@tangle-network/agent-eval 0.117.1 → 0.118.1

package/dist/run-record-BDH49H2E.d.ts

@@ -1,360 +0,0 @@
-import { AgentProfile } from '@tangle-network/agent-interface';
-import { V as ValidationError } from './errors-oeQrLqXC.js';
-import { F as FailureClass } from './schema-B3Q3l9Z_.js';
-
-type AgentProfileCellSchemaVersion = 'agent-profile-cell/v1';
-type AgentProfileJson = string | number | boolean | null | AgentProfileJson[] | {
-    [key: string]: AgentProfileJson;
-};
-type AgentProfileDimensionValue = string | number | boolean | null;
-interface AgentProfileSource {
-    /** Runtime/profile contract being fingerprinted, e.g. `agent-interface-profile`. */
-    kind: string;
-    /** sha256 over the canonical source profile object. */
-    hash: string;
-}
-interface AgentProfileSourceInput {
-    kind: string;
-    /** Precomputed sha256 for callers that already sign their profile artifact. */
-    hash?: string;
-    /** Full canonical runtime profile; hashed and then discarded from the cell. */
-    profile?: AgentProfileJson;
-}
-interface AgentProfileHarness {
-    id: string;
-    version?: string;
-    hash?: string;
-}
-interface AgentProfileCellInput {
-    profileId: string;
-    sourceProfile: AgentProfileSourceInput;
-    harness?: AgentProfileHarness;
-    model?: string;
-    promptHash?: string;
-    dimensions?: Record<string, AgentProfileDimensionValue>;
-}
-interface AgentProfileCell {
-    schemaVersion: AgentProfileCellSchemaVersion;
-    cellId: string;
-    profileId: string;
-    sourceProfile: AgentProfileSource;
-    harness?: AgentProfileHarness;
-    model?: string;
-    promptHash?: string;
-    dimensions?: Record<string, AgentProfileDimensionValue>;
-}
-declare class AgentProfileCellValidationError extends ValidationError {
-    readonly path: string;
-    constructor(message: string, path?: string);
-}
-declare function buildAgentProfileCell(input: AgentProfileCellInput): Promise<AgentProfileCell>;
-declare function agentProfileCellHashMaterial(cell: AgentProfileCell): Omit<AgentProfileCell, 'cellId'>;
-/**
- * Verify an `AgentProfileCell`'s `cellId` matches the sha256 of its hash-material fields, confirming the record has not been tampered with.
- */
-declare function verifyAgentProfileCell(cell: AgentProfileCell): Promise<boolean>;
-declare function validateAgentProfileCell(input: unknown): AgentProfileCell;
-declare function requireAgentProfileCell(record: {
-    runId: string;
-    agentProfile?: AgentProfileCell;
-}): AgentProfileCell;
-declare function agentProfileCellKey(record: {
-    runId: string;
-    agentProfile?: AgentProfileCell;
-}): string;
-declare function assertRunAgentProfileCell(record: {
-    runId: string;
-    model: string;
-    promptHash: string;
-    agentProfile?: AgentProfileCell;
-}): Promise<AgentProfileCell>;
-declare function groupRunsByAgentProfileCell<T extends {
-    runId: string;
-    agentProfile?: AgentProfileCell;
-}>(records: readonly T[]): Map<string, T[]>;
-/** Canonical `sourceProfile.kind` values. Two products fingerprinting the
- *  same canonical profile MUST use the same kind for their cells to share
- *  `sourceProfile.hash`. Extend rather than create new strings — adding a
- *  new kind is a deliberate cross-product schema change. */
-declare const AGENT_PROFILE_KINDS: {
-    /** A profile declared via `defineAgentProfile(...)` from
-     *  `@tangle-network/agent-interface`. The default kind for router-backed
-     *  and sandbox-backed products. */
-    readonly AGENT_INTERFACE_PROFILE: "agent-interface-profile";
-};
-type AgentProfileKind = (typeof AGENT_PROFILE_KINDS)[keyof typeof AGENT_PROFILE_KINDS];
-/** Canonicalize an arbitrary value into `AgentProfileJson` by JSON
- *  round-trip. Throws when the value contains anything not representable
- *  as JSON (functions, BigInt, cycles) — non-portable profiles fail loud
- *  rather than silently dropping fields. */
-declare function toAgentProfileJson(value: unknown): AgentProfileJson;
-/** Canonical AgentProfile shape required when deriving a stable cell id. */
-type AgentInterfaceProfileLike = AgentProfile & {
-    name: string;
-    version: string;
-};
-/** Higher-level helper that hard-codes the canonical
- *  `agent-interface-profile` kind plus the JSON canonicalization. Equivalent
- *  to calling `buildAgentProfileCell` with `profileId = \`${name}@${version}\``
- *  and `sourceProfile = { kind: AGENT_INTERFACE_PROFILE, profile: <round-tripped> }`.
