@tangle-network/agent-eval 0.123.8 → 0.125.0

package/CHANGELOG.md

@@ -6,7 +6,7 @@ All notable changes to `@tangle-network/agent-eval` and its sibling `agent-eval-
 
 ## [Unreleased]
 
-### Added
+### Changed
 
 - `selfImprove({ budget: { candidateConcurrency } })` exposes the existing `runOptimization()` control for scoring candidate campaigns in parallel; it remains opt-in and defaults to one candidate campaign at a time.
 - `llmPolicyEditProposer()` and `projectPolicyEditHistory()` accept `scenarioOrder: 'input'` when controlled comparisons must preserve first-occurrence caller order; ranked evidence selection remains the default.
@@ -14,9 +14,6 @@ All notable changes to `@tangle-network/agent-eval` and its sibling `agent-eval-
 - `llmPolicyEditProposer({ redactCurrentSurfaceForModel })` can remove credentials and unrelated fields from the current surface sent to the model while applying validated edits to the complete original surface.
 - `CostLedger.listPending()` exposes immutable pending paid calls and distinguishes calls that are active, late after cancellation, or interrupted by a prior process so durable workflows can reconcile exact reservations before resuming.
 - `traceAnalystProposer()` accepts an opt-in `resolvePriorFindings` callback that forwards canonical prior findings into the existing analyst registry.
-
-### Changed
-
 - Trace-analysis actors are instructed to emit one executable JavaScript program per turn, report named turn-limit exhaustion, and preserve per-analyst failure details when a proposer produces no findings.
 - Keep one live tip per lineage track when another track branches or merges from it, and compare those track tips when building the frontier.
 - Pass track identity, operation, vision, ancestry, and proposer choice to candidate generation so independent tracks can pursue distinct strategies.
@@ -53,6 +50,34 @@ All notable changes to `@tangle-network/agent-eval` and its sibling `agent-eval-
 - `runSkillOpt({ holdoutScenarios })` fails closed because those rows are adaptively reused.
   Pass `selectionScenarios`; selection result fields now use `Selection` instead of `Holdout`, and `lift` is now `selectionLift`.
 
+## [0.125.0] — 2026-07-24 — Claude Code supervision reader and path-bound policy edits
+
+### Added
+
+- `claudeCodeSupervisorRunReader()` reads Claude Code root and subagent transcripts into the existing supervision-tree report and rollout formats, while declaring source limits instead of fabricating unavailable spend, verdict, or deliverable data.
+- `llmPolicyEditProposer({ valueSchemaByJsonPath })` binds each allowed edit path to an exact JSON Schema in both the provider response contract and local admission.
+
+### Fixed
+
+- Deduplicate repeated Claude Code worker cancellation events and preserve available worker cache-token counts.
+- Reject policy-edit values that do not match their target path before a candidate can enter a campaign.
+
+## [0.124.0] — 2026-07-24 — rollout and supervisor-run subpaths
+
+### Added
+
+- `callLlm()` and `callLlmJson()` accept request-level or client-default `thinking: 'enabled' | 'disabled'`; GEPA, SkillOpt, and policy-edit authors expose the same per-proposer override, and the exact mode is preserved in cost bounds, raw request capture, and provider traffic.
+- `@tangle-network/agent-eval/rollout` subpath: the single owner of the `tangle.rollout.v1` serialization — canonical schema + fail-loud validation, ledger file API (`writeRolloutLedger`/`appendRolloutLines`/`readRolloutLedger`), harness-store intake readers (opencode sqlite, Claude Code project jsonl), exporters (SFT, reward rows, Prime Intellect verifiers `RolloutOutput`, OpenAI RFT), deterministic 9-rule scrubber, HuggingFace dataset-card generation, and the `agent-eval rollout-release` CLI (build + optional `--push`). Ported from the agent-runtime bench rollout-ledger and reconciled with the PR #410 row shape; see `docs/rollout.md` for the decision table.
+- `@tangle-network/agent-eval/supervisor-run` subpath: supervision-tree analysis alongside single-rollout trace analysis. `analyzeSupervisorRun(runDir | reader | sources)` returns a `SupervisorRunReport` — steer count with per-worker breakdown, spawn waves + sizes, max concurrency, respawns/repeated labels/delegation depth, supervisor wall + idle wall + worker utilization, accepted vs rejected vs empty-pass, evidence→respawn vs blind respawn, tokens/USD by role, judge verdict + patch stats — and `rollupSupervisorRuns` aggregates across runs. Every metric is `Measured<T> = T | {unavailable: reason}`, so a missing artifact never reads as a measured zero. The input contract is a `SupervisorRunReader` over already-read bytes; `loopsSupervisorRunReader` is one implementation (the loops `.loops/supervisor/*` on-disk layout). `supervisorRunRolloutLines` emits the tree as `tangle.rollout.v1` rows keyed by `parent_rollout_id`, so a supervision tree lands in the same ledger as solo rollouts. Ported from the agent-runtime bench run-report; byte-identical on both committed backfill fixtures.
+
+### Changed
+
+- `mintRolloutRows` now emits canonical `tangle.rollout.v1` lines (snake_case wire shape) instead of the interim `RolloutRow`; records without trace spans become labeled gap lines AND are listed in `missingTraces` instead of being skipped. `toSftRows`/`toRewardRows` operate on the new lines; `toSftRows` additionally enforces the trainable-split filter (holdout/dev/canary never export). The realness gate still forces reward 0 and SFT exclusion.
+
+### Fixed
+
+- `openOpencodeDb` composes the `node:sqlite` specifier at runtime so neither esbuild nor Vite rewrites it, keeping the opencode rollout reader working in the bundled package.
+
 ## [0.123.8] — 2026-07-23 — reasoning-token accounting
 
