@tangle-network/agent-runtime 0.133.8 → 0.134.2

package/dist/{runtime-D-26vxJt.d.ts → runtime-DlxoqEj2.d.ts}

@@ -1,10 +1,10 @@
-import { E as SandboxClient, i as ExecCtx } from "./types-6zHo67jn.js";
-import { Q as UsageEvent, U as Spend, d as ExecutorFactory, gn as ExecutorProgress, ln as TraceSource, m as ExecutorRegistry } from "./types-DJ7k6vHE.js";
-import { i as AgentEnvironmentProvider, o as AgentEnvironmentProviderRegistry, w as ProviderExecutorOptions } from "./environment-provider-B-w9VYcF.js";
+import { E as SandboxClient, i as ExecCtx, k as Validator } from "./types-6zHo67jn.js";
+import { Q as UsageEvent, U as Spend, d as ExecutorFactory, gn as ExecutorProgress, ln as TraceSource, m as ExecutorRegistry } from "./types-D4BfurNs.js";
+import { i as AgentEnvironmentProvider, o as AgentEnvironmentProviderRegistry, w as ProviderExecutorOptions } from "./environment-provider-qDOhdE9M.js";
 import { AgentProfile, AgentProfileMcpServer, HarnessType, ReasoningEffort } from "@tangle-network/agent-interface";
 import { WorkspacePlanReceipt } from "@tangle-network/agent-profile-materialize";
 import { ChildProcess } from "node:child_process";
-import { BackendType } from "@tangle-network/sandbox";
+import { BackendType, SandboxEvent } from "@tangle-network/sandbox";
 //#region src/runtime/router-retry-policy.d.ts
 /** Exact retry controls accepted at `AgentProfile.model.metadata.retry`. */
 interface RouterRetryPolicy {
@@ -702,6 +702,22 @@ interface SandboxSeam {
   /** Hard cap on the composed loop's iterations. The budget pool reserves against
    *  the spawn `Budget.maxIterations`; this is the leaf's own ceiling. Default 1. */
   maxIterations?: number;
+  /**
+   * OPT-IN executable score for this worker. Forwarded to the composed
+   * `runAgentRounds` as its `validator`, so the kernel calls `validate` while the
+   * iteration's box is still alive: `ValidationCtx.box` is a LIVE `SandboxInstance`
+   * and the check can run commands or read files in the container it is scoring.
+   * Every other supervised hook fires after teardown and can only read the artifact.
+   *
+   * The resulting verdict becomes the winner's verdict, which this executor already
+   * surfaces on its `ExecutorResult`. Absent, nothing changes: the loop runs
+   * unscored and the leaf falls back to its own settle verdict.
+   *
+   * Not representable with `steering` — a steerable session is a multi-turn session
+   * on one box, not a `runAgentRounds` composition, so the pair is rejected instead
+   * of silently dropping the score.
