@tangle-network/agent-eval 0.136.0 → 0.137.0

package/dist/{client-DcvgkaZi.d.ts → insight-report-B9ooYH_g.d.ts}

@@ -1,8 +1,7 @@
 import { o as FailureClass } from "./schema-BtVldJ3T.js";
-import { l as RunTerminalOutcome } from "./run-record-DcObtIGh.js";
-import { _ as GateDecision, j as MutableSurface } from "./types-DiWLru6Z.js";
-import { bt as ContinuousAgreement } from "./statistics-CKOqre5S.js";
-import { i as ParetoFigureSpec, t as GainDistributionBin } from "./summary-report-CPMINBqs.js";
+import { l as RunTerminalOutcome } from "./run-record-BPCa2rQ8.js";
+import { bt as ContinuousAgreement } from "./statistics-_7P642CN.js";
+import { i as ParetoFigureSpec, t as GainDistributionBin } from "./summary-report-DHipz9Kx.js";
 //#region src/contract/insight-report.d.ts
 interface InsightReport {
   /** Number of runs analyzed. */
@@ -403,202 +402,5 @@ interface Recommendation {
   evidencePath?: string;
 }
 //#endregion
-//#region src/hosted/types.d.ts
-declare const HOSTED_WIRE_VERSION: '2026-07-24.v1';
-type HostedWireVersion = typeof HOSTED_WIRE_VERSION;
-/** Every ingest request carries these. */
-interface HostedIngestHeaders {
-  /** Bearer token. The orchestrator validates against the tenant key. */
-  authorization: `Bearer ${string}`;
-  /** Stable tenant id (the orchestrator-side primary key for the tenant). */
-  'x-tangle-tenant-id': string;
-  /** Wire-version pin so the server can reject incompatible payloads. */
-  'x-tangle-wire-version': HostedWireVersion;
-  /** Stable request key generated once and reused across retries. */
-  'idempotency-key': string;
-}
-/** Lifecycle stages of an eval-run as the substrate reports them. */
-type EvalRunStatus = 'started' | 'baseline-complete' | 'generation-complete' | 'gate-decided' | 'finished' | 'errored';
-interface EvalRunCellScore {
-  /** Stable scenario id from the consumer's scenario set. */
-  scenarioId: string;
-  /** Repetition index when reps > 1; 0 for the default. */
-  rep: number;
-  /** Composite score across successful judges, or null when unscored. */
-  compositeMean: number | null;
-  /** Per-judge and per-dimension scores; failed or missing judges are absent. */
-  dimensions: Record<string, Record<string, number>>;
-  /** Root execution result, kept separate from task quality. */
-  terminalOutcome: RunTerminalOutcome;
-  /** Canonical execution-error count, or null when the producer did not measure it. */
-  executionErrorCount: number | null;
-  /** Per-cell dispatch or judge error. Missing on success. */
-  errorMessage?: string;
-}
-interface EvalRunGenerationSnapshot {
-  /** Generation index. 0 is baseline. */
-  index: number;
-  /** Candidate surface fingerprint (stable hash) — pivot key into the
-   *  trace stream to fetch the underlying execution. */
-  surfaceHash: string;
-  /** The candidate surface itself. May be omitted to avoid PII when the
-   *  consumer prefers not to ship verbatim prompts. */
-  surface?: MutableSurface;
-  /** Per-cell scores for this generation. */
-  cells: EvalRunCellScore[];
-  /** Mean across scored cells, or null when no cell has a task-quality label. */
-  compositeMean: number | null;
-  /** Total $ spent across this generation. */
-  costUsd: number;
-  /** Wall-clock duration of this generation. */
-  durationMs: number;
-}
-/**
- * The top-level eval-run event. One ingest call per logical eval-run;
- * generations stream in incrementally via repeated calls with the same
- * `runId`. The orchestrator deduplicates by `(runId, generation.index)`.
- */
-interface EvalRunEvent {
-  /** Stable run id (the substrate's `runId`). UUID or substrate-generated. */
-  runId: string;
-  /** Where this run was happening — derived from `RunCampaignOptions.runDir`. */
-  runDir: string;
-  /** ISO-8601 timestamp the substrate recorded the event. */
-  timestamp: string;
-  /** Lifecycle stage this event represents. */
-  status: EvalRunStatus;
-  /** Free-form consumer tags (env, branch, model id, etc.). Searchable. */
-  labels: Record<string, string>;
-  /** Baseline campaign snapshot. Present when status >= baseline-complete. */
-  baseline?: EvalRunGenerationSnapshot;
-  /** Per-generation snapshots. Streams in; orchestrator appends. */
-  generations: EvalRunGenerationSnapshot[];
-  /** Final gate decision. Present when status >= gate-decided. */
-  gateDecision?: GateDecision;
-  /** Held-out lift = winner-on-holdout - baseline-on-holdout. */
-  holdoutLift?: number;
-  /** Total $ spent across baseline + every generation. */
-  totalCostUsd: number;
-  /** Total wall-clock duration. */
-  totalDurationMs: number;
-  /** Error message if status === 'errored'. */
-  errorMessage?: string;
-  /** Rigor packet emitted alongside the run — distributional summary,
-   *  paired-bootstrap lift CI, judge stats, inter-rater agreement,
-   *  contamination check, failure clusters (when an analyst is wired),
-   *  outcome correlation (when downstream signal is supplied), and the
-   *  recommendations the dashboard surfaces verbatim. */
-  insightReport?: InsightReport;
-}
-/**
- * Canonical unsigned 64-bit integer encoded as a base-10 string.