 ### Fixed

package/dist/analyst/index.d.ts

@@ -1020,6 +1020,7 @@ interface LlmMessage {
         };
     }>;
 }
+type LlmThinkingMode = 'enabled' | 'disabled';
 interface LlmCallRequest {
     model: string;
     messages: LlmMessage[];
@@ -1032,6 +1033,8 @@ interface LlmCallRequest {
     };
     temperature?: number;
     maxTokens?: number;
+    /** OpenAI-compatible reasoning mode. Omitted when the provider default should apply. */
+    thinking?: LlmThinkingMode;
     /** Per-call timeout, default 300s. */
     timeoutMs?: number;
 }
@@ -1125,6 +1128,8 @@ interface LlmClientOptions {
      * Default: `extract`.
      */
     jsonPayloadMode?: 'extract' | 'exact';
+    /** Default provider reasoning mode. A per-call request value takes precedence. */
+    thinking?: LlmThinkingMode;
     /** Fetch implementation — defaults to global `fetch`. Override for custom transport (e.g. tests). */
     fetch?: typeof fetch;
     /**

package/dist/analyst/index.js

@@ -9,13 +9,13 @@ import {
   diffFindings,
   emitSkillUsageFindings,
   runSemanticConceptJudge
-} from "../chunk-RQ5TP2TV.js";
+} from "../chunk-W5B3ZGP3.js";
 import {
   behavioralAnalyst,
   buildDefaultAnalystRegistry,
   createChatClient,
   deriveEfficiencyFindings
-} from "../chunk-BUR5R4R4.js";
+} from "../chunk-A6GT67HT.js";
 import "../chunk-HHWE3POT.js";
 import {
   ANALYST_SEVERITIES,
@@ -71,8 +71,8 @@ import {
   validatePolicyEdit,
   validatePolicyEditCandidateRecord,
   validateUsageSettlementTimeout
-} from "../chunk-EEHHBAWH.js";
-import "../chunk-PDHIOKRE.js";
+} from "../chunk-DT7OXY3C.js";
+import "../chunk-EQUK3RFS.js";
 import {
   CostLedger
 } from "../chunk-GRCDRKII.js";

package/dist/benchmarks/index.js

@@ -17,12 +17,12 @@ import {
   runBenchmarkAdapter,
   summarizeBenchmarkCampaign
 } from "../chunk-JKDNAOF5.js";
-import "../chunk-ZU3QWGZE.js";
-import "../chunk-QVGVJQMR.js";
+import "../chunk-A62YMFWA.js";
+import "../chunk-PMITBABE.js";
 import "../chunk-D5JZ7UDZ.js";
-import "../chunk-EEHHBAWH.js";
+import "../chunk-DT7OXY3C.js";
 import "../chunk-ARU2PZFM.js";
-import "../chunk-PDHIOKRE.js";
+import "../chunk-EQUK3RFS.js";
 import "../chunk-PJQFMIOX.js";
 import "../chunk-GRCDRKII.js";
 import "../chunk-VI2UW6B6.js";

package/dist/campaign/index.d.ts

@@ -1113,6 +1113,7 @@ interface LlmMessage {
         };
     }>;
 }
+type LlmThinkingMode = 'enabled' | 'disabled';
 interface LlmCallRequest {
     model: string;
     messages: LlmMessage[];
@@ -1125,6 +1126,8 @@ interface LlmCallRequest {
     };
     temperature?: number;
     maxTokens?: number;
+    /** OpenAI-compatible reasoning mode. Omitted when the provider default should apply. */
+    thinking?: LlmThinkingMode;
     /** Per-call timeout, default 300s. */
     timeoutMs?: number;
 }
@@ -1219,6 +1222,8 @@ interface LlmClientOptions {
      * Default: `extract`.
      */
     jsonPayloadMode?: 'extract' | 'exact';
+    /** Default provider reasoning mode. A per-call request value takes precedence. */
+    thinking?: LlmThinkingMode;
     /** Fetch implementation — defaults to global `fetch`. Override for custom transport (e.g. tests). */
     fetch?: typeof fetch;
     /**
@@ -5455,6 +5460,8 @@ interface SkillOptProposerOptions {
     editBudget?: number;
     temperature?: number;
     maxTokens?: number;
+    /** Provider reasoning mode for each patch-authoring call. Omitted to preserve the provider default. */
+    thinking?: LlmCallRequest['thinking'];
     /** Top-K weak scenarios/dimensions surfaced as evidence. Default 3. */
     evidenceK?: number;
 }
@@ -5706,6 +5713,8 @@ interface GepaProposerOptions {
     temperature?: number;
     /** Reflection max tokens. Default 6000. */
     maxTokens?: number;
+    /** Provider reasoning mode for each reflection call. Omitted to preserve the provider default. */
+    thinking?: LlmCallRequest['thinking'];
     /** Structured-doc constraints. Candidates violating any are rejected
      *  post-parse and dropped from the returned population. */
     constraints?: GepaProposerConstraints;
@@ -5832,6 +5841,8 @@ type PolicyEditFindingSource = {
     kind: 'global';
     label: string;
 };
+/** Exact JSON Schema applied to authored `change.value` for one allowed path. */
+type PolicyEditJsonValueSchema = Readonly<Record<string, unknown>>;
 /** Trace-derived findings must name the measured profile that produced them.
  *  Cross-run doctrine can be explicitly global; an unwrapped finding is rejected. */
 interface PolicyEditFindingInput {
@@ -5927,6 +5938,12 @@ interface LlmPolicyEditProposerOptions {
     targetSurface: JsonPolicyEditTargetSurface;
     /** Exact JSON paths the author may change. Prefix or fuzzy matches are not accepted. */
     allowedJsonPaths: readonly string[];
+    /**
+     * Optional exact JSON Schema for `change.value` at every allowed path.
+     * When present, keys must exactly match `allowedJsonPaths`; both the provider
+     * response schema and local validation bind each path to its own value shape.
+     */
+    valueSchemaByJsonPath?: Readonly<Record<string, PolicyEditJsonValueSchema>>;
     /** Caller-owned search objective bound into every authored forecast. */
     objectives: readonly PolicyEditObjective[];
     /** Default: evidence-only, so uncertain edits are measured rather than
@@ -5937,6 +5954,8 @@ interface LlmPolicyEditProposerOptions {
     maxCandidates?: number;
     temperature?: number;
     maxTokens?: number;
+    /** Provider reasoning mode for each policy-authoring call. Omitted to preserve the provider default. */
+    thinking?: LlmCallRequest['thinking'];
     timeoutMs?: number;
     /** Number of most recent scored generations sent to the author. Default: 4. */
     maxHistoryGenerations?: number;
@@ -7754,4 +7773,4 @@ declare function verifyCodeSurface(surface: CodeSurface, worktreeDir?: string):
  *  identity against the checkout at `worktreeRef`. */
 declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
 
