@tangle-network/agent-runtime 0.70.0 → 0.71.0

package/dist/{delegates-DqAgo32T.d.ts → delegates-CsXJPZDH.d.ts}

@@ -2,7 +2,7 @@ import { ChildProcess } from 'node:child_process';
 import { DefaultVerdict, AgentEvalError } from '@tangle-network/agent-eval';
 import { AgentProfile } from '@tangle-network/agent-interface';
 import { BackendType, SandboxEvent, SandboxInstance } from '@tangle-network/sandbox';
-import { d as LoopTokenUsage, b as RuntimeHooks, e as LoopTraceEmitter, L as LoopTraceEvent, S as SandboxClient, A as AgentRunSpec } from './types-BYa2ZOAx.js';
+import { d as LoopTokenUsage, b as RuntimeHooks, e as LoopTraceEmitter, L as LoopTraceEvent, S as SandboxClient, A as AgentRunSpec } from './types-CdnEAE3U.js';
 import { a as UiLens, U as UiFinding, C as CoderTask } from './substrate-BoRXgvka.js';
 
 /**
@@ -535,6 +535,10 @@ type SupervisedResult<Out> = {
     reason: 'all-children-down' | 'budget-exhausted' | 'aborted';
     tree: TreeView;
     downCount: number;
+    /** The conserved spend incurred before the run failed — real cost is paid even when no
+     *  worker delivers, so the caller always learns what the delegation actually spent. Summed
+     *  off the same journal the `winner` path reads. */
+    spentTotal: Spend;
 };
 /** Live root handle — the substrate a chat/pi-viz client attaches to (Q2). `signal`
  *  delivers an out-of-band message to the running root; `view()` materializes the tree. */
@@ -1656,6 +1660,48 @@ declare class DelegationTaskQueue {
  */
 declare function hashIdempotencyInput(value: unknown): string;
 
+/**
+ * @experimental
+ *
+ * The completion-oracle: **settled ⟺ DELIVERED.**
+ *
+ * Foreman's one hard lesson (0/18 self-improvement deliverables) — "done" must mean a check
+ * PASSED, not the agent's say-so. `gateOnDeliverable` wraps an `Executor` so its settlement
+ * is `valid` ONLY when the deliverable check passes. The child still RUNS and settles (its
+ * spend is conserved into the pool either way), but a child that ran WITHOUT delivering
+ * settles `valid:false` — so a keep-best driver never counts it as done, and a gate never
+ * inflates with self-judged wins.
+ *
+ * Dual-purpose by construction:
+ *  - product: the agent fleet only advances on real, checked deliverables.
+ *  - proof: the gate's `valid` is the honest settle — equal-k comparisons can't be gamed by an
+ *    arm that "ran" without producing the artifact.
+ *
+ * The check is a DEPLOYABLE oracle — a test command, a state verifier, the commit0 judge —
+ * read off the child's output, never the model judging itself. A throwing check is
+ * fail-closed (not delivered), never a crash.
+ */
+
+/**
+ * The deployable completion oracle passed to {@link gateOnDeliverable}: a `check` that
+ * decides DELIVERED (settles `valid` ⟺ it resolves true) plus an optional `describe` of
+ * what the spawn was supposed to produce. The check reads the child's output — never the
+ * model judging itself.
+ */
+interface DeliverableSpec<Out = unknown> {
+    /** The deployable check that decides DELIVERED. `settled.valid ⟺ this resolves true`. */
+    check: (out: Out) => boolean | Promise<boolean>;
+    /** What the spawn was supposed to produce — surfaced in traces/reports. */
+    describe?: string;
+}
+/**
+ * Wrap an `Executor` so its settlement `valid` reflects the deliverable check, not the
+ * inner verdict. Handles both `execute` shapes (one-shot `Promise<ExecutorResult>` and
+ * streaming `AsyncIterable<UsageEvent>` + `resultArtifact()`); the check runs once the inner
+ * executor has produced its output. The inner `score` is preserved; only `valid` is gated.
+ */
+declare function gateOnDeliverable<Out>(inner: Executor<Out>, deliverable: DeliverableSpec<Out>): Executor<Out>;
+
 /**
  * @experimental
  *
@@ -2239,4 +2285,4 @@ interface SettleDetachedCoderTurnOptions {
  */
 declare function settleDetachedCoderTurn(turn: DetachedTurn, options: SettleDetachedCoderTurnOptions): Promise<CoderOutput>;
 
