@tangle-network/agent-eval 0.117.0 → 0.118.0

package/dist/storage-DrX3v_5B.d.ts

@@ -1,50 +0,0 @@
-import { C as CostLedger } from './cost-ledger-DWy3XdJc.js';
-
-/**
- * `CampaignStorage` — the filesystem seam `runCampaign` writes through
- * (run/cell dirs, the resumability cache, per-cell artifacts, trace spans).
- *
- * The default (`fsCampaignStorage`) is the Node filesystem — identical
- * behavior to the inline `node:fs` calls it replaces, so existing CLI
- * consumers are unaffected. `inMemoryCampaignStorage` keeps everything in a
- * `Map`, so the substrate runs in environments WITHOUT a filesystem
- * (Cloudflare Workers, Deno Deploy, other edge runtimes) — the campaign
- * still produces its `CampaignResult` (cells + aggregates) in memory;
- * artifacts/traces simply aren't persisted to disk.
- *
- * Paths are opaque keys to the in-memory adapter — it does not parse them,
- * so the same `join(...)`-built paths work unchanged across both adapters.
- */
-interface CampaignStorage {
-    /** Ensure a directory exists (recursive). No-op for in-memory. */
-    ensureDir(dir: string): void;
-    /** Does this path exist (as a written file or an ensured dir)? */
-    exists(path: string): boolean;
-    /** Read a UTF-8 file; `undefined` when missing or unreadable. */
-    read(path: string): string | undefined;
-    /** Write a file (string or bytes). Parent dir is assumed ensured. */
-    write(path: string, content: string | Uint8Array): void;
-    /** Append only when the current UTF-8 byte length matches `expectedBytes`.
-     * Returns the new length, or undefined when another writer won. */
-    append?(path: string, content: string, expectedBytes: number): number | undefined;
-}
-/** Node-filesystem storage — the default. Lazily requires `node:fs` so the
- *  module imports cleanly in non-Node runtimes (where the caller passes
- *  `inMemoryCampaignStorage` instead and never constructs this).
- *
- *  `createRequire(import.meta.url)` is the ESM-native lazy require — a bare
- *  `require` is a ReferenceError under `"type": "module"`, which is exactly
- *  the shape this package publishes. */
-declare function fsCampaignStorage(): CampaignStorage;
-/** In-memory storage for filesystem-less runtimes. Artifacts + trace spans
- *  live in a `Map` for the duration of the run; the `CampaignResult` is
- *  fully populated, but nothing is persisted to disk. */
-declare function inMemoryCampaignStorage(): CampaignStorage;
-/** Open the durable spend account stored beside a logical run. */
-declare function createRunCostLedger(input: {
-    storage: CampaignStorage;
-    runDir: string;
-    costCeilingUsd?: number;
-}): CostLedger;
-
-export { type CampaignStorage as C, createRunCostLedger as c, fsCampaignStorage as f, inMemoryCampaignStorage as i };

package/dist/store-C1YxJDEK.d.ts

@@ -1,248 +0,0 @@
-/**
- * Shared types for the trace-analyst module.
- *
- * Wire format. The store interface speaks `OtlpSpanLike` rows — one JSONL
- * line per span, OTLP-shaped. We do NOT depend on a specific tracing
- * vendor at the type level. Adapter
- * layers map upstream shapes onto this interface.
- *
- * Design constraint. Every read operation that can return arbitrary
- * payload must carry a byte budget so the agent's tool result stays
- * bounded regardless of input trace size. Oversized responses
- * substitute a deterministic summary instead of bytes — see
- * `ViewTraceOversized`.
