@sverka/arena 0.2.25 → 0.2.26

package/dist/index.d.mts

@@ -440,9 +440,11 @@ export declare function judgeAllRuns(results: RunResult[], tasks: Task[], config
 export declare function compareCombos(baseline: RunResult, candidate: RunResult): ComboComparison;
 //#endregion
 //#region src/config.d.ts
+type ArenaErrorCode = "CONFIG_INVALID" | "CONFIG_NOT_FOUND" | "UNKNOWN_AGENT" | "SCHEMA_INVALID" | "PUBLISH_CONFLICT" | "REGISTRY_UNAVAILABLE" | "PACK_NOT_FOUND" | "PACK_INVALID";
 export declare class ArenaError extends Error {
-  readonly code: "CONFIG_INVALID" | "CONFIG_NOT_FOUND" | "UNKNOWN_AGENT";
-  constructor(message: string, code: "CONFIG_INVALID" | "CONFIG_NOT_FOUND" | "UNKNOWN_AGENT");
+  readonly code: ArenaErrorCode;
+  readonly cause: unknown;
+  constructor(message: string, code: ArenaErrorCode, cause?: unknown);
 }
 export declare function resolveAdapter(name: string): AgentAdapter;
 declare const configFileSchema: z.ZodObject<{
@@ -561,5 +563,299 @@ export declare function arenaResultToEvents(result: ArenaResult): RunEvent[];
 /** Render the whole arena matrix as a standalone sverka HTML report. */
 export declare function writeAggregateReport(result: ArenaResult, outputPath: string, context?: ReportContext): void;
 //#endregion
-export type { ActionStep, AgentAdapter, AgentProcess, AgentSpawnConfig, AggregateMetrics, ArenaConfig, ArenaConfigFile, ArenaResult, CaseAnalysis, CheckResult, ComboComparison, DeterministicCheck, JudgeConfig, JudgeVerdict, ModelConfig, Observation, PluginConfig, RunMetrics, RunResult, Task, ToolCall, ToolCallMessage, ToolCallUpdateMessage, TraceData, TraceStep };
+//#region src/registry.d.ts
+interface TaskResult {
+  readonly task: string;
+  /** sha256 of the task prompt — a prompt edit breaks comparability. */
+  readonly promptHash: string;
+  readonly score: {
+    readonly passed: boolean;
+    readonly findings: number;
+  };
+  readonly metrics: {
+    readonly tokens?: number;
+    readonly toolCalls?: number;
+    readonly durationMs: number;
+    readonly stopReason?: string;
+  };
+  readonly traceRef?: string;
+}
+interface ArenaResultV1 {
+  readonly schema: "arena.result/v1";
+  readonly runId: string;
+  readonly pack: string;
+  readonly agent: string;
+  readonly model: string;
+  readonly plugins: readonly string[];
+  readonly sverkaVersion: string;
+  readonly startedAt: string;
+  readonly tasks: readonly TaskResult[];
+}
+export declare const arenaResultV1Schema: z.ZodObject<{
+  schema: z.ZodLiteral<"arena.result/v1">;
+  runId: z.ZodString;
+  pack: z.ZodString;
+  agent: z.ZodString;
+  model: z.ZodString;
+  plugins: z.ZodArray<z.ZodString>;
+  sverkaVersion: z.ZodString;
+  startedAt: z.ZodISODateTime;
+  tasks: z.ZodArray<z.ZodObject<{
+    task: z.ZodString;
+    promptHash: z.ZodString;
+    score: z.ZodObject<{
+      passed: z.ZodBoolean;
+      findings: z.ZodNumber;
+    }, z.core.$strip>;
+    metrics: z.ZodObject<{
+      tokens: z.ZodOptional<z.ZodNumber>;
+      toolCalls: z.ZodOptional<z.ZodNumber>;
+      durationMs: z.ZodNumber;
+      stopReason: z.ZodOptional<z.ZodString>;
+    }, z.core.$strip>;
+    traceRef: z.ZodOptional<z.ZodString>;
+  }, z.core.$strip>>;
+}, z.core.$strip>;
+/** Validate a parsed value as arena.result/v1 — names failing fields. */
+export declare function parseArenaResultV1(doc: unknown): ArenaResultV1;
+/** Canonical result path: results/<pack>/<agent>/<YYYY-MM-DD>/<runId>.json */
+export declare function resultPath(doc: ArenaResultV1, date?: string): string;
+/** sha256 of a task prompt — the comparability cohort key. */
+export declare function promptHash(prompt: string): string;
+/** A trace payload: a filesystem path to copy, or inline data. */
+type TraceInput = string | {
+  name: string;
+  data: unknown;
+};
+interface PublishOptions {
+  /** Trace files/payloads — written under traces/<runId>/<name>. */
+  traces?: readonly TraceInput[];
+  /** Override the YYYY-MM-DD partition (default: startedAt day). */
+  date?: string;
+}
+interface ListQuery {
+  pack?: string;
+  agent?: string;
+  /** ISO date/datetime lower bound on startedAt (inclusive). */
+  since?: string;
+}
+interface ArenaRegistry {
+  publish(result: ArenaResultV1, opts?: PublishOptions): Promise<string>;
+  list(query?: ListQuery): Promise<ArenaResultV1[]>;
+}
+interface GitRegistryConfig {
