h as withDriverExecutor, ln as assertBoxlessPromptOptions, lt as createExecutor, on as createSandboxLineage, pn as canonicalObservedModelParts, rn as defaultSelectWinner, sn as probeSandboxCapabilities, un as readPromptOptions, ut as createExecutorRegistry } from "./redact-BOU77QfQ.js";
 import { E as zeroSpend, _ as randomSuffix, b as sleep, l as chargedTokens, m as isAbortError, p as hasCompleteCacheBreakdown, s as addSpend, u as cloneSpend, x as stringifySafe } from "./environment-provider-Dr-wfnHg.js";
 import { i as notifyRuntimeHookEvent } from "./runtime-hooks-tXpAarhW.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-BcPXFzVb.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-DxEQ77H7.js";
 import { b as mapExecutorResult, n as supervise, y as gateOnDeliverable } from "./supervise-C0V3pZXK.js";
 import "./coordination-driver-iSN-m3VX.js";
 import "./run-layout-B8I_LXN-.js";
@@ -27,6 +28,7 @@ import { tmpdir } from "node:os";
 import { createAgentRunOutcomeTracker } from "@tangle-network/sandbox/runtime";
 import { collectAgentResponseText } from "@tangle-network/sandbox";
 import { createInterface } from "node:readline";
+import { pathToFileURL } from "node:url";
 import { gzipSync } from "node:zlib";
 import { createContext, runInContext } from "node:vm";
 import { stuckLoopView, toolWasteView } from "@tangle-network/agent-eval/pipelines";
@@ -2696,23 +2698,25 @@ function applyFilter(records, filter) {
 var InMemoryCorpus = class {
 	byId = /* @__PURE__ */ new Map();
 	async append(record) {
+		let snapshot;
 		try {
 			assertCorpusRecord(record, "append: record");
+			snapshot = detachedFrozen(record);
 		} catch (err) {
 			return {
 				succeeded: false,
 				error: err instanceof Error ? err.message : String(err)
 			};
 		}
-		const existing = this.byId.get(record.id);
+		const existing = this.byId.get(snapshot.id);
 		if (existing) {
-			if (recordsEqual(existing, record)) return { succeeded: true };
+			if (recordsEqual(existing, snapshot)) return { succeeded: true };
 			return {
 				succeeded: false,
 				error: `corpus conflict: id '${record.id}' is already stored with a different record; a learned fact is append-only — re-mint the id or reconcile before re-appending`
 			};
 		}
-		this.byId.set(record.id, freeze(record));
+		this.byId.set(snapshot.id, snapshot);
 		return { succeeded: true };
 	}
 	async query(filter) {
@@ -2737,40 +2741,51 @@ var FileCorpus = class {
 	async append(record) {
 		try {
 			assertCorpusRecord(record, "append: record");
+			const snapshot = detachedFrozen(record);
+			const fs = await import("node:fs/promises");
+			const { dirname } = await import("node:path");
+			const { tryAcquireAtomicFileLock } = await import("@tangle-network/agent-eval/ledger-core");
+			await fs.mkdir(dirname(this.path), { recursive: true });
+			await (await fs.open(this.path, "a")).close();
+			const canonicalPath = await fs.realpath(this.path);
+			const deadline = Date.now() + 5e3;
+			for (;;) {
+				const acquisition = tryAcquireAtomicFileLock({ lockPath: `${canonicalPath}.lock` });
+				if (!acquisition.acquired) {
+					if (Date.now() >= deadline) throw new Error(`corpus append lock unavailable after 5000ms: ${canonicalPath}`);
+					await new Promise((resolve) => setTimeout(resolve, 25));
+					continue;
+				}
+				try {
+					const existing = (await this.load(canonicalPath)).get(snapshot.id);
+					if (existing) {
+						if (recordsEqual(existing, snapshot)) return { succeeded: true };
+						return {
+							succeeded: false,
+							error: `corpus conflict: id '${snapshot.id}' is already stored in ${canonicalPath} with a different record; a learned fact is append-only — re-mint the id or reconcile before re-appending`
+						};
+					}
+					await this.appendLine(canonicalPath, snapshot);
+					return { succeeded: true };
+				} finally {
+					acquisition.lock.release();
+				}
+			}
 		} catch (err) {