- */
-/** OTLP span kind (subset we actually use). */
-type TraceAnalystSpanKind = 'AGENT' | 'LLM' | 'TOOL' | 'CHAIN' | 'GUARDRAIL' | 'SPAN' | 'UNKNOWN';
-type TraceAnalystSpanStatus = 'OK' | 'ERROR' | 'UNSET';
-/** Subset of OTLP span fields the analyst exposes to the agent. The
- *  store's job is to project upstream's full span shape down to this
- *  view — the analyst never sees vendor extensions directly. */
-interface TraceAnalystSpan {
-    trace_id: string;
-    span_id: string;
-    parent_span_id: string | null;
-    name: string;
-    kind: TraceAnalystSpanKind;
-    start_time: string;
-    end_time: string;
-    duration_ms: number;
-    status: TraceAnalystSpanStatus;
-    status_message?: string;
-    service_name: string | null;
-    agent_name: string | null;
-    model_name: string | null;
-    tool_name: string | null;
-    /** Raw JSON-serialisable attribute map. May contain large strings;
-     *  callers must respect the per-attribute byte cap. */
-    attributes: Record<string, unknown>;
-}
-interface TraceAnalystTraceSummary {
-    trace_id: string;
-    service_name: string | null;
-    agent_name: string | null;
-    span_count: number;
-    has_errors: boolean;
-    start_time: string;
-    end_time: string;
-    duration_ms: number;
-    raw_jsonl_bytes: number;
-    models: string[];
-    tools: string[];
-}
-interface TraceAnalystFilters {
-    /** Restrict to traces that contain at least one error span. */
-    has_errors?: boolean;
-    /** Match if any span's `service.name` is in this list. */
-    service_names?: string[];
-    /** Match if any span's `agent.name` is in this list. */
-    agent_names?: string[];
-    /** Match if any LLM span's `llm.model_name` is in this list. */
-    model_names?: string[];
-    /** Match if any tool span's `tool.name` is in this list. */
-    tool_names?: string[];
-    /** ISO-8601 lower bound on the trace's earliest start time. */
-    start_time_after?: string;
-    /** ISO-8601 upper bound on the trace's earliest start time. */
-    start_time_before?: string;
-    /** Single regex applied to raw JSONL bytes for the trace. Opt-in;
-     *  expensive on large datasets. Use the indexed filters above first. */
-    regex_pattern?: string;
-}
-/** One distinct error signature across the dataset — the deterministic unit of
- *  failure coverage. Signatures normalize volatile tokens (digits, hex/uuids,
- *  paths, durations) out of the span `status_message` so semantically identical
- *  failures collapse into one cluster. An analyst that accounts for every
- *  cluster has, by construction, covered every distinct failure mode. */
-interface ErrorCluster {
-    /** Normalized status_message — the cluster key. */
-    signature: string;
-    /** A verbatim, un-normalized exemplar message (for exact-string citation). */
-    status_message_sample: string;
-    /** The span name that most often carries this signature, if any. */
-    span_name: string | null;
-    /** The tool that most often carries this signature, if any. */
-    tool_name: string | null;
-    trace_count: number;
-    span_count: number;
-    /** trace_count / total error traces in the matched set (0..1). */
-    prevalence: number;
-    /** Real trace ids carrying this signature (capped), passable to view/search. */
-    exemplar_trace_ids: string[];
-    /** Real span ids carrying this signature (capped). */
-    exemplar_span_ids: string[];
-}
-interface DatasetOverview {
-    total_traces: number;
-    raw_jsonl_bytes: number;
-    services: string[];
-    agents: string[];
-    models: string[];
-    tool_names: string[];
-    /** Up to 20 real trace ids the agent may pass to view/search tools. */
-    sample_trace_ids: string[];
-    errors: {
-        trace_count: number;
-        span_count: number;
-    };
-    /** The COMPLETE deterministic error-signature population, sorted by
-     *  trace_count desc. This is the failure-coverage checklist: an analysis is
-     *  complete only when every cluster here is accounted for. Empty when the
-     *  matched set has no error spans. */
-    error_clusters: ErrorCluster[];
-    time_range: {
-        earliest: string;
-        latest: string;
-    } | null;
-}
-interface QueryTracesPage {
-    traces: TraceAnalystTraceSummary[];
-    total: number;
-    has_more: boolean;
-}
-/** Full-trace view. When the response would exceed the per-call byte
- *  budget, `oversized` is populated INSTEAD of `spans` so the agent
- *  knows to switch to `searchTrace` / `viewSpans`. */
-interface ViewTraceResult {
-    trace_id: string;
-    spans?: TraceAnalystSpan[];
-    oversized?: ViewTraceOversized;
-}
-interface ViewTraceOversized {
-    span_count: number;
-    /** Names with their counts, sorted desc. Capped at 20 entries. */
-    top_span_names: Array<[string, number]>;
-    /** Largest single span body (bytes after attribute-cap projection). */
-    span_response_bytes_max: number;
-    error_span_count: number;
-}
-interface ViewSpansResult {
-    trace_id: string;
-    spans: TraceAnalystSpan[];
-    /** Number of requested span ids that were not found in the trace. */
-    missing_span_ids: string[];
-    /** Number of attribute fields truncated to fit the per-attribute cap. */
-    truncated_attribute_count: number;
-}
-interface SpanMatchRecord {
-    trace_id: string;
-    span_id: string;
-    span_name: string;
-    span_kind: TraceAnalystSpanKind;
-    /** JSON pointer-style path to the matched value, e.g.
-     *  `attributes."llm.input_messages"[2].content`. */
-    attribute_path: string;
-    matched_text: string;
-    context_before: string;
-    context_after: string;
-    match_offset: number;
-}
-interface SearchTraceResult {
-    trace_id: string;
-    hits: SpanMatchRecord[];
-    total_matches: number;
-    has_more: boolean;
-}
-interface SearchSpanResult {
-    trace_id: string;
-    span_id: string;
-    hits: SpanMatchRecord[];
-    total_matches: number;
-    has_more: boolean;
-}
-/** Tunable byte budgets for bounded RLM tool output. */
-interface TraceAnalystByteBudgets {
-    /** Max bytes any single tool response may emit. Hard ceiling enforced
-     *  by the store; oversized → summary. Default 150_000. */
-    perCallByteCeiling: number;
-    /** Per-attribute string truncation cap on `viewTrace` (discovery scan).