+   */
+  validator?: Validator<SandboxLeafOut>;
   /**
    * OPT-IN: run this worker as a multi-turn, STEERABLE session instead of the historical
    * single-shot `runAgentRounds` composition. Setting it gives the sandbox worker an `Executor.deliver`
@@ -909,6 +925,11 @@ interface RouterToolsSeam {
     durationMs?: number;
   }) => void;
 }
+/** Parsed output of the sandbox leaf: the iteration's raw event stream. What a
+ *  `SandboxSeam.validator` receives as its `output` argument. */
+interface SandboxLeafOut {
+  events: SandboxEvent[];
+}
 /**
  * The leaf `createWorktreeCliExecutor` as a backend-as-data factory: a supervisor-authored
  * `AgentProfile` driving claude / codex / opencode on its own worktree. `budgetExempt` like
@@ -956,5 +977,5 @@ declare function createExecutor(config: ExecutorConfig): ExecutorFactory<unknown
  */
 declare function createExecutorRegistry(): ExecutorRegistry;
 //#endregion
-export { removeWorktree as $, WorktreeCommandResult as A, RunLocalHarnessOptions as B, ResolvedMcpServerLaunch as C, resolveSecretEnv as D, resolveMcpServerLaunch as E, CodexTokenUsage as F, CreateWorktreeOptions as G, localHarnessExecutable as H, DEFAULT_LOCAL_HARNESS as I, GitRunner as J, DiffOptions as K, LOCAL_HARNESSES as L, WorktreeProfileMaterializationReceipt as M, CodexExecutionEvidence as N, secretEnvOfMcpServer as O, CodexExecutionPolicy as P, createWorktree as Q, LocalHarness as R, KeyProvider as S, mcpSecretEnvMetadataKey as T, parseCodexTokenUsage as U, harnessSupportsReasoningEffort as V, runLocalHarness as W, WorktreeHandle as X, RemoveWorktreeOptions as Y, captureWorktreeDiff as Z, SteerableSandboxSession as _, CliWorktreeSeam as a, ToolLoopCompactionOptions as at, InboxMessage as b, RouterSeam as c, cliWorktreeExecutor as d, RouterTransportConfig as et, createExecutor as f, SteerableSandboxArgs as g, SandboxSteeringOptions as h, CliWorktreeBridgeSeam as i, ToolLoopCompaction as it, WorktreeHarnessResult as j, WorktreeCheckRunner as k, RouterToolsSeam as l, DEFAULT_SANDBOX_STEERING_MAX_TURNS as m, BridgeSeam as n, ToolLoopCallContext as nt, ExecutorConfig as o, ToolLoopMessageRecord as ot, createExecutorRegistry as p, DiffResult as q, CliSeam as r, ToolLoopChat as rt, ProviderSeam as s, ToolLoopToolCall as st, BridgeModelCredential as t, ToolSpec as tt, SandboxSeam as u, createSteerableSandboxSession as v, envKeyProvider as w, createInbox as x, Inbox as y, LocalHarnessResult as z };
-//# sourceMappingURL=runtime-D-26vxJt.d.ts.map
+export { createWorktree as $, WorktreeCheckRunner as A, LocalHarnessResult as B, KeyProvider as C, resolveMcpServerLaunch as D, mcpSecretEnvMetadataKey as E, CodexExecutionPolicy as F, runLocalHarness as G, harnessSupportsReasoningEffort as H, CodexTokenUsage as I, DiffResult as J, CreateWorktreeOptions as K, DEFAULT_LOCAL_HARNESS as L, WorktreeHarnessResult as M, WorktreeProfileMaterializationReceipt as N, resolveSecretEnv as O, CodexExecutionEvidence as P, captureWorktreeDiff as Q, LOCAL_HARNESSES as R, createInbox as S, envKeyProvider as T, localHarnessExecutable as U, RunLocalHarnessOptions as V, parseCodexTokenUsage as W, RemoveWorktreeOptions as X, GitRunner as Y, WorktreeHandle as Z, SteerableSandboxArgs as _, CliWorktreeSeam as a, ToolLoopCompaction as at, Inbox as b, RouterSeam as c, ToolLoopToolCall as ct, SandboxSeam as d, removeWorktree as et, cliWorktreeExecutor as f, SandboxSteeringOptions as g, DEFAULT_SANDBOX_STEERING_MAX_TURNS as h, CliWorktreeBridgeSeam as i, ToolLoopChat as it, WorktreeCommandResult as j, secretEnvOfMcpServer as k, RouterToolsSeam as l, createExecutorRegistry as m, BridgeSeam as n, ToolSpec as nt, ExecutorConfig as o, ToolLoopCompactionOptions as ot, createExecutor as p, DiffOptions as q, CliSeam as r, ToolLoopCallContext as rt, ProviderSeam as s, ToolLoopMessageRecord as st, BridgeModelCredential as t, RouterTransportConfig as tt, SandboxLeafOut as u, SteerableSandboxSession as v, ResolvedMcpServerLaunch as w, InboxMessage as x, createSteerableSandboxSession as y, LocalHarness as z };
