@tangle-network/agent-runtime 0.348.1 → 0.350.2

package/dist/{structural-rollout-Wl_rCuzM.js → structural-rollout-B86bSixU.js}

@@ -1,7 +1,7 @@
-import { $i as mapSandboxToolEvent, At as createExecutor, Ci as projectSandboxOutcome, Dt as createExecutorRegistry, Qi as mapSandboxEvent, Ti as runtimeFailureCause, Zi as createSandboxToolPartState, _ as withDriverExecutor, _i as runtimeOwnedPendingExecutorMaterialization, _t as InMemorySpawnJournal, ai as authoredProfileDigest, bi as unknownMaterializationReceipt, ca as awaitAbortable, ci as knownExecutionBindingReceipt, da as providerMessageText, ea as terminalTextFromSandboxEvent, gt as InMemoryResultBlobStore, hi as runtimeOwnedExecutorMaterialization, kt as captureReusableExecutorConfig, la as promptFromAgentTurnInput, li as knownMaterializationReceipt, mi as runtimeOwnedExecutorExecutionBinding, n as createSupervisor, ra as createSandboxUsageLedger, ta as canonicalStreamEventFromSandboxEvent, ua as promptOptionsFromAgentTurnInput, wi as readSandboxOutcome, xi as executableAgentProfileSnapshot, yi as unknownExecutionBindingReceipt } from "./supervisor-DDy0aEtR.js";
-import { m as ValidationError, p as SessionMismatchError, r as BackendTransportError } from "./errors-DodWX-cb.js";
-import { B as sha256Bytes$1, Ct as armDeadlineTimer, X as errorMessage, et as isAsyncIterable } from "./sandbox-profile-compat-_fEtHPCj.js";
-import { c as profileModelExecutionSettings, g as observedModelMatchesDeclared, i as concreteModelId, l as profileProviderModel, o as enforceTokenLimits, s as profileBridgeWireModel, t as assertExecutableAgentProfile } from "./model-policy-CyCUwHed.js";
+import { $i as createSandboxToolPartState, At as createExecutor, Dt as createExecutorRegistry, Ei as runtimeFailureCause, Si as executableAgentProfileSnapshot, Ti as readSandboxOutcome, _ as withDriverExecutor, _t as InMemorySpawnJournal, aa as createSandboxUsageLedger, bi as unknownExecutionBindingReceipt, da as promptFromAgentTurnInput, ea as mapSandboxEvent, fa as promptOptionsFromAgentTurnInput, gi as runtimeOwnedExecutorMaterialization, gt as InMemoryResultBlobStore, hi as runtimeOwnedExecutorExecutionBinding, kt as captureReusableExecutorConfig, li as knownExecutionBindingReceipt, n as createSupervisor, na as terminalTextFromSandboxEvent, oi as authoredProfileDigest, pa as providerMessageText, ra as canonicalStreamEventFromSandboxEvent, ta as mapSandboxToolEvent, ua as awaitAbortable, ui as knownMaterializationReceipt, vi as runtimeOwnedPendingExecutorMaterialization, wi as projectSandboxOutcome, xi as unknownMaterializationReceipt } from "./supervisor-C1OePYE-.js";
+import { m as ValidationError, p as SessionMismatchError, r as BackendTransportError } from "./errors-wuhGvpE3.js";
+import { B as sha256Bytes$1, Z as errorMessage, tt as isAsyncIterable, wt as armDeadlineTimer } from "./sandbox-profile-compat-BxMQ1VXv.js";
+import { c as profileModelExecutionSettings, g as observedModelMatchesDeclared, i as concreteModelId, l as profileProviderModel, o as enforceTokenLimits, s as profileBridgeWireModel, t as assertExecutableAgentProfile } from "./model-policy-GSxWwP-b.js";
 import { agentProfileSchema, canonicalAgentProfileDigest, canonicalCandidateDigest, renderInputPartsAsText } from "@tangle-network/agent-interface";
 import { costForTokenPricing, makeProposalFinding, scoreKnowledgeReadiness } from "@tangle-network/agent-eval";
 import { createAgentRunOutcomeTracker } from "@tangle-network/sandbox/runtime";
@@ -10,7 +10,7 @@ import { existsSync, mkdirSync, renameSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import { assertProposalFindings } from "@tangle-network/agent-eval/analyst";
 import { pathToFileURL } from "node:url";
-//#region src/sessions.ts
+// src/sessions.ts
 /**
 *
 * Session helpers + an in-memory `RuntimeSessionStore` implementation suitable
@@ -101,8 +101,7 @@ var InMemoryRuntimeSessionStore = class {
 		return [...this.events.get(sessionId) ?? []];
 	}
 };
-//#endregion
-//#region src/backends.ts
+// src/backends.ts
 /**
 * Provider-neutral backend adapters for `runAgentTaskStream`.