-export { type DelegationProfile as $, type Agent as A, type Budget as B, type CoderReviewer as C, type DelegateCodeArgs as D, type ExecutorRegistry as E, type FleetHandle as F, type GitRunner as G, type DelegationStatusArgs as H, type CappedDelegationTrace as I, type CoderReview as J, type CreateWorktreeOptions as K, type LocalHarness as L, DELEGATION_TRACE_MAX_BYTES as M, type NodeId as N, DELEGATION_TRACE_MAX_SPANS as O, type DelegateCodeConfig as P, type DelegateResearchConfig as Q, type ResearcherDelegate as R, type Scope as S, type TraceContext as T, type UiAuditorDelegate as U, type DelegateRunCtx as V, type DelegateUiAuditConfig as W, type DelegateUiAuditRoute as X, type DelegationError as Y, type DelegationHistoryEntry as Z, DelegationPersistenceError as _, type DetachedWinnerSelection as a, type EvalRunsExportResult as a$, type DelegationProgress as a0, type DelegationRecord as a1, type DelegationResultPayload as a2, type DelegationResumeContext as a3, type DelegationResumeDriver as a4, type DelegationResumeTick as a5, type DelegationRunContext as a6, DelegationStateCorruptError as a7, type DelegationStatus as a8, type DelegationStore as a9, type SubmitOutput as aA, type UiAuditorDelegationOutput as aB, type WorktreeHandle as aC, buildDelegationTraceSpans as aD, capDelegationTrace as aE, captureWorktreeDiff as aF, coderTaskFromArgs as aG, composeLoopTraceEmitters as aH, createDelegationTraceCollector as aI, createDetachedTurnResumeDriver as aJ, createFleetWorkspaceExecutor as aK, createPropagatingTraceEmitter as aL, createSiblingSandboxExecutor as aM, createWorktree as aN, detachedSessionDelegate as aO, detachedTurnEvents as aP, formatDetachedSessionRef as aQ, hashIdempotencyInput as aR, parseDetachedSessionRef as aS, readTraceContextFromEnv as aT, removeWorktree as aU, runDetachedTurn as aV, settleDetachedCoderTurn as aW, traceContextToEnv as aX, type EvalRunEvent as aY, type EvalRunGeneration as aZ, type EvalRunsExportConfig as a_, type DelegationTaskQueueOptions as aa, type DelegationTraceCaps as ab, type DelegationTraceCollector as ac, type DelegationTraceSpan as ad, type DetachedSessionDelegateOptions as ae, type DetachedSessionRefParts as af, type DetachedTurn as ag, type DetachedTurnResumeDriverOptions as ah, type DiffOptions as ai, type DiffResult as aj, type DriveTurnCapableBox as ak, type DriveTurnTick as al, type FeedbackRating as am, type FeedbackRefersTo as an, FileDelegationStore as ao, type FileDelegationStoreOptions as ap, type FleetWorkspaceExecutorOptions as aq, InMemoryDelegationStore as ar, type LocalHarnessResult as as, type RemoveWorktreeOptions as at, type ResearchOutputShape as au, type RunDetachedTurnOptions as av, type RunLocalHarnessOptions as aw, type SettleDetachedCoderTurnOptions as ax, type SiblingSandboxExecutorOptions as ay, type SubmitInput as az, type CoderOutput as b, INTELLIGENCE_WIRE_VERSION as b0, type LoopSpanNode as b1, type OtelAttribute as b2, type OtelExportConfig as b3, type OtelExporter as b4, type OtelSpan as b5, buildLoopOtelSpans as b6, buildLoopSpanNodes as b7, createOtelExporter as b8, exportEvalRuns as b9, loopEventToOtelSpan as ba, type SpawnEvent as bb, type ExecutorFactory as bc, type UsageEvent as bd, type Supervisor as be, type ExecutorContext as bf, type ExecutorResult as bg, type SupervisorOpts as bh, type TreeView as bi, type WidenGate as bj, type DelegateFeedbackArgs as c, type DelegationFeedbackSnapshot as d, DelegationTaskQueue as e, type CoderDelegate as f, type ResultBlobStore as g, type AgentSpec as h, type SpawnJournal as i, type RootHandle as j, type SupervisedResult as k, type Settled as l, type Spend as m, type Executor as n, type DelegationExecutor as o, type DelegateCodeResult as p, type DelegateFeedbackResult as q, runLocalHarness as r, type ResearchSource as s, type DelegateResearchArgs as t, type DelegateResearchResult as u, type DelegateUiAuditArgs as v, type DelegateUiAuditResult as w, type DelegationHistoryResult as x, type DelegationHistoryArgs as y, type DelegationStatusResult as z };