- * JSON numbers cannot represent OTLP nanosecond timestamps exactly.
- */
-type UnixNanoTimestamp = string;
-/**
- * OTel-shape span with a few additional attributes for eval-run pivoting.
- * Compatible with any OTLP collector — `name`, `traceId`, `spanId`,
- * `startTimeUnixNano`, `endTimeUnixNano`, `attributes` are stock OTel.
- */
-interface TraceSpanEvent {
-  traceId: string;
-  spanId: string;
-  parentSpanId?: string;
-  name: string;
-  startTimeUnixNano: UnixNanoTimestamp;
-  endTimeUnixNano: UnixNanoTimestamp;
-  attributes: Record<string, string | number | boolean>;
-  events?: Array<{
-    timeUnixNano: UnixNanoTimestamp;
-    name: string;
-    attributes?: Record<string, string | number | boolean>;
-  }>;
-  status?: {
-    code: 'OK' | 'ERROR' | 'UNSET';
-    message?: string;
-  };
-  /** Pivot back into the eval-run stream. */
-  'tangle.runId'?: string;
-  /** Pivot to the specific generation. */
-  'tangle.generation'?: number;
-  /** Pivot to the specific cell. */
-  'tangle.cellId'?: string;
-  /** Pivot to the specific scenario. */
-  'tangle.scenarioId'?: string;
-}
-interface IngestEvalRunsRequest {
-  wireVersion: HostedWireVersion;
-  events: EvalRunEvent[];
-}
-interface IngestTracesRequest {
-  wireVersion: HostedWireVersion;
-  spans: TraceSpanEvent[];
-}
-interface IngestResponse {
-  /** Accepted events / spans count. */
-  accepted: number;
-  /** Rejected events with reasons (validation failures, dup idempotency key, etc.). */
-  rejected: Array<{
-    index: number;
-    reason: string;
-  }>;
-}
-//#endregion
-//#region src/hosted/client.d.ts
-interface HostedTenant {
-  /** Orchestrator endpoint base URL (no trailing slash). Required. */
-  endpoint: string;
-  /** Bearer token issued by the orchestrator. Required. */
-  apiKey: string;
-  /** Tenant id — the orchestrator's primary key for this consumer. Required. */
-  tenantId: string;
-  /** Optional `fetch` override (auth wrappers, custom agent, test mocks). */
-  fetchImpl?: typeof fetch;
-  /** Per-call timeout in ms. Default 30s. */
-  timeoutMs?: number;
-  /** Retries on 5xx / network errors. Default 2. */
-  retries?: number;
-}
-interface HostedClient {
-  ingestEvalRun(event: EvalRunEvent, idempotencyKey?: string): Promise<IngestResponse>;
-  ingestEvalRuns(events: EvalRunEvent[], idempotencyKey?: string): Promise<IngestResponse>;
-  ingestTraces(spans: TraceSpanEvent[], idempotencyKey?: string): Promise<IngestResponse>;
-  readonly tenant: HostedTenant;
-  readonly wireVersion: HostedWireVersion;
-}
-declare function createHostedClient(tenant: HostedTenant): HostedClient;
-/**
- * Build a `HostedClient` from environment, or `undefined` when ingest is not
- * configured — the canonical, fail-soft wiring every product uses so eval-run +
- * trace provenance lands in the Intelligence dashboard with ONE call:
- *
- *   const hosted = hostedClientFromEnv()
- *   // ...run the loop...
- *   await emitLoopProvenance({ ..., hostedClient: hosted })  // no-op if undefined
- *
- * Returns `undefined` (NOT an error) when any of endpoint / apiKey / tenantId is
- * missing — so a product wires the ship call unconditionally and it stays a
- * no-op until the env is set. Env precedence:
- *   - endpoint:  `TANGLE_INGEST_URL` → `TANGLE_ORCHESTRATOR_URL`
- *   - apiKey:    `TANGLE_INGEST_API_KEY` → `TANGLE_API_KEY`
- *   - tenantId:  `TANGLE_TENANT_ID`
- * A trailing slash on the endpoint is stripped. Pass `overrides` to supply any
- * field directly (e.g. a fixed `tenantId` per product) — overrides win over env.
- */
-/**
- * Build a {@link HostedTenant} config from env — the input `selfImprove`'s
- * `hostedTenant` and `emitLoopProvenance` take. Same env precedence + overrides
- * as {@link hostedClientFromEnv}; returns `undefined` (not an error) when any of
- * endpoint / apiKey / tenantId is missing, so a product wires
- * `hostedTenant: hostedTenantFromEnv({ tenantId: 'my-agent' })` unconditionally
- * and it stays off until the env is set.