 *
@@ -240,8 +239,7 @@ function mapCommonBackendEvent(event, context) {
 function stringValue(value) {
 	return typeof value === "string" && value.length > 0 ? value : void 0;
 }
-//#endregion
-//#region src/runtime/agent-turn-backend-events.ts
+// src/runtime/agent-turn-backend-events.ts
 function turnBackendInput(task, input) {
 	return {
 		task,
@@ -687,8 +685,10 @@ function deriveTurnSignal(callerSignal, timeoutMs) {
 	const controller = new AbortController();
 	const clearTimer = timeoutMs > 0 ? armDeadlineTimer(timeoutMs, () => controller.abort(/* @__PURE__ */ new Error(`agent turn timed out after ${timeoutMs}ms`))) : void 0;
 	const onCallerAbort = () => controller.abort(callerSignal?.reason ?? /* @__PURE__ */ new Error("agent turn aborted"));
-	if (callerSignal) if (callerSignal.aborted) onCallerAbort();
-	else callerSignal.addEventListener("abort", onCallerAbort, { once: true });
+	if (callerSignal) {
+		if (callerSignal.aborted) onCallerAbort();
+		else callerSignal.addEventListener("abort", onCallerAbort, { once: true });
+	}
 	return {
 		signal: controller.signal,
 		dispose: () => {
@@ -697,8 +697,7 @@ function deriveTurnSignal(callerSignal, timeoutMs) {
 		}
 	};
 }
-//#endregion
-//#region src/runtime/stream-agent-turn.ts
+// src/runtime/stream-agent-turn.ts
 /**
 * `streamAgentTurn` — the ONE run-a-turn event-stream contract over every
 * execution substrate: a sandbox box (`SandboxInstance.streamPrompt`), a
@@ -1099,8 +1098,7 @@ function materializedModel(receipt, profile) {
 	const fallback = profile.model?.default;
 	return typeof fallback === "string" && fallback.length > 0 ? fallback : void 0;
 }
-//#endregion
-//#region src/runtime/profile-chat-client.ts
+// src/runtime/profile-chat-client.ts
 /** Profile-exact adapter for packages that consume agent-eval's ChatClient contract.
 * Every call still enters Runtime through createExecutor -> streamAgentTurn, and every
 * behavioral field is checked against the exact AgentProfile before any transport runs. */
@@ -1435,8 +1433,7 @@ function optimizerTokenCount(value, label) {
 	if (!Number.isSafeInteger(value) || value < 0) throw new Error(`profile optimizer ${label} must be a non-negative integer`);
 	return value;
 }
-//#endregion
-//#region src/runtime/observe.ts
+// src/runtime/observe.ts
 /**
 * The third-person observer — the connective tissue that closes the loop.
 *
@@ -1593,7 +1590,8 @@ async function analyzeWithProfile(input, opts) {
 	};
 	validateUsage(usage);
 	try {
-		const findings = assertProposalFindings(parseFindings(res.content).map((f) => makeProposalFinding({
+		const parsed = parseFindings(res.content);
+		const findings = assertProposalFindings(parsed.map((f) => makeProposalFinding({
 			analyst_id: observerId,
 			area: f.area,
 			severity: f.severity,
@@ -1687,8 +1685,7 @@ function renderReport(findings) {
 	const block = (title, fs) => fs.length === 0 ? "" : `**${title}**\n${fs.map((f) => `- [${f.severity}] ${f.claim}\n  → ${f.recommended_action ?? ""}`).join("\n")}\n`;
 	return [block("For the agent (fix skills / prompt / tools)", forAgent), block("For you (the operator)", forOperator)].filter(Boolean).join("\n");
 }
-//#endregion
-//#region src/runtime/strategy.ts
+// src/runtime/strategy.ts
 const taskNudge = "Use the available tools to bring the artifact to the required final state. Address EVERY distinct change the request implies. After each tool result, check what remains and continue. Re-read the values you set to confirm they took. Reply DONE only once every required change is made and verified.";
 function exactAgenticProfile(profile, context) {
 	const parsed = agentProfileSchema.safeParse(profile);
@@ -1702,7 +1699,11 @@ function requiredProfileModel(profile, context) {
 	return model;
 }
 function profileSystemPrompt(profile) {
-	const sections = [profile.prompt?.systemPrompt, ...profile.prompt?.instructions ?? []].filter((value) => typeof value === "string" && value.trim().length > 0);
+	const sections = [
+		profile.prompt?.systemPrompt,
+		profile.prompt?.appendSystemPrompt,
+		...profile.prompt?.instructions ?? []
+	].filter((value) => typeof value === "string" && value.trim().length > 0);
 	const instructions = profile.resources?.instructions;