- *
- *  Use this from any product consuming an agent-interface `AgentProfile`; the
- *  manual `buildAgentProfileCell` call is reserved for advanced cases
- *  (custom kinds, pre-computed source hashes, alternate profileId
- *  conventions). */
-declare function buildAgentInterfaceProfileCell(profile: AgentInterfaceProfileLike, input: Omit<AgentProfileCellInput, 'profileId' | 'sourceProfile'>): Promise<AgentProfileCell>;
-
-/**
- * Paper-grade RunRecord schema + runtime validator.
- *
- * Every run that participates in a promotion gate, paper table, or
- * researcher loop SHOULD be recorded as a `RunRecord`. The mandatory
- * fields are exactly those the paper "Two Loops, Three Roles" requires
- * for reproducibility: who/what/when/cost/seed/hash, plus the search vs
- * holdout split tag and either a `searchScore` or a `holdoutScore`.
- *
- * This is intentionally NOT a replacement for the rich `Run` /
- * `ProposeReviewReport` / `ScenarioResult` types already in the
- * package. Those are runtime structures with full provenance. A
- * `RunRecord` is the analysis-time projection — the JSON-friendly
- * row you'd put in a parquet file or paste into a notebook.
- *
- * Validate at the boundary:
- *
- *   const rec = validateRunRecord(rawJson)         // throws on missing
- *   const ok  = isRunRecord(rawJson)               // boolean check
- *   const rec = parseRunRecordSafe(rawJson)        // { ok, value | error }
- *
- * The validator runs in pure TS — zod is intentionally NOT a
- * dependency. Round-trip tested in `tests/run-record.test.ts`.
- */
-
-/** Search/dev/holdout split tag. 'search' is the paper-grade alias for the
- *  combined train+test pool that the optimizer is allowed to read. */
-type RunSplitTag = 'search' | 'dev' | 'holdout';
-interface RunTokenUsage {
-    input: number;
-    output: number;
-    cached?: number;
-}
-/**
- * How a run's USD amount was obtained.
- *
- * `costUsd` remains mandatory for wire compatibility. New producers should
- * always populate this discriminated union so a missing bill is never
- * mistaken for an observed zero-dollar run. For `uncaptured`, `costUsd` uses
- * the legacy `0` sentinel while this field carries the truthful null.
- */
-type RunCostProvenance = {
-    kind: 'observed';
-    usd: number;
-} | {
-    kind: 'estimated';
-    usd: number;
-} | {
-    kind: 'uncaptured';
-    usd: null;
-};
-interface RunJudgeMetadata {
-    model: string;
-    promptVersion: string;
-    /** [0,1] confidence the judge declared. Constant judge confidence
-     *  across many runs is a fallback signal (see `canary.ts`). */
-    confidence: number;
-    /** True if the judge degraded to a fallback path (rules-only,
-     *  prior-call cache, etc.). The canary uses this to alert. */
-    fallback: boolean;
-}
-/**
- * Per-judge / per-dimension breakdown for runs scored by an ensemble of
- * judges over a multi-dimensional rubric.
- *
- * The collapsed `outcome.searchScore` / `holdoutScore` carries the
- * composite the gate uses. The full breakdown belongs here so consumers
- * can answer "which judge disagreed?", "which dimension dragged the
- * composite down?", and "did half the panel fail?" without re-running.
- *
- * `perJudge[judgeId][dim]` is the canonical source; `perDimMean` and
- * `composite` are convenience projections — derivable but precomputed so
- * downstream IRR primitives (`interRaterReliability`,
- * `corpusInterRaterAgreement`) and reporters don't pay the same
- * aggregation twice.
- *
- * Fail-loud discipline: judges that errored out land in `failedJudges`
- * by id. A missing key in `perJudge` is ambiguous (silent zero vs not
- * run); the explicit list makes a partial-failure recorded as such.