+//# sourceMappingURL=runtime-DlxoqEj2.d.ts.map

package/dist/{runtime-Br49sl8u.js → runtime-Ox-uvtni.js}

@@ -1,18 +1,18 @@
-import { c as RetainedRunAdmissionError, f as ValidationError, l as RetainedRunDispatchBindingError, n as AnalystError, s as PlannerError } from "./errors-DRoXIc7E.js";
-import "./stream-agent-turn-D8YHY0oO.js";
-import { S as contentAddress, i as InMemorySpawnJournal, r as InMemoryResultBlobStore, y as isTraceAnalysisStore } from "./spawn-journal-BX3-4WDm.js";
-import { a as isAbortError, i as hasCompleteCacheBreakdown, l as sleep, m as zeroTokenUsage, n as cloneTokenUsage, s as randomSuffix, t as addTokenUsage, u as stringifySafe } from "./util-Dr--8wAL.js";
+import { c as RetainedRunAdmissionError, f as ValidationError, l as RetainedRunDispatchBindingError, n as AnalystError, s as PlannerError } from "./errors-CVIDr7GQ.js";
+import "./stream-agent-turn-CdIcf-bn.js";
+import { S as contentAddress, i as InMemorySpawnJournal, r as InMemoryResultBlobStore, y as isTraceAnalysisStore } from "./spawn-journal-TJSlDa2J.js";
+import { a as hasCompleteCacheBreakdown, c as randomSuffix, d as stringifySafe, h as zeroTokenUsage, n as chargedTokens, o as isAbortError, r as cloneTokenUsage, t as addTokenUsage, u as sleep } from "./util-CG3ghHfb.js";
 import { i as redactProtectedValue, r as redactProtectedReason } from "./protected-redaction--F3v1oo8.js";
 import { a as notifySandboxEventObserver, c as parseCanonicalTransportEvent, s as extractTransportEventIdentity } from "./sandbox-events-vIwmuPdl.js";
-import { c as profileProviderModel, i as concreteModelId, s as profileModelExecutionSettings, t as assertExecutableAgentProfile } from "./model-policy-Bv1DU2fw.js";
-import { T as executableAgentSpecSnapshot, w as executableAgentProfileSnapshot } from "./materialization-D2bdTfkK.js";
-import { F as createExecutor, I as createExecutorRegistry, K as createPushTraceSource, R as createWorktreeCliExecutor, T as rollingDispatch, at as routerBrain, et as defaultSelectWinner, h as settledToIteration, l as withDriverExecutor, n as createSupervisor, nt as createSandboxLineage, ot as runBrainLoop, rt as probeSandboxCapabilities, st as canonicalObservedModel, tt as runAgentRounds } from "./supervisor-BA87pb7g.js";
-import "./environment-provider-DTJA4ZiM.js";
+import { c as profileProviderModel, i as concreteModelId, s as profileModelExecutionSettings, t as assertExecutableAgentProfile } from "./model-policy-UTTc9Z-A.js";
+import { E as executableAgentSpecSnapshot, T as executableAgentProfileSnapshot } from "./materialization-BP5zspr2.js";
+import { F as createExecutor, I as createExecutorRegistry, K as createPushTraceSource, R as createWorktreeCliExecutor, T as rollingDispatch, at as routerBrain, et as defaultSelectWinner, h as settledToIteration, l as withDriverExecutor, n as createSupervisor, nt as createSandboxLineage, ot as runBrainLoop, rt as probeSandboxCapabilities, st as canonicalObservedModel, tt as runAgentRounds } from "./supervisor-4r4BYSti.js";
+import "./environment-provider-CPHwqLLm.js";
 import { i as notifyRuntimeHookEvent } from "./runtime-hooks-C7iJOWm3.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-C4_mifPH.js";
-import { It as gateOnDeliverable, Lt as mapExecutorResult, n as supervise } from "./supervise-BTwl4Cdi.js";