-export { type AcceptedEdit, type AceProposerOptions, type AnalystArtifact, type AnalystScenario, type AnalyzeCrossSurfaceInteractionsInput, type AnalyzeOtlpTraceFileOptions, type ApplySkillPatchResult, type AxisEvidence, type AxisVerdict, type BuildAnalystSurfaceDispatchOptions, type BuildEvidenceVectorOptions, type BuildLoopProvenanceArgs, type BuiltinOptimizationMethodConfig, type CampaignAggregates, type CampaignArtifactWriter, type CampaignBreakdown, type CampaignCellResult, type CampaignCostMeter, type CampaignResult, type CampaignRunPlan, type CampaignRunPlanCell, type CampaignScenarioIdentity, type CampaignStorage, type CampaignTokenUsage, type CampaignTraceWriter, type CodeSurface, type CodeSurfaceVerification, type CompareOptimizationMethodsOptions, type ComparisonCost, type CompositeProposerOptions, type CostLedgerHandle, type CrossSurfaceAdditionDecision, type CrossSurfaceAdditionRejectionReason, type CrossSurfaceAttemptCompleteness, type CrossSurfaceBestSingleSelection, type CrossSurfaceBootstrapPolicy, type CrossSurfaceCandidate, type CrossSurfaceCandidateComparison, type CrossSurfaceCandidateEvidence, type CrossSurfaceCandidateOutcome, type CrossSurfaceCandidateSummary, type CrossSurfaceComponent, type CrossSurfaceComponentEvidence, type CrossSurfaceCompositionStep, type CrossSurfaceDistribution, type CrossSurfaceEligibility, type CrossSurfaceEvidenceBreakdown, type CrossSurfaceIneligibilityReason, type CrossSurfaceInteractionAwareSelection, type CrossSurfaceInteractionEffect, type CrossSurfaceInteractionPath, type CrossSurfaceInteractionReport, type CrossSurfaceInteractionTask, type CrossSurfaceNaiveStackSelection, type CrossSurfacePairCompatibility, type CrossSurfacePairEvidence, type CrossSurfacePairIncompatibilityReason, type CrossSurfacePairwiseEntry, type CrossSurfaceRankedSingle, type CrossSurfaceRelativeCost, type CrossSurfaceSelectionPolicy, type CrossSurfaceSelections, type CrossSurfaceTaskRow, DEFAULT_POLICY_EDIT_HISTORY_LIMITS, type DefaultProductionGateOptions, type DimensionRegression, type DiscriminationScore, type DispatchContext, type DispatchFn, type EmitLoopProvenanceArgs, type EmitLoopProvenanceResult, type EvalFixture, type EvalFixtureFile, type EvalFixtureLoadOptions, type EvalFixtureRunPlan, type EvalFixtureScenario, type EvalFixtureValidationMode, type EvidenceVector, type EvolutionaryProposerOptions, type FailureModeRecallJudgeOptions, type FapoAttributionSignals, type FapoFailureCluster, type FapoOptimizationLevel, type FapoOptimizationMethodConfig, type FapoProposerOptions, type FapoReviewInput, type FapoReviewIssue, type FapoReviewResult, type FapoScopeContract, FileSearchLedger, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, type Gate, type GateContext, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaProposerConstraints, type GepaProposerOptions, type GepaRunnerCommand, type GitWorktreeAdapterOptions, type Governor, type GovernorContext, type GovernorOp, type HaloProposerOptions, type HeldOutGateOptions, type HeldoutSignificance, type HeldoutSignificanceOptions, type HeuristicGovernorOptions, type JsonPolicyEditTargetSurface, type JsonPrimitive, type JsonValue, type JudgeAggregate, type JudgeConfig, type JudgeDimension, type JudgeScore, type LabelTrust, type LabeledScenarioRecord, type LabeledScenarioSampleArgs, type LabeledScenarioSource, type LabeledScenarioStore, LabeledScenarioStoreError, type LabeledScenarioWrite, Lineage, type LineageEdge, type LineageGraph, type LineageNode, type LineageNodeInput, type LineageStore, LineageStoreConflictError, type LlmJudgeDimension, type LlmJudgeOptions, type LlmPolicyEditProposerOptions, type LoadEvalFixtureScenariosOptions, type LoopProvenanceArgsFromResult, type LoopProvenanceBackend, type LoopProvenanceCandidate, type LoopProvenanceEvidence, type LoopProvenanceRecord, type MemoryCurationProposerOptions, type MutableSurface, type Mutator, type NeutralizationGateOptions, type ObjectiveSource, type OpenAutoPrOptions, type OpenAutoPrResult, type OpenSearchLedgerOptions, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodPairwise, type OptimizationMethodResult, type OptimizationMethodRunOptions, type OptimizationMethodScore, type OptimizationProposer, type OptimizerConfig, POLICY_EDIT_CANDIDATE_RECORD_SCHEMA, type PairedHoldout, type ParameterCandidate, type ParameterChange, type ParameterSweepProposerOptions, type ParetoParent, type ParetoSignificanceGateOptions, type PendingCostCallView, type PlanCampaignRunOptions, type PlanEvalFixtureRunOptions, type PlaybackContext, type PlaybackDriver, type PlaybackStep, type PolicyEditAuthorScenarioOrder, type PolicyEditAuthorScenarioRow, type PolicyEditCandidateRecord, type PolicyEditCandidateSummary, type PolicyEditFindingInput, type PolicyEditFindingSource, type PolicyEditHistoryCandidateContext, type PolicyEditHistoryGenerationContext, type PolicyEditHistoryProjectionOptions, type PolicyEditObjective, type PolicyEditOutcomeContext, type PolicyEditProposerOptions, type PowerPreflight, type PowerPreflightOptions, type PremeasuredOptimizationBaseline, type