- */
-interface JudgeScoresRecord {
-    /** Per-judge per-dimension scores. `{ "kimi-k2.6": { helpfulness: 0.8, clarity: 0.7 }, ... }`. */
-    perJudge: Record<string, Record<string, number>>;
-    /** Per-dim mean across judges. Convenience — derivable from `perJudge`. */
-    perDimMean: Record<string, number>;
-    /** Composite mean across all dims and judges. Mirrors the score
-     *  the gate sees on `outcome.searchScore` / `holdoutScore`. */
-    composite: number;
-    /** Judges that errored or returned an unparseable verdict. Recorded
-     *  by id (e.g. `['glm-5.1']`) so a partial-failure case is explicit,
-     *  not inferred from missing keys in `perJudge`. */
-    failedJudges?: string[];
-    /** Free-form notes the judges emitted (joined across judges or
-     *  first-judge only — consumer's choice). */
-    notes?: string;
-}
-interface RunOutcome {
-    /** Score on the search/optimization split. Optional because a
-     *  holdout-only evaluation only fills `holdoutScore`. */
-    searchScore?: number;
-    /** Score on the held-out split. Optional because a search-only run
-     *  only fills `searchScore`. At least one must be present. */
-    holdoutScore?: number;
-    /** Bag of any other metric the run produced — judge dimensions,
-     *  pass/fail counters, latency stats, etc. Numeric only — keeps
-     *  reporters honest. */
-    raw: Record<string, number>;
-    /** Per-judge / per-dim breakdown. Consumers writing ensemble
-     *  judgements populate this; substrate primitives like
-     *  `interRaterReliability` and `corpusInterRaterAgreement` accept
-     *  these records as input. Optional — single-judge or scalar-only
-     *  runs leave it unset. */
-    judgeScores?: JudgeScoresRecord;
-    /** Authenticity / realness verdict — did the run build the REAL thing on the
-     *  intended infra, or fake it (see `./authenticity`)? Optional: only domains
-     *  with an authenticity config populate it. Carried in the corpus so the
-     *  flywheel / off-policy learning can optimize for real completion, not gamed
-     *  pass-rate. `score` is 0-1; `gated` is the anti-Goodhart flag — a gated run
-     *  must not count as a real success regardless of `score`. */
-    realness?: {
-        score: number;
-        gated: boolean;
-        reason?: string;
-    };
-}
-/**
- * Mandatory paper-grade fields for a single evaluation run. Optional
- * fields are extension points; mandatory fields throw if missing.
- *
- * Hash discipline:
- *   - `promptHash` is the sha256 of the EFFECTIVE prompt sent to the
- *     model (after any steering bundle merge).
- *   - `configHash` is the sha256 of the effective run config (model,
- *     temperature, tools, judges, splits). The pair (promptHash,
- *     configHash) uniquely identifies an experiment cell.
- *
- * Model snapshot discipline:
- *   - `model` MUST encode a snapshot version. Bare aliases like
- *     `claude-sonnet-4` or `gpt-4o` are banned — they remap silently.
- *     Use `claude-sonnet-4-6@2025-04-15` or `gpt-4o-2024-11-20`.
- */
-interface RunRecord {
-    /** UUID for the run. */
-    runId: string;
-    /** Logical experiment grouping (a treatment vs a baseline within
-     *  the same sweep should share `experimentId`). */
-    experimentId: string;
-    /** Stable identifier for the candidate (variant) being run. The
-     *  promotion gate compares two `candidateId`s on matched items. */
-    candidateId: string;
-    /** RNG seed for the run. Always recorded — silent re-seeding is
-     *  the most common cause of non-reproducible numbers. */
-    seed: number;
-    /** Model identifier WITH snapshot version. */
-    model: string;
-    /** sha256 of the effective prompt (post-steering). */
-    promptHash: string;
-    /** sha256 of the effective config. */
-    configHash: string;
-    /** Git SHA the harness was run from. */
-    commitSha: string;
-    /** End-to-end wall-clock duration in milliseconds. */
-    wallMs: number;
-    /** Time spent queued before execution started, if known. */
-    queueMs?: number;
-    /** Total USD cost. Mandatory — runs without a cost number are
-     *  unbounded by definition and must not be admitted into the gate.