-     *  Default 4096. */
-    perAttributeViewBudget: number;
-    /** Per-attribute string truncation cap on `viewSpans` (surgical reads).
-     *  Default 16384. */
-    perAttributeSpanBudget: number;
-    /** Per-attribute cap on a single match record's `matched_text` and
-     *  context window. Default 1024. */
-    perMatchTextBudget: number;
-}
-declare const DEFAULT_TRACE_ANALYST_BUDGETS: TraceAnalystByteBudgets;
-/** Marker substituted in place of truncated string payloads. Callers
- *  parsing tool output can detect it deterministically. */
-declare const TRACE_ANALYST_TRUNCATION_MARKER_PREFIX = "[trace-analyst truncated:";
-
-/**
- * `TraceAnalysisStore` — read-side interface the trace-analyst calls
- * through. Six operations, all bounded:
- *
- *   - `getOverview(filters?)` — dataset rollup + sample trace ids.
- *   - `queryTraces(filters?, limit, offset)` — paginated summaries.
- *   - `countTraces(filters?)` — cheap count without materialisation.
- *   - `viewTrace(trace_id, perAttrCap)` — full span list, oversized → summary.
- *   - `viewSpans(trace_id, span_ids, perAttrCap)` — surgical span fetch.
- *   - `searchTrace(trace_id, regex, max_matches)` — bounded regex hits.
- *   - `searchSpan(trace_id, span_id, regex, max_matches)` — single-span search.
- *
- * Multiple implementations ship in the core (`OtlpFileTraceStore`).
- * Downstream callers can supply their own — e.g. a DuckDB-backed
- * adapter or an in-memory adapter for tests — by implementing this
- * interface.
- *
- * Filters compose with AND semantics. Empty/undefined fields impose
- * no constraint. `regex_pattern` is the only opt-in raw-bytes scan —
- * implementations may skip it via `count`/`overview` when not set.
- */
-
-interface TraceAnalysisStore {
-    getOverview(filters?: TraceAnalystFilters): Promise<DatasetOverview>;
-    queryTraces(opts: {
-        filters?: TraceAnalystFilters;
-        limit: number;
-        offset?: number;
-    }): Promise<QueryTracesPage>;
-    countTraces(filters?: TraceAnalystFilters): Promise<number>;
-    viewTrace(opts: {
-        trace_id: string;
-        /** Override per-attribute byte cap. Defaults to discovery budget. */
-        per_attribute_byte_cap?: number;
-    }): Promise<ViewTraceResult>;
-    viewSpans(opts: {
-        trace_id: string;
-        span_ids: readonly string[];
-        /** Override per-attribute byte cap. Defaults to surgical budget. */
-        per_attribute_byte_cap?: number;
-    }): Promise<ViewSpansResult>;
-    searchTrace(opts: {
-        trace_id: string;
-        regex_pattern: string;
-        /** Hard cap on matches returned. Default 50. */
-        max_matches?: number;
-    }): Promise<SearchTraceResult>;
-    searchSpan(opts: {
-        trace_id: string;
-        span_id: string;
-        regex_pattern: string;
-        max_matches?: number;
-    }): Promise<SearchSpanResult>;
-}
-
-export { DEFAULT_TRACE_ANALYST_BUDGETS as D, type ErrorCluster as E, type QueryTracesPage as Q, type SearchSpanResult as S, type TraceAnalysisStore as T, type ViewSpansResult as V, type TraceAnalystSpan as a, type DatasetOverview as b, type SearchTraceResult as c, type SpanMatchRecord as d, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX as e, type TraceAnalystByteBudgets as f, type TraceAnalystFilters as g, type TraceAnalystSpanKind as h, type TraceAnalystSpanStatus as i, type TraceAnalystTraceSummary as j, type ViewTraceOversized as k, type ViewTraceResult as l };

package/dist/store-DGqD0Pyo.d.ts

@@ -1,116 +0,0 @@
-import { R as Run, S as Span, e as TraceEvent, A as Artifact, B as BudgetLedgerEntry, f as RunStatus, g as RunLayer, b as SpanKind, E as EventKind } from './schema-B3Q3l9Z_.js';
-
-interface RunFilter {
-    scenarioId?: string;
-    variantId?: string;
-    status?: RunStatus;
-    since?: number;
-    until?: number;
-    tag?: {
-        key: string;
-        value: string;
-    };
-    parentRunId?: string;
-    projectId?: string;
-    chatId?: string;
-    layer?: RunLayer;
-}
-interface SpanFilter {
-    runId?: string;
-    parentSpanId?: string;
-    kind?: SpanKind;
-    name?: string;