 			return {
 				succeeded: false,
 				error: err instanceof Error ? err.message : String(err)
 			};
 		}
-		let stored;
-		try {
-			stored = await this.load();
-		} catch (err) {
-			return {
-				succeeded: false,
-				error: err instanceof Error ? err.message : String(err)
-			};
-		}
-		const existing = stored.get(record.id);
-		if (existing) {
-			if (recordsEqual(existing, record)) return { succeeded: true };
-			return {
-				succeeded: false,
-				error: `corpus conflict: id '${record.id}' is already stored in ${this.path} with a different record; a learned fact is append-only — re-mint the id or reconcile before re-appending`
-			};
-		}
-		await this.appendLine(record);
-		return { succeeded: true };
 	}
 	async query(filter) {
 		return applyFilter([...(await this.load()).values()], filter);
 	}
-	async load() {
+	async load(path = this.path) {
 		const fs = await import("node:fs/promises");
 		let text;
 		try {
-			text = await fs.readFile(this.path, "utf8");
+			text = await fs.readFile(path, "utf8");
 		} catch (err) {
 			if (isNoEntError(err)) return /* @__PURE__ */ new Map();
 			throw err;
@@ -2782,16 +2797,14 @@ var FileCorpus = class {
 			assertCorpusRecord(parsed, `corpus ${this.path} line ${i + 1}`);
 			const existing = byId.get(parsed.id);
 			if (existing && !recordsEqual(existing, parsed)) throw new Error(`corpus corrupted: ${this.path} has two different records for id '${parsed.id}'; an append-only corpus must never hold a conflicting re-append under one id`);
-			byId.set(parsed.id, freeze(parsed));
+			byId.set(parsed.id, detachedFrozen(parsed));
 		}
 		return byId;
 	}
-	async appendLine(record) {
+	async appendLine(path, record) {
 		const fs = await import("node:fs/promises");
-		const path = await import("node:path");
-		await fs.mkdir(path.dirname(this.path), { recursive: true });
-		const needsSeparator = await prepareJsonlAppend(this.path);
-		const fh = await fs.open(this.path, "a");
+		const needsSeparator = await prepareJsonlAppend(path);
+		const fh = await fs.open(path, "a");
 		try {
 			await writeAllBytes(fh, `${needsSeparator ? "\n" : ""}${JSON.stringify(record)}\n`);
 			await fh.sync();
@@ -2843,9 +2856,6 @@ async function renderCorpusToInstructions(opts) {
 function renderLine(record) {
 	return record.rationale ? `${record.claim} (${record.rationale})` : record.claim;
 }
-function freeze(record) {
-	return Object.freeze({ ...record });
-}
 //#endregion
 //#region src/runtime/personify/registry.ts
 /**
@@ -4029,9 +4039,10 @@ function errorMessage(error) {
 * optimization strategy as code; the caller gates it like any human-built candidate
 * (runBenchmark + a frozen holdout).
 *
-* Structurally safe by construction: the authored body composes shot()/critique() and
-* spends through the Supervisor's conserved pool — it can be wrong, but it cannot
-* Goodhart the check (it never sees the verifiers) and it cannot win by overspending.
+* Brokered shot()/critique() calls spend through the Supervisor's conserved pool.
+* Source validation checks obvious forbidden operations; it is not an execution sandbox.
+* Code that uses other execution paths can access state or spend outside that pool.
+* Callers must provide isolation when authored code requires an enforced boundary.
 *
 * The authored module is written to `outDir` and dynamically imported — run under a
 * TS-capable loader (tsx) since models often emit type annotations.
@@ -4041,16 +4052,20 @@ const strategyAuthorContract = `
 You author an OPTIMIZATION STRATEGY for an agentic loop system. A strategy decides how to
 spend a compute budget to beat a task's deployable check. You compose exactly two steps:
 
-  shot(spec?: { handle?, messages?, steer?, persona?, tools? }): Promise<ShotResult | null>
+  shot(spec?: { handle?, messages?, steer?, profile?, tools? }): Promise<ShotResult | null>
     Runs ONE worker attempt (a bounded tool loop) over an artifact.