-     *  `0` is retained as the compatibility sentinel for an uncaptured amount;
-     *  inspect `costProvenance` before treating it as observed. */
-    costUsd: number;
-    /** Observed, model-priced estimate, or genuinely uncaptured USD amount.
-     *  Optional only so existing serialized RunRecords remain valid. */
-    costProvenance?: RunCostProvenance;
-    /** Token usage breakdown. */
-    tokenUsage: RunTokenUsage;
-    /** Judge-side metadata, if a judge was used. */
-    judgeMetadata?: RunJudgeMetadata;
-    /** Per-split scores + raw bag. */
-    outcome: RunOutcome;
-    /** Canonical, cross-agent failure class drawn from the shared
-     *  `FAILURE_CLASSES` taxonomy. This is the aggregation key that makes
-     *  "which failure dominates across the whole fleet" answerable in ONE
-     *  vocabulary — every agent classifies against the same enum. Producers
-     *  set it via the substrate classifier; leave unset only when the failure
-     *  genuinely can't be classified. */
-    failureClass?: FailureClass;
-    /** Free-form domain-specific failure detail, scoped UNDER `failureClass`
-     *  (e.g. failureClass='tool_recovery_failure', failureMode='forge_build_unsatisfied').
-     *  The within-agent drill-down; `failureClass` is the cross-agent key. */
-    failureMode?: string;
-    /** Which split this run was drawn from. */
-    splitTag: RunSplitTag;
-    /**
-     * Stable scenario identifier the run was scored against. Optional for
-     * backwards compatibility, but **strongly recommended**: every primitive
-     * that pairs runs by scenario (preferences, paired stats, BT tournament)
-     * keys on this. The campaign artifact populates it canonically; legacy
-     * runs without it fall back to inference from `outcome.raw.scenario_id`
-     * or `experimentId`.
-     */
-    scenarioId?: string;
-    /**
-     * Canonical identity for the agent profile cell that produced this row:
-     * profile artifact hash plus optional harness/model/prompt/reporting
-     * dimensions. Use `agentProfile.cellId` to group persona sweeps and
-     * longitudinal reports by the complete source profile, not by a loose
-     * candidate label or opaque config hash.
-     */
-    agentProfile?: AgentProfileCell;
-}
-declare class RunRecordValidationError extends ValidationError {
-    readonly path: string;
-    constructor(message: string, path?: string);
-}
-/**
- * Strict validator. Throws `RunRecordValidationError` on the first
- * missing or wrongly-typed field. Returns the input cast to
- * `RunRecord` on success — the validator does not coerce.
- */
-declare function validateRunRecord(input: unknown): RunRecord;
-/**
- * Resolve provenance for both new and legacy records.
- *
- * Legacy producers sometimes set `outcome.raw.cost_estimated = 1`. A positive
- * unlabeled amount is treated as observed, matching the historical contract.
- * Zero without an explicit label is conservatively uncaptured: claiming an
- * observed $0 would be stronger than the serialized evidence supports.
- */
-declare function resolveRunCostProvenance(run: Pick<RunRecord, 'costUsd' | 'costProvenance' | 'outcome'>): RunCostProvenance;
-/** Boolean validator — convenience for filtering arrays. */
-declare function isRunRecord(input: unknown): input is RunRecord;
-/** Non-throwing validator — returns a discriminated union. */
-declare function parseRunRecordSafe(input: unknown): {
-    ok: true;
-    value: RunRecord;
-} | {
-    ok: false;
-    error: RunRecordValidationError;
-};
-/** Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. */
-declare function roundTripRunRecord(record: RunRecord): RunRecord;
-/**
- * Heuristic snapshot check. Accepts:
- *   - `name@YYYY-MM-DD` (Anthropic style: `claude-sonnet-4-6@2025-04-15`)
- *   - `name-YYYYMMDD`   (OpenAI style: `gpt-4o-2024-11-20`)
- *   - `name@<arbitrary-token>` (allow opaque snapshots like `@v3`)
- *   - explicit `:date-...` Vertex-style tags
- *
- * Rejects bare aliases like `claude-sonnet-4` or `gpt-4o` that remap
- * silently as providers ship new snapshots.
