@tangle-network/agent-eval 0.125.0 → 0.126.0

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.125.0",
+    "version": "0.126.0",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/rl.d.ts

@@ -2410,7 +2410,12 @@ declare function detectRewardHacking(input: DetectRewardHackingInput): RewardHac
 type CostChannel = 'agent' | 'judge' | 'verifier' | 'analyst' | 'driver' | (string & {});
 /** Per-million token rates for a model or endpoint not covered by package pricing. */
 interface CustomTokenPricing {
+    /** Non-cached input tokens. */
     inputUsdPerMillion: number;
+    /** Cache-read tokens. Falls back to the normal input rate when omitted. */
+    cachedInputUsdPerMillion?: number;
+    /** Cache-creation or cache-write tokens. Falls back to the normal input rate when omitted. */
+    cacheWriteUsdPerMillion?: number;
     outputUsdPerMillion: number;
 }
 interface ChannelRollup {
@@ -3501,104 +3506,6 @@ interface RLCampaignResult<V> {
 }
 declare function runRLCampaign<V>(opts: RunRLCampaignOptions<V>): Promise<RLCampaignResult<V>>;
 
-/**
- * Analyst contract — the missing orchestration layer over agent-eval's
- * existing analyzers (analyzeTraces, MultiLayerVerifier, RunCritic,
- * SemanticConceptJudge, JudgeFn, ...).
- *
- * Each existing primitive returns its own output shape. The Analyst
- * contract is the single envelope every primitive lifts into, so a
- * registry can run N analysts against a run and a single renderer can
- * compose findings without knowing which analyzer produced them.
- *
- * The contract is intentionally domain-agnostic: nothing here knows
- * about code, voice, RAG, or any particular agent stack. Analysts
- * declare what INPUT KIND they need (a trace store, an artifact dir,
- * a RunRecord, a JudgeInput, or `custom`), and the registry routes
- * the matching input from `AnalystRunInputs`.
- */
-
-interface EvidenceRef {
-    /**
-     * Where the evidence lives. `span` and `event` refer to OTLP trace
-     * elements; `artifact` to a file inside the run's artifact tree;
-     * `finding` to another AnalystFinding (cross-analyst chaining);
-     * `metric` to a named scalar reading the renderer knows how to read.
-     */
-    kind: 'span' | 'event' | 'artifact' | 'finding' | 'metric';
-    uri: string;
-    excerpt?: string;
-}
-
-type PolicyEditSchemaVersion = 'policy-edit/v1';
-declare const POLICY_EDIT_AXES: readonly ["carrier", "representation", "budget", "sampling", "output_contract", "tool_contract", "routing", "memory", "agent_profile", "deployment_target"];
-type PolicyEditAxis = (typeof POLICY_EDIT_AXES)[number];
-declare const POLICY_EDIT_TARGET_SURFACES: readonly ["prompt", "tool-contract", "runtime-config", "memory", "agent-profile", "code", "deployment"];
-type PolicyEditTargetSurface = (typeof POLICY_EDIT_TARGET_SURFACES)[number];
-type PolicyEditRisk = 'low' | 'medium' | 'high' | 'unknown';
-type PolicyEditGainDirection = 'increase' | 'decrease';
-type PolicyEditGainUnit = 'absolute' | 'relative' | 'percent' | 'score';
-interface PolicyEditTarget {
-    surface: PolicyEditTargetSurface;
-    /** Stable path inside the target surface, for example `system-prompt:tools`
-     * or `budget.maxTurns`. */
-    path?: string;
-    /** Optional canonical deployment identity. Store the existing cell, not a
-     * local profile shape. */
-    agentProfileCell?: AgentProfileCell;
-    /** Human label when the path is not enough for a readable audit trail. */
-    label?: string;
-}
-type PolicyEditChange = {
-    kind: 'text';
-    mode: 'append' | 'prepend' | 'replace';
-    value: string;
-    /** Required when `mode === 'replace'`; exact match only. */
-    find?: string;
-} | {
-    kind: 'json';
-    mode: 'set' | 'merge' | 'remove';
-    path: string;
-    value?: AgentProfileJson;
-};
-interface PolicyEditExpectedGain {
-    /** Metric this edit is expected to move, e.g. `holdout.composite`. */
-    metric: string;
-    direction: PolicyEditGainDirection;
-    /** Positive magnitude in the metric's native units. */
-    amount: number;
-    unit?: PolicyEditGainUnit;
-    rationale?: string;
-}
-interface PolicyEditSource {
-    findingIds: string[];
-    analystIds: string[];
-    evidenceRefs: EvidenceRef[];
-    /** Mirrors `AnalystFinding.derived_from_judge`; admission rejects it. */
-    derivedFromJudge?: boolean;
-}
-interface PolicyEdit {
-    schemaVersion: PolicyEditSchemaVersion;
-    editId: string;
-    axis: PolicyEditAxis;
-    target: PolicyEditTarget;
-    change: PolicyEditChange;
-    claim: string;
-    expectedGain: PolicyEditExpectedGain;
-    confidence: number;
-    risk: PolicyEditRisk;
-    source: PolicyEditSource;
-    rationale?: string;
-    validationPlan?: string;
-    metadata?: Record<string, unknown>;
-}
-declare const POLICY_EDIT_CANDIDATE_RECORD_SCHEMA: "tangle.policy-edit-candidate.v1";
-/** JSON-safe attribution carried with a measured candidate and its scores. */
-interface PolicyEditCandidateRecord {
-    schema: typeof POLICY_EDIT_CANDIDATE_RECORD_SCHEMA;
-    policyEdit: PolicyEdit;
-}
-
 /**
  * Pass A substrate types — `runCampaign` is the one primitive every
  * eval flow composes from. Three contracts in this file:
@@ -3776,8 +3683,6 @@ interface GenerationCandidate {
      *  "because rationale Z" the audit requires to survive to the result.
      *  Present when the proposer returned a `ProposedCandidate`. */
     rationale?: string;
-    /** Exact structured cause threaded from the proposer, when available. */
-    candidateRecord?: PolicyEditCandidateRecord;
 }
 interface CampaignAggregates {
     byJudge: Record<string, JudgeAggregate>;

package/dist/rl.js

@@ -10,14 +10,14 @@ import {
 } from "./chunk-3RF76KTD.js";
 import {
   runEvalCampaign
-} from "./chunk-GID26AN4.js";
+} from "./chunk-P22LJ3Y2.js";
 import {
   detectRewardHacking,
   extractVerifiableReward,
   extractVerifiableRewardsFromRecords,
   filterDeterministicallyRewarded
 } from "./chunk-ARU2PZFM.js";
-import "./chunk-EQUK3RFS.js";
+import "./chunk-J5SQWP6Y.js";
 import {
   rubricPredictiveValidity
 } from "./chunk-X4UCIOTZ.js";
@@ -34,14 +34,13 @@ import {
   thompsonCurriculum,
   varianceBasedCurriculum
 } from "./chunk-VZSRQ272.js";
-import "./chunk-GRCDRKII.js";
+import "./chunk-WS3NZZQQ.js";
 import "./chunk-VI2UW6B6.js";
 import "./chunk-TT4KNT67.js";
 import "./chunk-PC4UYEBM.js";
 import "./chunk-VQMK5FMP.js";
-import "./chunk-LOW3U7JZ.js";
+import "./chunk-IILEIWGW.js";
 import "./chunk-MA6HLL3S.js";
-import "./chunk-GC4ATIKK.js";
 import "./chunk-VSMTAMNK.js";
 import {
   ValidationError