 	if (typeof instructions === "string" && instructions.trim()) sections.push(instructions);
 	else if (instructions && typeof instructions === "object" && instructions.kind === "inline" && instructions.content.trim()) sections.push(instructions.content);
@@ -2176,8 +2177,7 @@ async function runAgentic(opts) {
 		ms: Date.now() - started
 	};
 }
-//#endregion
-//#region src/improvement/optimizer-prompt.ts
+// src/improvement/optimizer-prompt.ts
 /**
 * The senior scientific-method optimizer doctrine — the ONE substantial prompt
 * core shared by every builder/author surface (tool build, MCP build, codebase
@@ -2252,8 +2252,7 @@ const strategyAuthorMethod = [
 	"6. PREDICT, THEN REFLECT: put the hypothesis, the mechanism, and the predicted lift in a",
 	"   comment at the top of the module — the holdout verdict will be read against it."
 ].join("\n");
-//#endregion
-//#region src/runtime/authored-code.ts
+// src/runtime/authored-code.ts
 /**
 * The ONE lint over model-authored source.
 *
@@ -2284,8 +2283,7 @@ function assertAuthoredCode(code, options = {}) {
 		[/child_process|node:fs|node:net|node:http|worker_threads/, "node builtin access"]
 	]) if (pattern.test(code)) throw new ValidationError(`${context} rejected: ${what}`);
 }
-//#endregion
-//#region src/runtime/strategy-author.ts
+// src/runtime/strategy-author.ts
 /**
 * authorStrategy — the agent-authored layer as a package primitive (software-3.0): an
 * LLM reads a benchmark's per-task LOSSES + the defineStrategy contract and writes a NEW
@@ -2313,10 +2311,10 @@ spend a compute budget to beat a task's deployable check. You compose exactly tw
     - steer        => a corrective instruction injected before the shot.
     - profile      => a complete AgentProfile — give THIS shot its own instructions,
       model, skills, tools, hooks, and subagents. Include name, harness, and
-      model: { provider, default }; put standing instructions in prompt.systemPrompt.
+      model: { provider, default }; put standing instructions in prompt.appendSystemPrompt.
       For example: { ...opts.workerProfile, name: 'researcher',
         prompt: { ...opts.workerProfile.prompt,
-          systemPrompt: 'Inspect the evidence before proposing a change.' } }.
+          appendSystemPrompt: 'Inspect the evidence before proposing a change.' } }.
       Choose an available model from the current execution setup. Omit profile to use
       the worker's exact profile. Every shot spends from the same conserved budget.
     - tools        => string[] — restrict THIS shot to a subset of the task's tools by
@@ -2387,7 +2385,7 @@ async function requestStrategySource(opts) {
 			context: "strategy author"
 		}).chat({ messages: [{
 			role: "user",
-			content: `${opts.contract ?? strategyAuthorContract}\n\nBASELINE RESULTS on the "${opts.environmentName}" environment (budget=${opts.budget}) — the per-task losses are your gradient:\n${opts.lossesJson}\n\nAuthor ONE new strategy that you expect to beat the baselines on THIS environment at the same budget.\n${strategyAuthorMethod}\n\nOutput only the module code block.`
+			content: `${opts.contract ?? "\nYou author an OPTIMIZATION STRATEGY for an agentic loop system. A strategy decides how to\nspend a compute budget to beat a task's deployable check. You compose exactly two steps:\n\n  shot(spec?: { handle?, messages?, steer?, profile?, tools? }): Promise<ShotResult | null>\n    Runs ONE worker attempt (a bounded tool loop) over an artifact.\n    - omit handle  => the shot opens its OWN fresh artifact and closes it after (a sample).\n    - pass handle  => the shot CONTINUES that artifact (state accumulates across shots).\n    - messages     => the carried conversation (pass the previous ShotResult.messages to continue).\n    - steer        => a corrective instruction injected before the shot.\n    - profile      => a complete AgentProfile — give THIS shot its own instructions,\n      model, skills, tools, hooks, and subagents. Include name, harness, and\n      model: { provider, default }; put standing instructions in prompt.appendSystemPrompt.\n      For example: { ...opts.workerProfile, name: 'researcher',\n        prompt: { ...opts.workerProfile.prompt,\n          appendSystemPrompt: 'Inspect the evidence before proposing a change.' } }.