- */
-declare function modelHasSnapshot(model: string): boolean;
-
-export { type AgentProfileCell as A, requireAgentProfileCell as B, resolveRunCostProvenance as C, roundTripRunRecord as D, toAgentProfileJson as E, validateAgentProfileCell as F, validateRunRecord as G, verifyAgentProfileCell as H, type JudgeScoresRecord as J, type RunRecord as R, type RunSplitTag as a, type RunCostProvenance as b, type RunTokenUsage as c, type RunJudgeMetadata as d, type AgentProfileCellInput as e, type AgentProfileJson as f, AGENT_PROFILE_KINDS as g, type AgentInterfaceProfileLike as h, type AgentProfileCellSchemaVersion as i, AgentProfileCellValidationError as j, type AgentProfileDimensionValue as k, type AgentProfileHarness as l, type AgentProfileKind as m, type AgentProfileSource as n, type AgentProfileSourceInput as o, type RunOutcome as p, RunRecordValidationError as q, agentProfileCellHashMaterial as r, agentProfileCellKey as s, assertRunAgentProfileCell as t, buildAgentInterfaceProfileCell as u, buildAgentProfileCell as v, groupRunsByAgentProfileCell as w, isRunRecord as x, modelHasSnapshot as y, parseRunRecordSafe as z };

package/dist/runtime-trajectory-DGBIUt4B.d.ts

@@ -1,49 +0,0 @@
-import { a as RunSplitTag } from './run-record-BDH49H2E.js';
-
-interface RuntimeTrajectoryHookEvent {
-    id: string;
-    runId: string;
-    scenarioId?: string;
-    target: string;
-    phase: string;
-    timestamp: number;
-    stepIndex?: number;
-    parentId?: string;
-    payload?: unknown;
-    metadata?: Record<string, unknown>;
-}
-interface RuntimeTrajectoryRecord {
-    id?: string;
-    scenarioId?: string;
-    splitTag?: RunSplitTag;
-    runtimeEvents?: unknown;
-    [key: string]: unknown;
-}
-interface RuntimeTrajectoryRunRecord {
-    runId: string;
-    scenarioId?: string;
-    splitTag: RunSplitTag;
-}
-interface RuntimeTrajectoryEvidenceSummary {
-    recordCount: number;
-    recordWithRuntimeEventsCount: number;
-    runtimeRunCount: number;
-    lifecycleEventCount: number;
-    defaultedSplitCount: number;
-}
-interface RuntimeTrajectoryEvidenceProjection {
-    runs: RuntimeTrajectoryRunRecord[];
-    events: RuntimeTrajectoryHookEvent[];
-    summary: RuntimeTrajectoryEvidenceSummary;
-    diagnostics: string[];
-}
-interface ProjectRuntimeTrajectoryEvidenceOptions<TRecord extends RuntimeTrajectoryRecord = RuntimeTrajectoryRecord> {
-    records: TRecord[];
-    defaultSplitTag?: RunSplitTag;
-    recordIdOf?: (record: TRecord, index: number) => string | undefined;
-    scenarioIdOf?: (record: TRecord, index: number) => string | undefined;
-}
-declare function projectRuntimeTrajectoryEvidence<TRecord extends RuntimeTrajectoryRecord>(options: ProjectRuntimeTrajectoryEvidenceOptions<TRecord>): RuntimeTrajectoryEvidenceProjection;
-declare function parseRuntimeTrajectoryHookEvent(input: unknown): RuntimeTrajectoryHookEvent | null;
-
-export { type ProjectRuntimeTrajectoryEvidenceOptions as P, type RuntimeTrajectoryRecord as R, type RuntimeTrajectoryEvidenceProjection as a, type RuntimeTrajectoryEvidenceSummary as b, type RuntimeTrajectoryHookEvent as c, type RuntimeTrajectoryRunRecord as d, projectRuntimeTrajectoryEvidence as e, parseRuntimeTrajectoryHookEvent as p };

package/dist/schema-B3Q3l9Z_.d.ts

@@ -1,201 +0,0 @@
-/**
- * TraceSchema v1 — the canonical data model for agent-eval.
- *
- * Every score, every failure class, every pipeline in the framework is
- * a view over this data. Shape it once, live with it.
- *
- * Wire-compatible with OpenTelemetry span semantics (see trace/otel.ts)
- * but extended with agent-specific span kinds (llm, tool, retrieval,
- * judge, sandbox) and first-class BudgetLedger / Artifact / JudgeVerdict
- * entities that OTEL leaves as free-form attributes.