+export { type DelegationHistoryEntry as $, type Agent as A, type Budget as B, type CoderReviewer as C, type DelegateCodeArgs as D, type ExecutorFactory as E, type FleetHandle as F, type GitRunner as G, type DelegationHistoryArgs as H, type DelegationStatusResult as I, type DelegationStatusArgs as J, type CappedDelegationTrace as K, type LocalHarness as L, type CoderReview as M, type NodeId as N, type CreateWorktreeOptions as O, DELEGATION_TRACE_MAX_BYTES as P, DELEGATION_TRACE_MAX_SPANS as Q, type ResearcherDelegate as R, type Spend as S, type TraceContext as T, type UiAuditorDelegate as U, type DelegateCodeConfig as V, type DelegateResearchConfig as W, type DelegateRunCtx as X, type DelegateUiAuditConfig as Y, type DelegateUiAuditRoute as Z, type DelegationError as _, type DetachedWinnerSelection as a, type EvalRunGeneration as a$, DelegationPersistenceError as a0, type DelegationProfile as a1, type DelegationProgress as a2, type DelegationRecord as a3, type DelegationResultPayload as a4, type DelegationResumeContext as a5, type DelegationResumeDriver as a6, type DelegationResumeTick as a7, type DelegationRunContext as a8, DelegationStateCorruptError as a9, type SiblingSandboxExecutorOptions as aA, type SubmitInput as aB, type SubmitOutput as aC, type UiAuditorDelegationOutput as aD, type WorktreeHandle as aE, buildDelegationTraceSpans as aF, capDelegationTrace as aG, captureWorktreeDiff as aH, coderTaskFromArgs as aI, composeLoopTraceEmitters as aJ, createDelegationTraceCollector as aK, createDetachedTurnResumeDriver as aL, createFleetWorkspaceExecutor as aM, createPropagatingTraceEmitter as aN, createSiblingSandboxExecutor as aO, createWorktree as aP, detachedSessionDelegate as aQ, detachedTurnEvents as aR, formatDetachedSessionRef as aS, hashIdempotencyInput as aT, parseDetachedSessionRef as aU, readTraceContextFromEnv as aV, removeWorktree as aW, runDetachedTurn as aX, settleDetachedCoderTurn as aY, traceContextToEnv as aZ, type EvalRunEvent as a_, type DelegationStatus as aa, type DelegationStore as ab, type DelegationTaskQueueOptions as ac, type DelegationTraceCaps as ad, type DelegationTraceCollector as ae, type DelegationTraceSpan as af, type DetachedSessionDelegateOptions as ag, type DetachedSessionRefParts as ah, type DetachedTurn as ai, type DetachedTurnResumeDriverOptions as aj, type DiffOptions as ak, type DiffResult as al, type DriveTurnCapableBox as am, type DriveTurnTick as an, type FeedbackRating as ao, type FeedbackRefersTo as ap, FileDelegationStore as aq, type FileDelegationStoreOptions as ar, type FleetWorkspaceExecutorOptions as as, InMemoryDelegationStore as at, type LocalHarnessResult as au, type RemoveWorktreeOptions as av, type ResearchOutputShape as aw, type RunDetachedTurnOptions as ax, type RunLocalHarnessOptions as ay, type SettleDetachedCoderTurnOptions as az, type CoderOutput as b, type EvalRunsExportConfig as b0, type EvalRunsExportResult as b1, INTELLIGENCE_WIRE_VERSION as b2, type LoopSpanNode as b3, type OtelAttribute as b4, type OtelExportConfig as b5, type OtelExporter as b6, type OtelSpan as b7, buildLoopOtelSpans as b8, buildLoopSpanNodes as b9, createOtelExporter as ba, exportEvalRuns as bb, loopEventToOtelSpan as bc, type SpawnEvent as bd, type UsageEvent as be, type Supervisor as bf, type ExecutorContext as bg, type ExecutorResult as bh, type SupervisorOpts as bi, type TreeView as bj, type WidenGate as bk, gateOnDeliverable as bl, type DelegateFeedbackArgs as c, type DelegationFeedbackSnapshot as d, type ExecutorRegistry as e, type DeliverableSpec as f, DelegationTaskQueue as g, type CoderDelegate as h, type Scope as i, type ResultBlobStore as j, type AgentSpec as k, type SpawnJournal as l, type RootHandle as m, type SupervisedResult as n, type Settled as o, type Executor as p, type DelegationExecutor as q, runLocalHarness as r, type DelegateCodeResult as s, type DelegateFeedbackResult as t, type ResearchSource as u, type DelegateResearchArgs as v, type DelegateResearchResult as w, type DelegateUiAuditArgs as x, type DelegateUiAuditResult as y, type DelegationHistoryResult as z };