     - omit handle  => the shot opens its OWN fresh artifact and closes it after (a sample).
     - pass handle  => the shot CONTINUES that artifact (state accumulates across shots).
     - messages     => the carried conversation (pass the previous ShotResult.messages to continue).
     - steer        => a corrective instruction injected before the shot.
-    - persona      => { systemPrompt?, model? } — give THIS shot its own role and/or model
-      (multi-agent strategies: a researcher shot then an engineer shot, a panel of k
-      personas over one budget). On a fresh shot the systemPrompt replaces the task's; on
-      a carried conversation it arrives as a hand-off message. Same conserved budget.
+    - profile      => a complete AgentProfile — give THIS shot its own instructions,
+      model, skills, tools, hooks, and subagents. Include name, harness, and
+      model: { provider, default }; put standing instructions in prompt.systemPrompt.
+      For example: { ...opts.workerProfile, name: 'researcher',
+        prompt: { ...opts.workerProfile.prompt,
+          systemPrompt: 'Inspect the evidence before proposing a change.' } }.
+      Choose an available model from the current execution setup. Omit profile to use
+      the worker's exact profile. Every shot spends from the same conserved budget.
     - tools        => string[] — restrict THIS shot to a subset of the task's tools by
       name (focus an explore shot on read-only tools, an execute shot on write tools).
       Restriction-only; unknown names make the shot fail. ALWAYS select from
@@ -4088,7 +4103,7 @@ Rules:
 - The module must be EXACTLY this shape (no other imports, no commentary outside code):
 
 import { defineStrategy } from '@tangle-network/agent-runtime/kernel'
-export default defineStrategy('your-strategy-name', async ({ surface, task, budget, shot, critique, listTools }) => {
+export default defineStrategy('your-strategy-name', async ({ surface, task, opts, budget, shot, critique, listTools }) => {
   // your composition (listTools comes from the destructured context — it is NOT a global)
 })
 `;
@@ -4282,23 +4297,66 @@ async function runStrategyEvolution(cfg) {
 	const epsilon = cfg.championEpsilon ?? (cfg.objective === "cost" ? cfg.scoreTolerance ?? .05 : .01);
 	const byName = new Map(baselines.map((s) => [s.name, s]));
 	const codeByName = /* @__PURE__ */ new Map();
-	const fingerprint = {
+	for (const [label, count] of Object.entries({
+		trainN: cfg.trainN,
+		holdoutN: cfg.holdoutN
+	})) if (!Number.isSafeInteger(count) || count < 1) throw new Error(`evolution: ${label} must be a positive integer`);
+	const holdoutOffset = cfg.holdoutOffset ?? 0;
+	if (!Number.isSafeInteger(holdoutOffset) || holdoutOffset < 0) throw new Error("evolution: holdoutOffset must be a non-negative integer");
+	if (baselines.length === 0 || byName.size !== baselines.length) throw new Error("evolution: baselines must have non-empty, unique names");
+	if (cfg.checkpoint && !/^sha256:[0-9a-f]{64}$/.test(cfg.checkpoint.executionRef)) throw new Error("evolution checkpoint: executionRef must be a lowercase sha256:<64 hex> digest");
+	const fingerprint = cfg.checkpoint ? canonicalCandidateDigest$1({
+		schemaVersion: 2,
+		executionRef: cfg.checkpoint?.executionRef ?? null,
+		environment: cfg.environment.name,
 		trainN: cfg.trainN,
 		holdoutN: cfg.holdoutN,
+		holdoutOffset,
 		budget,
+		concurrency,
 		generations,
-		populationSize
-	};
+		populationSize,
+		baselines: baselines.map((baseline) => baseline.name),