-    toolName?: string;
-    judgeId?: string;
-    since?: number;
-    until?: number;
-}
-interface EventFilter {
-    runId?: string;
-    spanId?: string;
-    kind?: EventKind;
-    since?: number;
-    until?: number;
-}
-interface TraceStore {
-    appendRun(run: Run): Promise<void>;
-    updateRun(runId: string, patch: Partial<Run>): Promise<void>;
-    appendSpan(span: Span): Promise<void>;
-    updateSpan(spanId: string, patch: Partial<Span>): Promise<void>;
-    appendEvent(event: TraceEvent): Promise<void>;
-    appendArtifact(artifact: Artifact): Promise<void>;
-    appendBudgetEntry(entry: BudgetLedgerEntry): Promise<void>;
-    getRun(runId: string): Promise<Run | undefined>;
-    listRuns(filter?: RunFilter): Promise<Run[]>;
-    spans(filter?: SpanFilter): Promise<Span[]>;
-    events(filter?: EventFilter): Promise<TraceEvent[]>;
-    budget(runId: string): Promise<BudgetLedgerEntry[]>;
-    artifacts(runId: string): Promise<Artifact[]>;
-}
-declare class InMemoryTraceStore implements TraceStore {
-    private runs;
-    private allSpans;
-    private allEvents;
-    private allArtifacts;
-    private allBudget;
-    appendRun(run: Run): Promise<void>;
-    updateRun(runId: string, patch: Partial<Run>): Promise<void>;
-    appendSpan(span: Span): Promise<void>;
-    updateSpan(spanId: string, patch: Partial<Span>): Promise<void>;
-    appendEvent(event: TraceEvent): Promise<void>;
-    appendArtifact(artifact: Artifact): Promise<void>;
-    appendBudgetEntry(entry: BudgetLedgerEntry): Promise<void>;
-    getRun(runId: string): Promise<Run | undefined>;
-    listRuns(filter?: RunFilter): Promise<Run[]>;
-    spans(filter?: SpanFilter): Promise<Span[]>;
-    events(filter?: EventFilter): Promise<TraceEvent[]>;
-    budget(runId: string): Promise<BudgetLedgerEntry[]>;
-    artifacts(runId: string): Promise<Artifact[]>;
-}
-interface FileSystemTraceStoreOptions {
-    dir: string;
-    /** Roll over NDJSON files when they exceed this size in bytes. Default 32 MB. */
-    maxBytes?: number;
-}
-declare class FileSystemTraceStore implements TraceStore {
-    private dir;
-    private maxBytes;
-    /** Lazy in-memory index for queries — populated on first read. */
-    private index?;
-    /** Memoized index build — concurrent first reads share one build, and an
-     *  append racing an in-flight load awaits this so its row isn't lost. */
-    private indexPromise?;
-    /** Strictly-increasing rollover stamp. Date.now() alone collides when two
-     *  rollovers land in the same millisecond, overwriting a rolled file. */
-    private lastRolloverStamp;
-    /**
-     * Per-file append serialization. stat → conditional-rename → appendFile is a
-     * read-modify-write on the active file; without a lock two concurrent appends
-     * to the same `name` can both pass the size check (exceeding maxBytes) or one
-     * can append to a file the other just renamed away. Each entry chains the
-     * next append behind the prior one for that file.
-     */
-    private appendLocks;
-    constructor(options: FileSystemTraceStoreOptions);
-    private ensureDir;
-    private append;
-    private appendLocked;
-    private insertInto;
-    private load;
-    private buildIndex;
-    appendRun(run: Run): Promise<void>;
-    updateRun(runId: string, patch: Partial<Run>): Promise<void>;
-    appendSpan(span: Span): Promise<void>;
-    updateSpan(spanId: string, patch: Partial<Span>): Promise<void>;
-    appendEvent(event: TraceEvent): Promise<void>;
-    appendArtifact(artifact: Artifact): Promise<void>;
-    appendBudgetEntry(entry: BudgetLedgerEntry): Promise<void>;
-    getRun(runId: string): Promise<Run | undefined>;
-    listRuns(filter?: RunFilter): Promise<Run[]>;
-    spans(filter?: SpanFilter): Promise<Span[]>;
-    events(filter?: EventFilter): Promise<TraceEvent[]>;
-    budget(runId: string): Promise<BudgetLedgerEntry[]>;
-    artifacts(runId: string): Promise<Artifact[]>;
-}
-
-export { type EventFilter as E, FileSystemTraceStore as F, InMemoryTraceStore as I, type RunFilter as R, type SpanFilter as S, type TraceStore as T, type FileSystemTraceStoreOptions as a };