package/dist/{improvement-adapter-BVuMragr.d.ts → improvement-adapter-CioiEE2z.d.ts}

@@ -1,5 +1,5 @@
 import { FindingSubject, AnalystFinding } from '@tangle-network/agent-eval';
-import { I as ImprovementAdapter } from './types-DJu6TBGp.js';
+import { I as ImprovementAdapter } from './types-BC3bZpH0.js';
 
 /**
  * `AgentSurfaces` — declarative map of the mutable file/directory paths

package/dist/index.d.ts

@@ -1,21 +1,21 @@
 import { AgentProfile, AgentEvalError, AnalystFinding, KnowledgeReadinessReport, RunRecord, ControlEvalResult, KnowledgeRequirement } from '@tangle-network/agent-eval';
 export { AgentEvalError, AgentEvalErrorCode, ConfigError, ControlBudget, ControlDecision, ControlEvalResult, ControlRunResult, ControlStep, DataAcquisitionPlan, JudgeError, KnowledgeReadinessReport, KnowledgeRequirement, NotFoundError, RunRecord, ValidationError } from '@tangle-network/agent-eval';
-import { f as AgentBackendInput, g as AgentExecutionBackend, O as OpenAIChatTool, h as OpenAIChatToolChoice, i as AgentBackendContext, R as RuntimeStreamEvent, K as KnowledgeReadinessDecision, j as RunAgentTaskOptions, k as AgentTaskRunResult, l as RunAgentTaskStreamOptions, m as AgentRuntimeEvent, n as AgentTaskStatus, o as RuntimeSessionStore, p as RuntimeSession, b as RuntimeHooks } from './types-BYa2ZOAx.js';
-export { q as AgentAdapter, r as AgentKnowledgeProvider, s as AgentRuntimeEventSink, t as AgentTaskContext, u as AgentTaskSpec, B as BackendErrorDetail, v as RuntimeDecisionEvidenceRef, w as RuntimeDecisionKind, x as RuntimeDecisionPoint, y as RuntimeHookContext, z as RuntimeHookErrorContext, C as RuntimeHookEvent, D as RuntimeHookPhase, E as RuntimeHookTarget, F as RuntimeRunHandle, G as RuntimeRunPersistenceAdapter, H as RuntimeRunRow, J as composeRuntimeHooks, M as defineRuntimeHooks, N as notifyRuntimeDecisionPoint, P as notifyRuntimeHookEvent, Q as startRuntimeRun } from './types-BYa2ZOAx.js';
+import { f as AgentBackendInput, g as AgentExecutionBackend, O as OpenAIChatTool, h as OpenAIChatToolChoice, i as AgentBackendContext, R as RuntimeStreamEvent, K as KnowledgeReadinessDecision, j as RunAgentTaskOptions, k as AgentTaskRunResult, l as RunAgentTaskStreamOptions, m as AgentRuntimeEvent, n as AgentTaskStatus, o as RuntimeSessionStore, p as RuntimeSession, b as RuntimeHooks } from './types-CdnEAE3U.js';
+export { q as AgentAdapter, r as AgentKnowledgeProvider, s as AgentRuntimeEventSink, t as AgentTaskContext, u as AgentTaskSpec, B as BackendErrorDetail, v as RuntimeDecisionEvidenceRef, w as RuntimeDecisionKind, x as RuntimeDecisionPoint, y as RuntimeHookContext, z as RuntimeHookErrorContext, C as RuntimeHookEvent, D as RuntimeHookPhase, F as RuntimeHookTarget, G as RuntimeRunHandle, H as RuntimeRunPersistenceAdapter, J as RuntimeRunRow, M as composeRuntimeHooks, N as defineRuntimeHooks, P as notifyRuntimeDecisionPoint, Q as notifyRuntimeHookEvent, T as startRuntimeRun } from './types-CdnEAE3U.js';
 import { AgentProfile as AgentProfile$1 } from '@tangle-network/agent-interface';
 import { Scenario, ProfileDispatchFn, LabeledScenarioStore, WorktreeAdapter, SurfaceProposer } from '@tangle-network/agent-eval/campaign';