ProfileDispatchFn, ProfileMatrixError, type ProfileSummary, type PromotionObjective, type PromotionPolicy, type ProposalTrackContext, type ProposeContext, type ProposePatchesArgs, type ProposedCandidate, type RedactionStatus, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceScenario, type RejectedEdit, type RolloutArgumentDiff, type RolloutArgumentDiffOptions, type RolloutCall, type RunCampaignOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunLineageLoopOptions, type RunLineageLoopResult, type RunLineageLoopSeed, type RunLineageOptions, type RunLineageResult, type RunLineageSeed, type RunLineageStepResult, type RunOptimizationOptions, type RunOptimizationResult, type RunProfileMatrixOptions, type RunProfileMatrixResult, type RunSkillOptOptions, type RunSkillOptResult, SEARCH_LEDGER_SCHEMA, type Scenario, type ScenarioAggregate, type ScenarioRollup, type ScenarioSignal, type ScoreboardRenderOptions, type ScoreboardRow, type ScoreboardSummary, type ScoredRollout, type ScoredSurfaceOutcome, type SearchAccountingAudit, type SearchArtifactRef, type SearchAttemptAccounting, type SearchCandidateDecidedEvent, type SearchCandidateLineage, type SearchCandidateRegisteredEvent, type SearchCandidateSlot, type SearchCandidateSlotClosedEvent, type SearchCandidateSurface, type SearchCompletedEvent, type SearchCostAccounting, type SearchFailureReason, type SearchLedger, type SearchLedgerAppendResult, SearchLedgerConflictError, type SearchLedgerEntry, SearchLedgerError, type SearchLedgerEvent, type SearchLedgerHash, SearchLedgerIntegrityError, type SearchLedgerReplay, type SearchModelIdentity, type SearchOperationKind, type SearchOperationRecordedEvent, type SearchPlan, type SearchPlannedEvent, type SearchPlannedOperation, type SearchPlannedTask, type SearchSourceRef, type SearchSurfaceEffect, type SearchSurfaceEvidence, type SearchSurfaceKind, type SearchTaskAttemptedEvent, type SearchTaskOutcome, type SearchTokenAccounting, type SelectPolicyEditAuthorRowsOptions, type SequentialDecideFn, type SequentialDecideOptions, type SequentialDecision, type SequentialObservation, type SequentialPairedGate, type SequentialPairedGateOptions, type SerializedJsonBudget, type SessionScript, type SingleRunLock, type SingleRunLockOptions, type SkillOptEpochRecord, type SkillOptEvidence, type SkillOptProposer, type SkillOptProposerOptions, type SkillPatch, type SkillPatchOp, SkillPatchParseError, type SkillPatchRejection, type SurfaceProposer, type SurfaceScore, type TraceAnalystPriorFindings, type TraceAnalystProposerOptions, type TraceSpan, type TransientFailureOptions, type UngroundedLiteralReport, type UserStory, type UserStoryVerdict, type Worktree, type WorktreeAdapter, WorktreeAdapterError, aceProposer, acquireSingleRunLock, analyzeCrossSurfaceInteractions, analyzeOtlpTraceFile, applySkillPatch, assertCampaignDesign, assertCampaignSplitIdentity, assertCodeSurfaceIdentity, assertPolicyEditAuthorContextBudget, buildAnalystSurfaceDispatch, buildEvidenceVector, buildLoopProvenanceRecord, callbackGovernor, campaignBreakdown, campaignLineageStore, campaignMeanComposite, campaignMeasurementDigest, campaignScenarioIdentity, campaignSplitDigest, campaignSplitDigestFromIdentities, canonicalDigest, classifyUngroundedLiterals, codeSurfaceIdentityMaterial, compareOptimizationMethods, composeGate, compositeProposer, costFromLedgerSummary, countSentenceEdits, createReferenceEquivalenceJudge, createRunCostLedger, defaultProductionGate, detectScale, dimensionRegressions, discoverEvalFixtures, emitLoopProvenance, evolutionaryProposer, extractFapoAttributionSignals, extractH2Sections, failureModeRecallJudge, fapoEscalationMethod, fapoProposer, fsCampaignStorage, fsLineageStore, gepaOptimizationMethod, gepaParetoMethod, gepaProposer, gepaReflectionMethod, gitWorktreeAdapter, haloProposer, heldOutGate, heldoutSignificance, heuristicGovernor, inMemoryCampaignStorage, isProposedCandidate, isTransientTransportFailure, labelTrustRank, lineageNodeId, llmJudge, llmPolicyEditProposer, loadEvalFixture, loadEvalFixtureScenarios, loopProvenanceArgsFromResult, loopProvenanceSpans, makePlaybackDispatch, memLineageStore, memoryCurationProposer, neutralizationGate, neutralizeText, openAutoPr, openSearchLedger, pairHoldout, parameterSweepProposer, paretoPolicy, paretoSignificanceGate, parseSkillPatchResponse, patchEditCount, planCampaignRun, planEvalFixtureRun, policyEditProposer, powerPreflight, projectPolicyEditHistory, provenanceRecordPath, provenanceSpansPath, renderScoreboardMarkdown, renderSurfaceDiff, resolveRunDir, resolveWorktreePath, rolloutArgumentDiff, runCampaign, runEval, runImprovementLoop, runLineage, runLineageLoop, runOptimization, runProfileMatrix, runSkillOpt, scoreDiscrimination, scoreUserStory, scoreboardSummary, selectDiscriminative, selectPolicyEditAuthorRows, sequentialDecide, sequentialPairedGate, skillOptMethod, skillOptProposer, surfaceContentHash, surfaceHash, tangleTracesRoot, traceAnalystProposer, userStoryScoreboard, validatePolicyEditCandidateRecord, validateSearchLedgerEvent, verifyCodeSurface, verifyLoopProvenanceRecord };