-import "./authoring-DI2ZcORJ.js";
-import "./graph-CnLcp2Tl.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-CnHKUCVM.js";
+import { It as gateOnDeliverable, Lt as mapExecutorResult, n as supervise } from "./supervise-BNOccEtp.js";
+import "./authoring-BDICoQby.js";
+import "./graph-D5TKFG8R.js";
 import { CODING_HARNESSES, InMemoryTraceStore, OUTPUT_VALUE, benjaminiHochberg, buildTrajectory, computeFindingId as computeFindingId$1, confidenceInterval, expandProfileAxes, makeFinding as makeFinding$1, pairedBootstrap, paretoFrontier, wilcoxonSignedRank, wilson } from "@tangle-network/agent-eval";
 import { heldoutSignificance, runProfileMatrix } from "@tangle-network/agent-eval/campaign";
 import { AgentEnvironmentCapabilitiesSchema, AgentExactRunControlRefSchema, AgentNativeContextContinuationResultSchema, AgentRunCancellationAcknowledgementSchema, AgentRunCancellationRequestSchema, AgentTurnResultSchema, InteractionAcknowledgementSchema, InteractionResponseCommandSchema, NativeContextBoundaryProofSchema, NativeContextContinuationRequestSchema, RuntimeEventEnvelopeSchema, agentNativeContextContinuationResultMatchesRequest, agentProfileSchema, agentRunCancellationAcknowledgementMatchesRequest, agentRunCancellationRequestDigest, canonicalAgentProfileDigest, canonicalCandidateDigest, nativeContextContinuationTurnDigest, validateAgentProfileSecurity } from "@tangle-network/agent-interface";
@@ -1097,6 +1097,10 @@ function collectProviderModels(evidence, validate) {
 	const canonical = /* @__PURE__ */ new Set();
 	const observedRaw = /* @__PURE__ */ new Set();
 	for (const attempt of evidence.attempts) {
+		if (attempt.providerDispatch === "not_started") {
+			if (attempt.observations.length > 0 || attempt.identityConflict === true) return { kind: "unknown" };
+			continue;
+		}
 		if (attempt === null || typeof attempt !== "object" || !Array.isArray(attempt.observations) || attempt.observations.length === 0 || attempt.identityConflict === true) return { kind: "unknown" };
 		let attemptBase;
 		let attemptSnapshot;
@@ -3120,13 +3124,18 @@ async function trajectoryReport(journal, blobs, root, options = {}) {
 * within-tolerance when the per-channel spread (max − min across arms) over the median is
 * `≤ tolerance`. Pure over the reports — no I/O. Fails loud on an empty arm list (nothing to
 * compare) so a vacuous "equal" is never returned.
+*
+* The token channel uses `chargedTokens`, the same unit the conserved pool spends, so the cross-run
+* check and the within-run pool cannot disagree about what an arm cost. Charging the rolled-up
+* prompt total instead would rate an arm by how often it re-read a cached prefix: two arms given
+* identical work would read as unequal compute whenever their cache hit rates differed.
 */
 function equalKOnCost(arms, options = {}) {
 	if (arms.length === 0) throw new Error("equalKOnCost: no arms to compare");
 	const tolerance = options.tolerance ?? defaultEqualKTolerance;
 	const armCosts = arms.map((arm) => ({
 		label: arm.label,
-		tokens: arm.report.total.tokens.input + arm.report.total.tokens.output,
+		tokens: chargedTokens(arm.report.total.tokens),
 		usd: arm.report.total.usd,
 		iterations: arm.report.total.iterations
 	}));
@@ -3971,6 +3980,7 @@ async function startRetainedRunInEnvironment(options) {
 	if (!environment) throw new Error(`provider "${options.provider.name}" no longer holds environment "${options.environment.id}"`);
 	if (environment.id !== options.environment.id || environment.provider !== options.provider.name) throw new Error("provider reconstructed a different retained environment");