- */
-declare function hostedTenantFromEnv(overrides?: Partial<HostedTenant> & {
-  env?: Record<string, string | undefined>;
-}): HostedTenant | undefined;
-declare function hostedClientFromEnv(overrides?: Partial<HostedTenant> & {
-  env?: Record<string, string | undefined>;
-}): HostedClient | undefined;
-//#endregion
-export { ScalarDistribution as A, InsightReport as C, OutcomeCorrelationInsight as D, LiftInsight as E, Recommendation as O, FailureClusterInsight as S, JudgeInsight as T, UnixNanoTimestamp as _, hostedTenantFromEnv as a, ExecutionInsight as b, EvalRunGenerationSnapshot as c, HostedIngestHeaders as d, HostedWireVersion as f, TraceSpanEvent as g, IngestTracesRequest as h, hostedClientFromEnv as i, TokenUsageInsight as j, ReleaseSummary as k, EvalRunStatus as l, IngestResponse as m, HostedTenant as n, EvalRunCellScore as o, IngestEvalRunsRequest as p, createHostedClient as r, EvalRunEvent as s, HostedClient as t, HOSTED_WIRE_VERSION as u, CostProvenanceSummary as v, InterRaterInsight as w, FailureClassTally as x, ExecutionErrorOutcomeCell as y };
-//# sourceMappingURL=client-DcvgkaZi.d.ts.map
+export { FailureClusterInsight as a, JudgeInsight as c, Recommendation as d, ReleaseSummary as f, FailureClassTally as i, LiftInsight as l, TokenUsageInsight as m, ExecutionErrorOutcomeCell as n, InsightReport as o, ScalarDistribution as p, ExecutionInsight as r, InterRaterInsight as s, CostProvenanceSummary as t, OutcomeCorrelationInsight as u };
+//# sourceMappingURL=insight-report-B9ooYH_g.d.ts.map

package/dist/insight-report-B9ooYH_g.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"insight-report-B9ooYH_g.d.ts","names":[],"sources":["../src/contract/insight-report.ts"],"mappings":";;;;;UAqCiB;;EAEf;;;;EAKA,WAAW;;EAGX,WAAW;;;EAIX,cAAc,eAAe;;EAG7B;IACE,MAAM;IACN,QAAQ;;;;IAIR,aAAa;;;;;IAKb;MAAa;MAAe;;;;;;EAM9B,QAAQ,eAAe;;;;EAKvB,aAAa;;;;EAKb,OAAO;;;EAIP,kBAAkB;;;EAIlB,gBAAgB;;;;;EAMhB,qBAAqB;;;;;EAMrB,SAAS;;;;;;EAOT,wBAAwB;;;;EAKxB,iBAAiB;;;EAIjB,iBAAiB;;UAGF;EACf;IAAY;IAAW;;EACvB;IAAa;IAAW;;EACxB;IAAc;;EACd;;UAGe;;EAEf,YAAY;;EAEZ,SAAS;;;EAGT,YAAY;;;EAGZ;IACE;IACA,YAAY;IACZ,SAAS;IACT;;;EAGF,QAAQ;IAAQ;IAAe;;;;;EAI/B;IACE;IACA;IACA;;;;;EAKF;IACE;;;IAGA;;IAEA;;IAEA;;IAEA;;IAEA;;;;;IAKA,mBAAmB,OAAO,oBAAoB;;;;EAIhD;IACE;IACA;IACA;IACA;IACA;;;UAIa;;EAEf;;EAEA;;EAEA;;UAGe;EACf,OAAO;EACP,QAAQ;EACR,WAAW;EACX,QAAQ;EACR,YAAY;EACZ;IACE;IACA;IACA;IACA;IACA;;;;UAOa;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,WAAW;;;;;EAKX,WAAW;IAAQ;IAAe;;;UAGnB;;EAEf;;EAEA;;;EAGA,cAAc;;;EAGd;;;EAGA;;;EAGA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,SAAS;;EAET,mBAAmB;IACjB;IACA,SAAS;MAAQ;MAAe;;IAChC;;;UAIa;EACf;EACA;;EAEA;;EAEA;;;EAGA;;EAEA;;EAEA;;;;;;;;;;;EAWA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf,UAAU;IACR;IACA;;IAEA;;IAEA;;IAEA;;EAEF;;;;;UAMe;;EAEf,cAAc;;EAEd;;EAEA;;UAGe;;EAEf;;;EAGA;EACA,UAAU;IAAQ;IAAe;IAAgB;;;UAGlC;;;EAGf;;EAEA;;EAEA;;EAEA;;EAEA;IACE;IACA;IACA;;;UAIa;;;EAGf;EACA,MAAM;IACJ;IACA;IACA;;;;EAIF;;UAGQ;;EAER;;EAEA;;;;EAIA;;EAEA;EACA;;KAGU,eACP;EACC;;EAEA;;EAEA;;EAEA;;EAEA;MAED;EACC;;EAEA;EACA;EACA;EACA;;UAGW;;EAEf;EACA;;EAEA;;;;EAIA,SAAS,eAAe;;;EAGxB;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;;EAEA"}