-import { L as LocalHarness, r as runLocalHarness } from './delegates-DqAgo32T.js';
-export { aY as EvalRunEvent, aZ as EvalRunGeneration, a_ as EvalRunsExportConfig, a$ as EvalRunsExportResult, b0 as INTELLIGENCE_WIRE_VERSION, b1 as LoopSpanNode, b2 as OtelAttribute, b3 as OtelExportConfig, b4 as OtelExporter, b5 as OtelSpan, b6 as buildLoopOtelSpans, b7 as buildLoopSpanNodes, b8 as createOtelExporter, b9 as exportEvalRuns, ba as loopEventToOtelSpan } from './delegates-DqAgo32T.js';
+import { L as LocalHarness, r as runLocalHarness } from './delegates-CsXJPZDH.js';
+export { a_ as EvalRunEvent, a$ as EvalRunGeneration, b0 as EvalRunsExportConfig, b1 as EvalRunsExportResult, b2 as INTELLIGENCE_WIRE_VERSION, b3 as LoopSpanNode, b4 as OtelAttribute, b5 as OtelExportConfig, b6 as OtelExporter, b7 as OtelSpan, b8 as buildLoopOtelSpans, b9 as buildLoopSpanNodes, ba as createOtelExporter, bb as exportEvalRuns, bc as loopEventToOtelSpan } from './delegates-CsXJPZDH.js';
 import { Scenario as Scenario$1, SurfaceProposer as SurfaceProposer$1, JudgeConfig, MutableSurface, DispatchContext, SelfImproveBudget, SelfImproveLlm, SelfImproveResult } from '@tangle-network/agent-eval/contract';
-import { S as SurfaceImprovementEdit } from './improvement-adapter-BVuMragr.js';
-import { I as ImprovementAdapter } from './types-DJu6TBGp.js';
-export { C as CoderLoopRunnerOptions, D as DELEGATED_LOOP_MODES, a as DelegatedLoopMode, b as DelegatedLoopRegistry, c as DelegatedLoopResult, d as DelegatedLoopRunner, L as LoopRunnerCliArgs, e as LoopRunnerCliResult, R as ResearchLoopResult, f as ResearchLoopRunnerOptions, g as RunDelegatedLoopOptions, V as VetoedFact, W as WorktreeLoopRunnerOptions, h as auditLoopRunner, i as coderLoopRunner, j as isDelegatedLoopMode, p as parseLoopRunnerArgv, r as researchLoopRunner, k as runDelegatedLoop, l as runLoopRunnerCli, s as selfImproveLoopRunner, w as worktreeLoopRunner } from './loop-runner-bin-B0NeLTRd.js';
-export { m as mcpToolsForRuntimeMcp, a as mcpToolsForRuntimeMcpSubset } from './openai-tools-DPx9Gzvn.js';
+import { S as SurfaceImprovementEdit } from './improvement-adapter-CioiEE2z.js';
+import { I as ImprovementAdapter } from './types-BC3bZpH0.js';
+export { C as CoderLoopRunnerOptions, D as DELEGATED_LOOP_MODES, a as DelegatedLoopMode, b as DelegatedLoopRegistry, c as DelegatedLoopResult, d as DelegatedLoopRunner, L as LoopRunnerCliArgs, e as LoopRunnerCliResult, R as ResearchLoopResult, f as ResearchLoopRunnerOptions, g as RunDelegatedLoopOptions, V as VetoedFact, W as WorktreeLoopRunnerOptions, h as auditLoopRunner, i as coderLoopRunner, j as isDelegatedLoopMode, p as parseLoopRunnerArgv, r as researchLoopRunner, k as runDelegatedLoop, l as runLoopRunnerCli, s as selfImproveLoopRunner, w as worktreeLoopRunner } from './loop-runner-bin-DLM_bVQO.js';
+export { m as mcpToolsForRuntimeMcp, a as mcpToolsForRuntimeMcpSubset } from './openai-tools-kdCS-T12.js';
 import '@tangle-network/sandbox';
 import 'node:child_process';
 import './substrate-BoRXgvka.js';
 import './kb-gate-CuzMYGYM.js';
-import './worktree-fanout-gNfl0Byj.js';
+import './worktree-fanout-CK2ypmEm.js';
 