\n      Choose an available model from the current execution setup. Omit profile to use\n      the worker's exact profile. Every shot spends from the same conserved budget.\n    - tools        => string[] — restrict THIS shot to a subset of the task's tools by\n      name (focus an explore shot on read-only tools, an execute shot on write tools).\n      Restriction-only; unknown names make the shot fail. ALWAYS select from\n      await listTools(handle) — never hardcode. Omitted => the shot sees every tool.\n    ShotResult = { messages, score (0..1 on the task's check), passes, total, completions, toolErrors }\n    Returns null if the attempt failed infra-wise.\n\n  critique(messages): Promise<string | null>\n    A firewalled trace-analyst reads the attempt's trajectory and returns ONE corrective\n    instruction (or null when it judges the work complete). Costs ~1 completion.\n\n  consult(messages, instruction): Promise<string | null>\n    The RAW analyst channel: the same firewalled critic answers YOUR instruction over the\n    trajectory verbatim (no reformatting) — use it when you need a specific reply format\n    (a decision, a prediction). Costs ~1 completion.\n\n  surface.open(task) / surface.close(handle)\n    Open a persistent artifact you manage yourself (remember to close in a finally).\n    close is idempotent — closing an already-closed handle is a safe no-op.\n\n  listTools(handle): Promise<Array<{ name, description? }>>\n    The tools THIS task actually offers. TOOL SETS VARY PER TASK — if you restrict a\n    shot with `tools`, you MUST pick names from await listTools(handle); hardcoding\n    names from an example kills your shots on every task whose tools differ.\n\nRules:\n- ALWAYS await every shot/critique/surface call — a floating promise that rejects\n  crashes the whole benchmark run.\n- Stay within ~budget total shots; every shot/critique spends from a conserved pool.\n- For a FRESH attempt OMIT `messages` entirely (never pass `[]` — an empty array is a\n  fresh conversation too, but be explicit). To CONTINUE, pass the previous\n  ShotResult.messages unchanged.\n- Return { score, resolved, completions, progression, shots } — score = the BEST checkpoint\n  you reached (keep-best, never final-state), progression = score after each shot.\n- The module must be EXACTLY this shape (no other imports, no commentary outside code):\n\nimport { defineStrategy } from '@tangle-network/agent-runtime/kernel'\nexport default defineStrategy('your-strategy-name', async ({ surface, task, opts, budget, shot, critique, listTools }) => {\n  // your composition (listTools comes from the destructured context — it is NOT a global)\n})\n"}\n\nBASELINE RESULTS on the "${opts.environmentName}" environment (budget=${opts.budget}) — the per-task losses are your gradient:\n${opts.lossesJson}\n\nAuthor ONE new strategy that you expect to beat the baselines on THIS environment at the same budget.\n${strategyAuthorMethod}\n\nOutput only the module code block.`
 		}] }, { ...opts.signal ? { signal: opts.signal } : {} });
 		turns.push({
 			content: res.content,
@@ -2418,7 +2416,8 @@ async function requestStrategySource(opts) {
 *  module breaks the contract or exports no Strategy. */
 async function loadAuthoredStrategy(code, outDir) {
 	assertStrategyContract(code);
-	const file = join(outDir, `strategy-${sha256Bytes$1(Buffer.from(code, "utf8")).slice(7)}.mts`);
+	const digest = sha256Bytes$1(Buffer.from(code, "utf8")).slice(7);
+	const file = join(outDir, `strategy-${digest}.mts`);
 	if (!existsSync(file)) {
 		mkdirSync(outDir, { recursive: true });
 		const partial = `${file}.${process.pid}.tmp`;
@@ -2663,8 +2662,7 @@ const sampleThenRefine = defineStrategy("sampleThenRefine", async ({ surface, ta
 		for (const h of open) await surface.close(h);
 	}
 });
-//#endregion
-//#region src/runtime/structural-rollout.ts
+// src/runtime/structural-rollout.ts
 /**
 * structuralRollout — the measured structural lever as a fourth member of the
 * sample/refine/sampleThenRefine strategy family: k independent samples, selection by
@@ -3011,41 +3009,43 @@ function structuralRollout(config = {}) {
 		});