+export { type AcceptedEdit, type AceProposerOptions, type AnalystArtifact, type AnalystScenario, type AnalyzeCrossSurfaceInteractionsInput, type AnalyzeOtlpTraceFileOptions, type ApplySkillPatchResult, type AxisEvidence, type AxisVerdict, type BuildAnalystSurfaceDispatchOptions, type BuildEvidenceVectorOptions, type BuildLoopProvenanceArgs, type BuiltinOptimizationMethodConfig, type CampaignAggregates, type CampaignArtifactWriter, type CampaignBreakdown, type CampaignCellResult, type CampaignCostMeter, type CampaignResult, type CampaignRunPlan, type CampaignRunPlanCell, type CampaignScenarioIdentity, type CampaignStorage, type CampaignTokenUsage, type CampaignTraceWriter, type CodeSurface, type CodeSurfaceVerification, type CompareOptimizationMethodsOptions, type ComparisonCost, type CompositeProposerOptions, type CostLedgerHandle, type CrossSurfaceAdditionDecision, type CrossSurfaceAdditionRejectionReason, type CrossSurfaceAttemptCompleteness, type CrossSurfaceBestSingleSelection, type CrossSurfaceBootstrapPolicy, type CrossSurfaceCandidate, type CrossSurfaceCandidateComparison, type CrossSurfaceCandidateEvidence, type CrossSurfaceCandidateOutcome, type CrossSurfaceCandidateSummary, type CrossSurfaceComponent, type CrossSurfaceComponentEvidence, type CrossSurfaceCompositionStep, type CrossSurfaceDistribution, type CrossSurfaceEligibility, type CrossSurfaceEvidenceBreakdown, type CrossSurfaceIneligibilityReason, type CrossSurfaceInteractionAwareSelection, type CrossSurfaceInteractionEffect, type CrossSurfaceInteractionPath, type CrossSurfaceInteractionReport, type CrossSurfaceInteractionTask, type CrossSurfaceNaiveStackSelection, type CrossSurfacePairCompatibility, type CrossSurfacePairEvidence, type CrossSurfacePairIncompatibilityReason, type CrossSurfacePairwiseEntry, type CrossSurfaceRankedSingle, type CrossSurfaceRelativeCost, type CrossSurfaceSelectionPolicy, type CrossSurfaceSelections, type CrossSurfaceTaskRow, DEFAULT_POLICY_EDIT_HISTORY_LIMITS, type DefaultProductionGateOptions, type DimensionRegression, type DiscriminationScore, type DispatchContext, type DispatchFn, type EmitLoopProvenanceArgs, type EmitLoopProvenanceResult, type EvalFixture, type EvalFixtureFile, type EvalFixtureLoadOptions, type EvalFixtureRunPlan, type EvalFixtureScenario, type EvalFixtureValidationMode, type EvidenceVector, type EvolutionaryProposerOptions, type FailureModeRecallJudgeOptions, type FapoAttributionSignals, type FapoFailureCluster, type FapoOptimizationLevel, type FapoOptimizationMethodConfig, type FapoProposerOptions, type FapoReviewInput, type FapoReviewIssue, type FapoReviewResult, type FapoScopeContract, FileSearchLedger, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, type Gate, type GateContext, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaProposerConstraints, type GepaProposerOptions, type GepaRunnerCommand, type GitWorktreeAdapterOptions, type Governor, type GovernorContext, type GovernorOp, type HaloProposerOptions, type HeldOutGateOptions, type HeldoutSignificance, type HeldoutSignificanceOptions, type HeuristicGovernorOptions, type JsonPolicyEditTargetSurface, type JsonPrimitive, type JsonValue, type JudgeAggregate, type JudgeConfig, type JudgeDimension, type JudgeScore, type LabelTrust, type LabeledScenarioRecord, type LabeledScenarioSampleArgs, type LabeledScenarioSource, type LabeledScenarioStore, LabeledScenarioStoreError, type LabeledScenarioWrite, Lineage, type LineageEdge, type LineageGraph, type LineageNode, type LineageNodeInput, type LineageStore, LineageStoreConflictError, type LlmJudgeDimension, type LlmJudgeOptions, type LlmPolicyEditProposerOptions, type LoadEvalFixtureScenariosOptions, type LoopProvenanceArgsFromResult, type LoopProvenanceBackend, type LoopProvenanceCandidate, type LoopProvenanceEvidence, type LoopProvenanceRecord, type MemoryCurationProposerOptions, type MutableSurface, type Mutator, type NeutralizationGateOptions, type ObjectiveSource, type OpenAutoPrOptions, type OpenAutoPrResult, type OpenSearchLedgerOptions, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodPairwise, type OptimizationMethodResult, type OptimizationMethodRunOptions, type OptimizationMethodScore, type OptimizationProposer, type OptimizerConfig, POLICY_EDIT_CANDIDATE_RECORD_SCHEMA, type PairedHoldout, type ParameterCandidate, type ParameterChange, type ParameterSweepProposerOptions, type ParetoParent, type ParetoSignificanceGateOptions, type PendingCostCallView, type PlanCampaignRunOptions, type PlanEvalFixtureRunOptions, type PlaybackContext, type PlaybackDriver, type PlaybackStep, type PolicyEditAuthorScenarioOrder, type PolicyEditAuthorScenarioRow, type PolicyEditCandidateRecord, type PolicyEditCandidateSummary, type PolicyEditFindingInput, type PolicyEditFindingSource, type PolicyEditHistoryCandidateContext, type PolicyEditHistoryGenerationContext, type PolicyEditHistoryProjectionOptions, type PolicyEditJsonValueSchema, type PolicyEditObjective, type PolicyEditOutcomeContext, type PolicyEditProposerOptions, type PowerPreflight, type PowerPreflightOptions, type PremeasuredOptimizationBaseline, type