 /**
  * @stable

package/dist/index.js

@@ -1,7 +1,3 @@
-import {
-  mcpToolsForRuntimeMcp,
-  mcpToolsForRuntimeMcpSubset
-} from "./chunk-EDCVUZZC.js";
 import {
   DEFAULT_ROUTER_BASE_URL,
   cleanModelId,
@@ -21,26 +17,22 @@ import {
   runLoopRunnerCli,
   selfImproveLoopRunner,
   worktreeLoopRunner
-} from "./chunk-BYZCXQHF.js";
+} from "./chunk-K3RM4MPM.js";
+import "./chunk-P5OKDSLB.js";
+import {
+  mcpToolsForRuntimeMcp,
+  mcpToolsForRuntimeMcpSubset
+} from "./chunk-4KGQHS7U.js";
 import "./chunk-FNMGYYSS.js";
-import "./chunk-L5ZFBVT6.js";
+import "./chunk-INXDNX2W.js";
 import {
-  AgentEvalError,
-  BackendTransportError,
-  ConfigError,
-  JudgeError,
-  NotFoundError,
-  PlannerError,
-  RuntimeRunStateError,
-  SessionMismatchError,
-  ValidationError,
   assertModelAllowed,
   composeRuntimeHooks,
   defineRuntimeHooks,
   notifyRuntimeDecisionPoint,
   notifyRuntimeHookEvent,
   runLocalHarness
-} from "./chunk-QXWGSDAQ.js";
+} from "./chunk-5ISW5JUF.js";
 import {
   INTELLIGENCE_WIRE_VERSION,
   buildLoopOtelSpans,
@@ -48,7 +40,18 @@ import {
   createOtelExporter,
   exportEvalRuns,
   loopEventToOtelSpan
-} from "./chunk-G3RGMA7C.js";
+} from "./chunk-VMNEQHJR.js";
+import {
+  AgentEvalError,
+  BackendTransportError,
+  ConfigError,
+  JudgeError,
+  NotFoundError,
+  PlannerError,
+  RuntimeRunStateError,
+  SessionMismatchError,
+  ValidationError
+} from "./chunk-VLF5RHEQ.js";
 import "./chunk-DVQGYDN5.js";
 import "./chunk-WIR4HOOJ.js";
 import "./chunk-DGUM43GV.js";