 		let seq = candidates.length;
 		let repairStop = "already-passing";
-		if (!passesAllChecks(best.outcome)) if (best.outcome.crashed !== true && totalChecks(best.outcome) === 0) repairStop = "no-signal";
-		else {
-			repairStop = "rounds-exhausted";
-			for (let r = 0; r < policy.repairRounds; r += 1) {
-				const out = await shot({
-					messages: best.messages,
-					steer: repairSteer(best.outcome)
-				});
-				if (!out) break;
-				shots += 1;
-				completions += out.completions;
-				progression.push(out.score);
-				const artifact = extract(out.messages);
-				const outcome = await checkRunner.run(artifact, checks, runCtx);
-				const displaced = canDisplace(outcome, best.outcome);
-				const label = displaced ? "repair (displaced the incumbent)" : outcome.crashed !== true && outcome.passedOfficial < best.outcome.passedOfficial ? "repair (held out: passes fewer official checks than the incumbent)" : "repair (held out: no improvement)";
-				receipts.push({
-					candidateIndex: seq,
-					selected: false,
-					score: visibleCheckScore(outcome),
-					reason: describeOutcome(label, outcome),
-					selector: "driver"
-				});
-				if (displaced) best = {
-					index: seq,
-					messages: out.messages,
-					artifact,
-					outcome,
-					shotScore: out.score,
-					shotResolved: out.total > 0 && out.passes === out.total
-				};
-				seq += 1;
-				if (passesAllChecks(best.outcome)) {
-					repairStop = "repaired-pass";
-					break;
+		if (!passesAllChecks(best.outcome)) {
+			if (best.outcome.crashed !== true && totalChecks(best.outcome) === 0) repairStop = "no-signal";
+			else {
+				repairStop = "rounds-exhausted";
+				for (let r = 0; r < policy.repairRounds; r += 1) {
+					const out = await shot({
+						messages: best.messages,
+						steer: repairSteer(best.outcome)
+					});
+					if (!out) break;
+					shots += 1;
+					completions += out.completions;
+					progression.push(out.score);
+					const artifact = extract(out.messages);
+					const outcome = await checkRunner.run(artifact, checks, runCtx);
+					const displaced = canDisplace(outcome, best.outcome);
+					const label = displaced ? "repair (displaced the incumbent)" : outcome.crashed !== true && outcome.passedOfficial < best.outcome.passedOfficial ? "repair (held out: passes fewer official checks than the incumbent)" : "repair (held out: no improvement)";
+					receipts.push({
+						candidateIndex: seq,
+						selected: false,
+						score: visibleCheckScore(outcome),
+						reason: describeOutcome(label, outcome),
+						selector: "driver"
+					});
+					if (displaced) best = {
+						index: seq,
+						messages: out.messages,
+						artifact,
+						outcome,
+						shotScore: out.score,
+						shotResolved: out.total > 0 && out.passes === out.total
+					};
+					seq += 1;
+					if (passesAllChecks(best.outcome)) {
+						repairStop = "repaired-pass";
+						break;
+					}
 				}
 			}
 		}
@@ -3065,7 +3065,6 @@ function structuralRollout(config = {}) {
 		};
 	});
 }
-//#endregion
 export { defaultAnalystInstruction as A, InMemoryRuntimeSessionStore as B, optimizerMethod as C, runAgentic as D, refine as E, collectAgentTurn as F, nowIso as H, streamAgentTurn as I, createIterableBackend as L, renderReport as M, profileChatClient as N, sample as O, profileOptimizerModelCall as P, createSandboxPromptBackend as R, strategyAuthorSystemPrompt as S, depthStrategy as T, startOrResumeRuntimeSession as U, newRuntimeSession as V, touchSession as W, defineStrategy as _, defaultStructuralRolloutPolicy as a, sampleThenRefine as b, officialChecksFromMeta as c, selectBestIndex as d, structuralRollout as f, authorStrategy as g, assertStrategyContract as h, defaultExtractCandidate as i, observe as j, ObservationError as k, resolveEntrySymbol as l, adaptiveRefine as m, compareCheckOutcomes as n, filterAuthoredAsserts as o, visibleCheckScore as p, composeCheckSources as r, modelAuthoredChecks as s, canDisplace as t, sandboxCheckRunner as u, loadAuthoredStrategy as v, breadthStrategy as w, strategyAuthorContract as x, requestStrategySource as y, normalizeBackendStreamEvent as z };
 
-//# sourceMappingURL=structural-rollout-Wl_rCuzM.js.map
+//# sourceMappingURL=structural-rollout-B86bSixU.js.map