ProfileDispatchFn, ProfileMatrixError, type ProfileSummary, type PromotionObjective, type PromotionPolicy, type ProposalTrackContext, type ProposeContext, type ProposePatchesArgs, type ProposedCandidate, type RedactionStatus, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceScenario, type RejectedEdit, type RolloutArgumentDiff, type RolloutArgumentDiffOptions, type RolloutCall, type RunCampaignOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunLineageLoopOptions, type RunLineageLoopResult, type RunLineageLoopSeed, type RunLineageOptions, type RunLineageResult, type RunLineageSeed, type RunLineageStepResult, type RunOptimizationOptions, type RunOptimizationResult, type RunProfileMatrixOptions, type RunProfileMatrixResult, type RunSkillOptOptions, type RunSkillOptResult, SEARCH_LEDGER_SCHEMA, type Scenario, type ScenarioAggregate, type ScenarioRollup, type ScenarioSignal, type ScoreboardRenderOptions, type ScoreboardRow, type ScoreboardSummary, type ScoredRollout, type ScoredSurfaceOutcome, type SearchAccountingAudit, type SearchArtifactRef, type SearchAttemptAccounting, type SearchCandidateDecidedEvent, type SearchCandidateLineage, type SearchCandidateRegisteredEvent, type SearchCandidateSlot, type SearchCandidateSlotClosedEvent, type SearchCandidateSurface, type SearchCompletedEvent, type SearchCostAccounting, type SearchFailureReason, type SearchLedger, type SearchLedgerAppendResult, SearchLedgerConflictError, type SearchLedgerEntry, SearchLedgerError, type SearchLedgerEvent, type SearchLedgerHash, SearchLedgerIntegrityError, type SearchLedgerReplay, type SearchModelIdentity, type SearchOperationKind, type SearchOperationRecordedEvent, type SearchPlan, type SearchPlannedEvent, type SearchPlannedOperation, type SearchPlannedTask, type SearchSourceRef, type SearchSurfaceEffect, type SearchSurfaceEvidence, type SearchSurfaceKind, type SearchTaskAttemptedEvent, type SearchTaskOutcome, type SearchTokenAccounting, type SelectPolicyEditAuthorRowsOptions, type SequentialDecideFn, type SequentialDecideOptions, type SequentialDecision, type SequentialObservation, type SequentialPairedGate, type SequentialPairedGateOptions, type SerializedJsonBudget, type SessionScript, type SingleRunLock, type SingleRunLockOptions, type SkillOptEpochRecord, type SkillOptEvidence, type SkillOptProposer, type SkillOptProposerOptions, type SkillPatch, type SkillPatchOp, SkillPatchParseError, type SkillPatchRejection, type SurfaceProposer, type SurfaceScore, type TraceAnalystPriorFindings, type TraceAnalystProposerOptions, type TraceSpan, type TransientFailureOptions, type UngroundedLiteralReport, type UserStory, type UserStoryVerdict, type Worktree, type WorktreeAdapter, WorktreeAdapterError, aceProposer, acquireSingleRunLock, analyzeCrossSurfaceInteractions, analyzeOtlpTraceFile, applySkillPatch, assertCampaignDesign, assertCampaignSplitIdentity, assertCodeSurfaceIdentity, assertPolicyEditAuthorContextBudget, buildAnalystSurfaceDispatch, buildEvidenceVector, buildLoopProvenanceRecord, callbackGovernor, campaignBreakdown, campaignLineageStore, campaignMeanComposite, campaignMeasurementDigest, campaignScenarioIdentity, campaignSplitDigest, campaignSplitDigestFromIdentities, canonicalDigest, classifyUngroundedLiterals, codeSurfaceIdentityMaterial, compareOptimizationMethods, composeGate, compositeProposer, costFromLedgerSummary, countSentenceEdits, createReferenceEquivalenceJudge, createRunCostLedger, defaultProductionGate, detectScale, dimensionRegressions, discoverEvalFixtures, emitLoopProvenance, evolutionaryProposer, extractFapoAttributionSignals, extractH2Sections, failureModeRecallJudge, fapoEscalationMethod, fapoProposer, fsCampaignStorage, fsLineageStore, gepaOptimizationMethod, gepaParetoMethod, gepaProposer, gepaReflectionMethod, gitWorktreeAdapter, haloProposer, heldOutGate, heldoutSignificance, heuristicGovernor, inMemoryCampaignStorage, isProposedCandidate, isTransientTransportFailure, labelTrustRank, lineageNodeId, llmJudge, llmPolicyEditProposer, loadEvalFixture, loadEvalFixtureScenarios, loopProvenanceArgsFromResult, loopProvenanceSpans, makePlaybackDispatch, memLineageStore, memoryCurationProposer, neutralizationGate, neutralizeText, openAutoPr, openSearchLedger, pairHoldout, parameterSweepProposer, paretoPolicy, paretoSignificanceGate, parseSkillPatchResponse, patchEditCount, planCampaignRun, planEvalFixtureRun, policyEditProposer, powerPreflight, projectPolicyEditHistory, provenanceRecordPath, provenanceSpansPath, renderScoreboardMarkdown, renderSurfaceDiff, resolveRunDir, resolveWorktreePath, rolloutArgumentDiff, runCampaign, runEval, runImprovementLoop, runLineage, runLineageLoop, runOptimization, runProfileMatrix, runSkillOpt, scoreDiscrimination, scoreUserStory, scoreboardSummary, selectDiscriminative, selectPolicyEditAuthorRows, sequentialDecide, sequentialPairedGate, skillOptMethod, skillOptProposer, surfaceContentHash, surfaceHash, tangleTracesRoot, traceAnalystProposer, userStoryScoreboard, validatePolicyEditCandidateRecord, validateSearchLedgerEvent, verifyCodeSurface, verifyLoopProvenanceRecord };