+		objective: cfg.objective ?? "score",
+		scoreTolerance: cfg.scoreTolerance ?? .05,
+		champion: policy,
+		championEpsilon: epsilon,
+		minPairedTasks: cfg.minPairedTasks ?? null,
+		band: cfg.band ?? null,
+		lossesDetail: cfg.lossesDetail ?? "exact",
+		reproducerCheck: cfg.reproducerCheck ?? null,
+		modelPreflight: cfg.modelPreflight !== false,
+		modelPreflightTimeoutMs: cfg.modelPreflightTimeoutMs ?? null,
+		worker: {
+			routerBaseUrl: cfg.worker.routerBaseUrl,
+			profile: cfg.worker.workerProfile,
+			analystProfile: cfg.worker.analystProfile ?? cfg.worker.workerProfile,
+			corpusTags: cfg.worker.corpusTags ?? [],
+			corpusReadback: cfg.worker.corpusReadback ?? null
+		},
+		author: {
+			profile: cfg.author.profile,
+			fallbackProfile: cfg.author.fallbackProfile ?? null,
+			executorBackend: cfg.author.executor.backend
+		},
+		authorContract: strategyAuthorContract
+	}) : void 0;
 	let ckpt;
 	if (cfg.checkpoint?.resume && existsSync(cfg.checkpoint.path)) {
 		const raw = JSON.parse(readFileSync(cfg.checkpoint.path, "utf8"));
-		if (JSON.stringify(raw.fingerprint) !== JSON.stringify(fingerprint)) throw new Error(`evolution resume: checkpoint design mismatch — checkpoint ${JSON.stringify(raw.fingerprint)} vs config ${JSON.stringify(fingerprint)}; delete ${cfg.checkpoint.path} or match the config`);
+		if (raw.fingerprint !== fingerprint) throw new Error(`evolution resume: checkpoint design mismatch at ${cfg.checkpoint.path}; use a new checkpoint or restore the original execution dependencies`);
 		ckpt = raw;
 	}
+	const train = checkedTaskSlice(await cfg.tasks(0, cfg.trainN), cfg.trainN, "train", !!cfg.checkpoint);
+	const trainDigest = cfg.checkpoint ? canonicalCandidateDigest$1(train) : void 0;
+	if (ckpt && ckpt.trainDigest !== trainDigest) throw new Error("evolution resume: train task payloads changed");
+	let holdoutPoolDigest;
 	const save = (state) => {
 		if (cfg.checkpoint) writeFileSync(cfg.checkpoint.path, JSON.stringify({
 			...state,
-			fingerprint
+			fingerprint,
+			trainDigest,
+			holdoutPoolDigest
 		}, null, 1));
 	};
 	let modelsPreflighted = false;
@@ -4319,7 +4377,6 @@ async function runStrategyEvolution(cfg) {
 		modelsPreflighted = true;
 		return report;
 	};
-	const train = await cfg.tasks(0, cfg.trainN);
 	const probeTask = train[0];
 	if (!probeTask) throw new Error("runStrategyEvolution: empty train slice");
 	const probe = await cfg.environment.open(probeTask);
@@ -4346,11 +4403,16 @@ async function runStrategyEvolution(cfg) {
 		usd: gen0Champion.usd
 	}];
 	for (const row of generationRows) for (const c of row.candidates) {
-		if (!c.file || c.error) continue;
-		const mod = await import(`file://${c.file}`);
+		if (c.error) continue;
+		if (!c.file || !c.sourceSha256) throw new Error(`evolution resume: missing source identity for '${c.name}'`);
+		const bytes = readFileSync(c.file);
+		if (sha256Bytes(bytes) !== c.sourceSha256) throw new Error(`evolution resume: authored source changed for '${c.name}' (${c.file})`);
+		const sourceUrl = pathToFileURL(c.file);
+		sourceUrl.searchParams.set("sha256", c.sourceSha256);
+		const mod = await import(sourceUrl.href);
 		if (!mod.default || typeof mod.default.driver !== "function") throw new Error(`evolution resume: ${c.file} no longer exports a Strategy — cannot restore "${c.name}"`);
 		byName.set(c.name, renameStrategy(mod.default, c.name));
-		codeByName.set(c.name, readFileSync(c.file, "utf8"));
+		codeByName.set(c.name, bytes.toString("utf8"));
 	}