- */
-declare const TRACE_SCHEMA_VERSION = "1.0.0";
-type RunStatus = 'running' | 'completed' | 'failed' | 'aborted';
-interface BudgetSpec {
-    tokens?: number;
-    wallMs?: number;
-    calls?: number;
-    usd?: number;
-}
-interface RunOutcome {
-    score?: number;
-    pass?: boolean;
-    failureClass?: FailureClass;
-    notes?: string;
-}
-/**
- * Layer — optional classification in a nested build workflow.
- * `builder`: the meta-agent editing a project (e.g. agent-builder Forge chat).
- * `app-build`: sandbox harness that compiled + tested the generated scaffold.
- * `app-runtime`: a run of the generated agent against a domain scenario.
- * `meta`: any meta-eval (judge replay, correlation analysis).
- */
-type RunLayer = 'builder' | 'app-build' | 'app-runtime' | 'meta' | 'custom';
-interface Run {
-    runId: string;
-    /**
-     * Stable identifier of the scenario being executed.
-     *
-     * Always populated on the persisted Run — but `TraceEmitter.startRun` accepts
-     * input WITHOUT this field, substituting a sensible default
-     * (`run.layer ?? run.tags?.['kind'] ?? 'runtime'`) when the caller has no
-     * curated scenario to anchor to (runtime / operator / meta-eval runs). This
-     * keeps the persisted shape unambiguous for downstream filters + aggregations
-     * while removing the boilerplate of inventing placeholder ids at the call site.
-     */
-    scenarioId: string;
-    variantId?: string;
-    datasetVersion?: string;
-    /** Git SHA of agent code at run time. */
-    codeSha?: string;
-    /** Hash of the prompt template + any system prompt. */
-    promptSha?: string;
-    /** Model id + date + system-prompt hash, concatenated. */
-    modelFingerprint?: string;
-    seed?: number;
-    /** Arbitrary environment markers (shell, docker version, tz). */
-    envFingerprint?: Record<string, string>;
-    /** Version of the redaction rules applied to this run. */
-    redactionVersion?: string;
-    /** Parent run in a nested build workflow. A builder run's children are
-     *  app-build runs; those children are app-runtime runs. */
-    parentRunId?: string;
-    /** Stable project identifier — groups runs across chats + sessions. */
-    projectId?: string;
-    /** Chat/conversation identifier within a project. */
-    chatId?: string;
-    /** Layer classification — hint for aggregation; not enforced. */
-    layer?: RunLayer;
-    startedAt: number;
-    endedAt?: number;
-    status: RunStatus;
-    outcome?: RunOutcome;
-    budget?: BudgetSpec;
-    /** Free-form labels for downstream grouping. */
-    tags?: Record<string, string>;
-}
-type SpanKind = 'agent' | 'llm' | 'tool' | 'retrieval' | 'judge' | 'sandbox' | 'custom';
-type SpanStatus = 'ok' | 'error';
-interface SpanBase {
-    spanId: string;
-    parentSpanId?: string;
-    runId: string;
-    kind: SpanKind;
-    name: string;
-    startedAt: number;
-    endedAt?: number;
-    status?: SpanStatus;
-    error?: string;
-    /** Anything not covered by typed fields. Kept deliberately free-form. */
-    attributes?: Record<string, unknown>;
-}
-interface Message {
-    role: 'system' | 'user' | 'assistant' | 'tool';
-    content: string;
-    tokens?: number;
-    /** Multi-modal content descriptors; blobs themselves live in Artifacts. */
-    images?: Array<{
-        artifactId?: string;
-        url?: string;
-        mime?: string;
-    }>;
-}
-interface LlmSpan extends SpanBase {
-    kind: 'llm';
-    model: string;
-    messages: Message[];
-    output?: string;
-    inputTokens?: number;
-    outputTokens?: number;
-    cachedTokens?: number;
-    reasoningTokens?: number;
-    costUsd?: number;
-    finishReason?: string;
-}
-interface ToolSpan extends SpanBase {
-    kind: 'tool';
-    toolName: string;
-    args: unknown;
-    /** False when the source observed the call but did not capture its arguments. */