package/dist/campaign/index.js

@@ -71,7 +71,7 @@ import {
   userStoryScoreboard,
   validateSearchLedgerEvent,
   verifyCodeSurface
-} from "../chunk-ZU3QWGZE.js";
+} from "../chunk-A62YMFWA.js";
 import {
   assertCodeSurfaceIdentity,
   buildEvidenceVector,
@@ -112,7 +112,7 @@ import {
   surfaceContentHash,
   surfaceHash,
   verifyLoopProvenanceRecord
-} from "../chunk-QVGVJQMR.js";
+} from "../chunk-PMITBABE.js";
 import {
   SearchLedgerConflictError,
   SearchLedgerError,
@@ -133,9 +133,9 @@ import {
 import {
   POLICY_EDIT_CANDIDATE_RECORD_SCHEMA,
   validatePolicyEditCandidateRecord
-} from "../chunk-EEHHBAWH.js";
+} from "../chunk-DT7OXY3C.js";
 import "../chunk-ARU2PZFM.js";
-import "../chunk-PDHIOKRE.js";
+import "../chunk-EQUK3RFS.js";
 import "../chunk-PJQFMIOX.js";
 import "../chunk-GRCDRKII.js";
 import "../chunk-VI2UW6B6.js";

package/dist/{chunk-ZU3QWGZE.js → chunk-A62YMFWA.js}

@@ -15,7 +15,7 @@ import {
   runImprovementLoop,
   surfaceContentHash,
   surfaceHash
-} from "./chunk-QVGVJQMR.js";
+} from "./chunk-PMITBABE.js";
 import {
   SearchLedgerConflictError,
   SearchLedgerError,
@@ -51,14 +51,14 @@ import {
   mapConcurrent,
   policyEditsFromFindings,
   validatePolicyEditCandidateRecord
-} from "./chunk-EEHHBAWH.js";
+} from "./chunk-DT7OXY3C.js";
 import {
   callLlm,
   callLlmJson,
   costReceiptFromLlm,
   costReceiptFromLlmError,
   maximumChargeForLlmRequest
-} from "./chunk-PDHIOKRE.js";
+} from "./chunk-EQUK3RFS.js";
 import {
   eProcess,
   mcnemar,
@@ -4106,7 +4106,8 @@ function skillOptProposer(opts) {
       ],
       jsonMode: true,
       temperature: opts.temperature ?? 0.6,
-      maxTokens: opts.maxTokens ?? 4e3
+      maxTokens: opts.maxTokens ?? 4e3,
+      thinking: opts.thinking
     };
     const paid = await (args.costLedger ?? directCostLedger).runPaidCall({
       channel: "driver",
@@ -6328,7 +6329,7 @@ var POLICY_EDIT_AUTHOR_JSON_SCHEMA = {
     }
   }
 };
-function policyEditAuthorJsonSchema(maxItems, allowedJsonPaths) {
+function policyEditAuthorJsonSchema(maxItems, allowedJsonPaths, valueSchemaByJsonPath) {
   const schema = JSON.parse(JSON.stringify(POLICY_EDIT_AUTHOR_JSON_SCHEMA));
   const properties = schema.properties;
   const edits = properties.edits;
@@ -6338,8 +6339,32 @@ function policyEditAuthorJsonSchema(maxItems, allowedJsonPaths) {
   const target = itemProperties.target;
   const targetProperties = target.properties;
   targetProperties.path = { type: "string", enum: [...allowedJsonPaths] };
+  if (valueSchemaByJsonPath) {
+    edits.items = {
+      anyOf: allowedJsonPaths.map(
+        (path) => policyEditAuthorItemJsonSchema(item, path, valueSchemaByJsonPath.get(path))
+      )
+    };
+  }
   return schema;
 }
+function policyEditAuthorItemJsonSchema(item, path, valueSchema) {
+  const branch = structuredClone(item);
+  const properties = branch.properties;
+  const target = properties.target;
+  const targetProperties = target.properties;
+  targetProperties.path = { const: path };
+  const change = properties.change;
+  const changeBranches = change.anyOf;
+  for (const changeBranch of changeBranches) {
+    const changeProperties = changeBranch.properties;
+    const mode = changeProperties.mode;
+    if (mode.const !== "remove") {
+      changeProperties.value = structuredClone(valueSchema);
+    }
+  }
+  return branch;
+}
 var DEFAULT_POLICY_EDIT_HISTORY_LIMITS = Object.freeze({
   generations: 4,
   candidatesPerGeneration: 16,
@@ -6369,6 +6394,7 @@ function policyEditAuthorSystem(responseSchema) {
 function llmPolicyEditProposer(opts) {
   const allowedJsonPaths = validateAllowedJsonPaths(opts.allowedJsonPaths);
   const allowedPathSet = new Set(allowedJsonPaths);
+  const pathValueSchemas = validatePathValueSchemas(opts.valueSchemaByJsonPath, allowedJsonPaths);
   const objectives = validateObjectives(opts.objectives);
   const objectiveByKey = new Map(objectives.map((objective) => [objective.key, objective]));
   requireNonEmpty(opts.model, "model");
@@ -6463,7 +6489,11 @@ function llmPolicyEditProposer(opts) {
           objectiveByKey
         )
       };
-      const responseSchema = policyEditAuthorJsonSchema(limit, allowedJsonPaths);
+      const responseSchema = policyEditAuthorJsonSchema(
+        limit,
+        allowedJsonPaths,
+        pathValueSchemas?.schemas