 	let authoredOk = generationRows.reduce((n, row) => n + row.candidates.filter((c) => !c.error).length, 0);
 	const lastRow = generationRows[generationRows.length - 1];
@@ -4398,6 +4460,7 @@ async function runStrategyEvolution(cfg) {
 				candidates.push({
 					name: unique,
 					file: authored.file,
+					sourceSha256: sha256Bytes(Buffer.from(authored.code)),
 					gzipBits: gzipSync(Buffer.from(authored.code)).length * 8,
 					codeChars: authored.code.length
 				});
@@ -4445,13 +4508,17 @@ async function runStrategyEvolution(cfg) {
 		});
 	}
 	if (authoredOk === 0) throw new Error("runStrategyEvolution: every author attempt failed across all generations — no search happened; see the candidates[].error entries");
-	const holdoutOffset = cfg.trainN + (cfg.holdoutOffset ?? 0);
+	const poolN = cfg.band?.holdoutPoolN ?? cfg.holdoutN;
+	const pool = checkedTaskSlice(await cfg.tasks(cfg.trainN + holdoutOffset, poolN), poolN, "holdout", !!cfg.checkpoint);
+	const trainIds = new Set(train.map((task) => task.id));
+	if (pool.some((task) => trainIds.has(task.id))) throw new Error("evolution: train and holdout task IDs must be disjoint");
+	holdoutPoolDigest = cfg.checkpoint ? canonicalCandidateDigest$1(pool) : void 0;
+	if (ckpt?.holdout && ckpt.holdoutPoolDigest !== holdoutPoolDigest) throw new Error("evolution resume: holdout task payloads changed");
 	let holdoutTasks = [];
 	let bandInfo;
 	if (ckpt?.holdout && ckpt.verdict) {
 		bandInfo = ckpt.band;
 		if (cfg.reproducerCheck && codeByName.has(incumbent.name)) {
-			const pool = await cfg.tasks(holdoutOffset, cfg.band?.holdoutPoolN ?? cfg.holdoutN);
 			const gateIds = new Set(ckpt.holdout.perTask.map((r) => r.taskId));
 			holdoutTasks = pool.filter((t) => gateIds.has(t.id));
 		}
@@ -4459,7 +4526,6 @@ async function runStrategyEvolution(cfg) {
 		const maxRef = cfg.band.maxRefScore ?? .99;
 		const reference = baselines[0];
 		if (!reference) throw new Error("evolution band: baselines[0] required as the screening reference");
-		const pool = await cfg.tasks(holdoutOffset, cfg.band.holdoutPoolN);
 		const refScores = (await bench("band-screen", pool, [reference])).perTask.filter((r) => r.cells?.[reference.name]).map((r) => ({
 			taskId: r.taskId,
 			score: r.cells?.[reference.name]?.score ?? 0
@@ -4473,7 +4539,7 @@ async function runStrategyEvolution(cfg) {
 			inBand: kept.length,
 			refScores
 		};
-	} else holdoutTasks = await cfg.tasks(holdoutOffset, cfg.holdoutN);
+	} else holdoutTasks = pool;
 	let holdout;
 	let verdict;
 	if (ckpt?.holdout && ckpt.verdict) {
@@ -4573,6 +4639,15 @@ async function runStrategyEvolution(cfg) {
 		trajectory
 	};
 }
+function checkedTaskSlice(tasks, count, label, snapshot) {
+	if (!Number.isSafeInteger(count) || count < 1 || tasks.length !== count) throw new Error(`evolution: ${label} must supply exactly ${count} tasks; received ${tasks.length}`);
+	const ids = /* @__PURE__ */ new Set();
+	for (const task of tasks) {
+		if (typeof task.id !== "string" || !task.id.trim() || ids.has(task.id)) throw new Error(`evolution: ${label} task IDs must be non-empty and unique`);
+		ids.add(task.id);
+	}
+	return snapshot ? immutableCandidateValue(tasks) : tasks;
+}
 //#endregion
 //#region src/runtime/supervise/chat-transport-executor.ts
 /**
@@ -6361,4 +6436,4 @@ function tail(s) {
 //#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, 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 };
 
-//# sourceMappingURL=runtime-BVrqSu0-.js.map
+//# sourceMappingURL=runtime-CuNJ0opC.js.map