-    argsCaptured?: boolean;
-    result?: unknown;
-    latencyMs?: number;
-}
-interface RetrievalSpan extends SpanBase {
-    kind: 'retrieval';
-    query: string;
-    hits: Array<{
-        docId: string;
-        score: number;
-        content?: string;
-    }>;
-}
-interface JudgeSpan extends SpanBase {
-    kind: 'judge';
-    judgeId: string;
-    /** Span this judgment applies to. */
-    targetSpanId: string;
-    dimension: string;
-    /** Numeric score (free-range; interpretation up to the judge). */
-    score: number;
-    rationale?: string;
-    evidence?: string;
-}
-interface SandboxSpan extends SpanBase {
-    kind: 'sandbox';
-    image?: string;
-    command?: string;
-    exitCode?: number;
-    testsTotal?: number;
-    testsPassed?: number;
-    stdoutHash?: string;
-    stderrHash?: string;
-    /** Duration in ms; the harness fills this explicitly (endedAt - startedAt may miss setup). */
-    wallMs?: number;
-}
-interface GenericSpan extends SpanBase {
-    kind: 'agent' | 'custom';
-}
-type Span = LlmSpan | ToolSpan | RetrievalSpan | JudgeSpan | SandboxSpan | GenericSpan;
-type EventKind = 'log' | 'error' | 'budget_decrement' | 'budget_breach' | 'state_mutation' | 'policy_violation' | 'redaction_applied' | 'custom';
-interface TraceEvent {
-    eventId: string;
-    runId: string;
-    spanId?: string;
-    kind: EventKind;
-    timestamp: number;
-    payload: Record<string, unknown>;
-}
-interface BudgetLedgerEntry {
-    runId: string;
-    dimension: keyof BudgetSpec;
-    limit: number;
-    consumed: number;
-    remaining: number;
-    timestamp: number;
-    breached: boolean;
-    /** Span that triggered this entry, if any. */
-    spanId?: string;
-}
-interface Artifact {
-    artifactId: string;
-    runId: string;
-    spanId?: string;
-    contentType: string;
-    sizeBytes: number;
-    /** sha256 in hex. */
-    hash: string;
-    /** External storage URL (R2, S3, filesystem path). */
-    storageUrl?: string;
-    /** Inline content for small blobs — keep under ~64KB. */
-    inlineContent?: string;
-}
-type FailureClass = 'success' | 'reasoning_error' | 'tool_selection_error' | 'tool_argument_error' | 'tool_recovery_failure' | 'hallucination' | 'instruction_following' | 'safety_refusal_miss' | 'policy_violation' | 'budget_exceeded' | 'format_drift' | 'permission_escalation' | 'pii_leak' | 'cost_overrun' | 'timeout' | 'sandbox_failure' | 'missing_user_data' | 'missing_domain_data' | 'missing_codebase_context' | 'missing_runtime_context' | 'missing_credentials' | 'missing_integration_connection' | 'missing_integration_scope' | 'integration_approval_required' | 'integration_auth_expired' | 'integration_provider_failure' | 'bad_integration_manifest' | 'unsafe_integration_write_denied' | 'stale_external_data' | 'bad_retrieval' | 'insufficient_evidence' | 'contradictory_evidence' | 'ambiguous_user_intent' | 'knowledge_readiness_blocked' | 'unknown';
-declare const FAILURE_CLASSES: readonly FailureClass[];
-declare function isLlmSpan(s: Span): s is LlmSpan;
-declare function isToolSpan(s: Span): s is ToolSpan;
-declare function isRetrievalSpan(s: Span): s is RetrievalSpan;
-declare function isJudgeSpan(s: Span): s is JudgeSpan;
-declare function isSandboxSpan(s: Span): s is SandboxSpan;
-
-export { type Artifact as A, type BudgetLedgerEntry as B, type EventKind as E, type FailureClass as F, type GenericSpan as G, type JudgeSpan as J, type LlmSpan as L, type Message as M, type Run as R, type Span as S, type ToolSpan as T, type RunOutcome as a, type SpanKind as b, type RetrievalSpan as c, type SandboxSpan as d, type TraceEvent as e, type RunStatus as f, type RunLayer as g, type BudgetSpec as h, FAILURE_CLASSES as i, type SpanBase as j, type SpanStatus as k, TRACE_SCHEMA_VERSION as l, isJudgeSpan as m, isLlmSpan as n, isRetrievalSpan as o, isSandboxSpan as p, isToolSpan as q };