+      );
       const system = policyEditAuthorSystem(responseSchema);
       assertPolicyEditAuthorContextBudget(
         { system, authorContext, responseSchema },
@@ -6482,6 +6512,7 @@ function llmPolicyEditProposer(opts) {
         },
         temperature: opts.temperature ?? 0.2,
         maxTokens: opts.maxTokens ?? 6e3,
+        thinking: opts.thinking,
         timeoutMs: opts.timeoutMs
       };
       const paid = await (ctx.costLedger ?? directCostLedger).runPaidCall({
@@ -6515,6 +6546,7 @@ function llmPolicyEditProposer(opts) {
           findingByKey,
           opts.targetSurface,
           allowedPathSet,
+          pathValueSchemas?.validators,
           objectives[0],
           ctx.incumbentOutcome?.composite ?? ctx.baselineOutcome?.composite
         )
@@ -6961,12 +6993,23 @@ function parseAuthorResponse(value) {
   const issues = parsed.error.issues.map((issue) => `${issue.path.join(".") || "<root>"}: ${issue.message}`).join("; ");
   throw new Error(`llmPolicyEditProposer: invalid PolicyEdit response: ${issues}`);
 }
-function bindAuthoredEdit(draft, findingByKey, targetSurface, allowedPaths, objective, currentComposite) {
+function bindAuthoredEdit(draft, findingByKey, targetSurface, allowedPaths, valueValidatorByJsonPath, objective, currentComposite) {
   if (!allowedPaths.has(draft.target.path)) {
     throw new Error(
       `llmPolicyEditProposer: JSON path '${draft.target.path}' is outside allowedJsonPaths`
     );
   }
+  if (draft.change.mode !== "remove") {
+    const valueValidator = valueValidatorByJsonPath?.get(draft.target.path);
+    if (valueValidator) {
+      const parsedValue = valueValidator.safeParse(draft.change.value);
+      if (!parsedValue.success) {
+        throw new Error(
+          `llmPolicyEditProposer: change.value for JSON path '${draft.target.path}' does not match valueSchemaByJsonPath (${formatJsonValidationError(parsedValue.error)})`
+        );
+      }
+    }
+  }
   const cited = draft.source.findingKeys.map((findingKey) => {
     const finding = findingByKey.get(findingKey);
     if (!finding) {
@@ -7219,6 +7262,44 @@ function validateAllowedJsonPaths(paths) {
   }
   return [...paths];
 }
+function validatePathValueSchemas(input, allowedJsonPaths) {
+  if (input === void 0) return void 0;
+  if (!input || typeof input !== "object" || Array.isArray(input)) {
+    throw new Error("llmPolicyEditProposer: valueSchemaByJsonPath must be a JSON object");
+  }
+  const allowed = new Set(allowedJsonPaths);
+  const suppliedPaths = Object.keys(input);
+  const missing = allowedJsonPaths.filter((path) => !Object.hasOwn(input, path));
+  const unexpected = suppliedPaths.filter((path) => !allowed.has(path));
+  if (missing.length > 0 || unexpected.length > 0) {
+    throw new Error(
+      `llmPolicyEditProposer: valueSchemaByJsonPath keys must exactly match allowedJsonPaths (missing: ${missing.join(", ") || "none"}; unexpected: ${unexpected.join(", ") || "none"})`
+    );
+  }
+  const schemas = /* @__PURE__ */ new Map();
+  const validators = /* @__PURE__ */ new Map();
+  for (const path of allowedJsonPaths) {
+    const parsed = JsonValueSchema.safeParse(input[path]);
+    if (!parsed.success || !parsed.data || typeof parsed.data !== "object" || Array.isArray(parsed.data)) {
+      const detail = parsed.success ? "expected a JSON object" : formatJsonValidationError(parsed.error);
+      throw new Error(
+        `llmPolicyEditProposer: valueSchemaByJsonPath['${path}'] must be a valid JSON Schema object (${detail})`
+      );
+    }
+    const schema = parsed.data;
+    let validator;
+    try {
+      validator = z.fromJSONSchema(schema);
+    } catch (error) {
+      throw new Error(
+        `llmPolicyEditProposer: valueSchemaByJsonPath['${path}'] is unsupported (${error instanceof Error ? error.message : String(error)})`
+      );
+    }
+    schemas.set(path, schema);
+    validators.set(path, validator);
+  }
+  return { schemas, validators };
+}
 function requireNonEmpty(value, field) {
   if (!value || value.trim() !== value) {
     throw new Error(`llmPolicyEditProposer: ${field} must be a trimmed non-empty string`);
@@ -9185,4 +9266,4 @@ export {
   verifyCodeSurface,
   resolveWorktreePath
 };
-//# sourceMappingURL=chunk-ZU3QWGZE.js.map
+//# sourceMappingURL=chunk-A62YMFWA.js.map