 	if (!environment.dispatch || !environment.session) throw new Error(`provider "${options.provider.name}" does not expose detached session control`);
+	await assertRetainedEnvironmentOwnership(options.provider, environment.id, options.environment.idempotencyKey);
 	return dispatchRetainedRun({
 		provider: options.provider,
 		environment,
@@ -3982,6 +3992,11 @@ async function startRetainedRunInEnvironment(options) {
 		now: options.now
 	});
 }
+async function assertRetainedEnvironmentOwnership(provider, environmentId, idempotencyKey) {
+	if (!provider.list) throw new Error(`provider "${provider.name}" cannot prove retained environment ownership by metadata`);
+	const matches = (await provider.list({ metadata: { retainedIdempotencyKey: idempotencyKey } })).filter((summary) => summary.id === environmentId && summary.provider === provider.name);
+	if (matches.length !== 1 || matches[0]?.metadata?.retainedIdempotencyKey !== idempotencyKey) throw new Error(`provider "${provider.name}" could not bind environment "${environmentId}" to its retained idempotency key`);
+}
 async function dispatchRetainedRun(options) {
 	const { environment, identity } = options;
 	await admitDurably(options.onAdmission, {
@@ -6654,4 +6669,4 @@ function tail(s) {
 //#endregion
 export { FileCorpus as $, pickChampion as A, renderLeaderboardMarkdown as At, runBenchmark as B, envKeyProvider as Bt, analyzeTrace as C, completionAuthorizes as Ct, chatWorkerSeam as D, leaderboard as Dt, chatTransportExecutor as E, stopSentinel as Et, strategyAuthorContract as F, McpSpawnFault as Ft, promotionGate as G, recoverRetainedRun as H, resolveMcpServerLaunch as Ht, strategyAuthorSystemPrompt as I, connectStdioMcp as It, definePersona as J, equalKOnCost as K, SandboxRunAbortError as L, materializeLocalMcp as Lt, selectChampion as M, renderPairwiseMarkdown as Mt, assertStrategyContract as N, auditIntent as Nt, createChatSessionStore as O, pairwiseSignificance as Ot, authorStrategy as P, defaultAuditorInstruction as Pt, registerShape as Q, openSandboxRun as R, createMcpEnvironment as Rt, withUntrackedArtifacts as S, superviseDispatch as St, runCoderChecks as T, sentinelCompletion as Tt, startRetainedRun as U, resolveSecretEnv as Ut, reconnectRetainedRun as V, mcpSecretEnvMetadataKey as Vt, startRetainedRunInEnvironment as W, secretEnvOfMcpServer as Wt, builtinShapes as X, runPersonified as Y, createShapeRegistry as Z, VERIFY_TAIL_CHARS as _, resolveSandboxClient as _t, localShell as a, panel as at, settledWorkerOut as b, loopCampaignDispatch as bt, createVerifierEnvironment as c, verify as ct, harvestSurfaceDiffs as d, buildSteerContext as dt, InMemoryCorpus as et, failuresAnalyst as f, createScopeAnalyst as ft, NOTE_MAX_CHARS as g, defineLeaderboard as gt, EVIDENCE_MAX_CHARS as h, harvestCorpus as ht, jjWorkspace as i, loopUntil as it, runStrategyEvolution as j, renderLeaderboardSvg as jt, discriminatingMeans as k, renderLeaderboardHtml as kt, boxSurfaceReader as l, widen as lt, worktreeFanout as m, inProcessSandboxClient as mt, makeFinding$1 as n, fanout as nt, runInWorkspace as o, pipeline as ot, superviseSurface as p, registryScopeAnalyst as pt, trajectoryReport as q, gitWorkspace as r, flatWidenGate as rt, createWaterfallCollector as s, selectValidWinner as st, computeFindingId$1 as t, renderCorpusToInstructions as tt, fsSurfaceReader as u, assertTraceDerivedFindings as ut, closingWorkerNote as v, localSandboxClient as vt, patchDelivered as w, deterministicCompletion as wt, copyUntrackedIntoClone as x, loopDispatch as xt, composeWorkerEvidence as y, inlineSandboxClient as yt, printBenchmarkReport as z, sanitizeMcpToolSchema as zt };
 
-//# sourceMappingURL=runtime-Br49sl8u.js.map
+//# sourceMappingURL=runtime-Ox-uvtni.js.map