+  /** Clone URL — https or git@. */
+  url: string;
+  /** Branch to publish on (default: "main"). */
+  branch?: string;
+  /** Local checkout dir (default: a tmp cache keyed by URL). */
+  dir?: string;
+  /** Bearer token for https remotes (header-scoped, never persisted). */
+  token?: string;
+}
+export declare function createGitRegistry(cfg: GitRegistryConfig): ArenaRegistry;
+/** Minimal structural subset of the AWS SDK S3 client we depend on. */
+interface S3ClientLike {
+  putObject(input: {
+    Bucket: string;
+    Key: string;
+    Body: string | Uint8Array;
+  }): Promise<unknown>;
+  getObject?(input: {
+    Bucket: string;
+    Key: string;
+  }): Promise<{
+    Body?: unknown;
+  }>;
+  listObjectsV2?(input: {
+    Bucket: string;
+    Prefix?: string;
+    ContinuationToken?: string;
+  }): Promise<{
+    Contents?: {
+      Key?: string;
+    }[];
+    IsTruncated?: boolean;
+    NextContinuationToken?: string;
+  }>;
+}
+interface S3RegistryConfig {
+  bucket: string;
+  /** Key prefix for the whole registry tree (default: ""). */
+  prefix?: string;
+  /** Client — injectable so tests need no AWS SDK. When omitted, the
+   * optional `@aws-sdk/client-s3` peer is imported lazily. */
+  client?: S3ClientLike;
+}
+export declare function createS3Registry(cfg: S3RegistryConfig): ArenaRegistry;
+/** Registry over a local directory — a git checkout or a plain tree. */
+export declare function createFileRegistry(dir: string): ArenaRegistry;
+/**
+ * Publish several documents under ONE finalize — a matrix explosion
+ * produces one commit+push, not N. Internal: package code only.
+ */
+export declare function publishBatch(registry: ArenaRegistry, docs: readonly {
+  doc: ArenaResultV1;
+  opts?: PublishOptions;
+}[], message: string): Promise<string[]>;
+/**
+ * Rebuild index.json by scanning the results/ tree — the CI job that
+ * denormalizes after publish pushes (`arena reindex`).
+ */
+export declare function reindexRegistry(registry: ArenaRegistry): Promise<{
+  runs: number;
+}>;
+/**
+ * Resolve a registry reference to an ArenaRegistry:
+ *   s3://bucket[/prefix]        → s3 backend
+ *   file:///abs/path | ./dir    → file backend
+ *   git::<url> | <git-url>      → git backend
+ */
+export declare function openRegistry(ref: string, opts?: {
+  token?: string;
+  dir?: string;
+  client?: S3ClientLike;
+}): ArenaRegistry;
+/**
+ * Resolve a registry ref to a local directory — file refs return their
+ * path, git refs return the (cloned) checkout dir. s3 registries have no
+ * local dir — callers that need one must error themselves.
+ */
+export declare function resolveRegistryDir(ref: string, opts?: {
+  token?: string;
+  dir?: string;
+}): Promise<string>;
+//#endregion
+//#region src/publish.d.ts
+interface PublishContext {
+  /** Task-pack name — the leaderboard's first-level grouping. */
+  pack: string;
+  /** The run's AgentAdapter.id (matrix results carry no agent field). */
+  agent: string;
+  /** sverka version stamp — part of the comparability cohort. */
+  sverkaVersion: string;
+  /** startedAt override (default: result.timestamp). */
+  startedAt?: string;
+  /** taskId → prompt supplement — used when the file's analysis section
+   * does not carry the prompt (e.g. hand-shaped results). */
+  prompts?: Record<string, string>;
+  /** Trace file paths attached to v1-document publishes (--trace). */
+  traces?: readonly string[];
+  /** Deterministic runId generator for tests: called per cell. */
+  runId?: (cell: {
+    model: string;
+    plugins: string[];
+  }, index: number) => string;
+}
+/** Unique per-document run id — readable, sortable, collision-free. */
+export declare function newRunId(): string;
+/**
+ * Explode a matrix {@link ArenaResult} into one `arena.result/v1`
+ * document per (model, plugin-set) cell.
+ */
+export declare function explodeResult(result: ArenaResult, ctx: PublishContext): ArenaResultV1[];
+/**
+ * Explode a matrix result and publish every cell document (with its
+ * traces) to the registry. Returns the registry-relative result paths.
+ */
+export declare function publishResult(result: ArenaResult, registry: ArenaRegistry, ctx: PublishContext): Promise<string[]>;
+/**
+ * Publish a results file — accepts either an `arena.result/v1` document
+ * (or an array of them, published as-is) or a matrix `results.json`
+ * (exploded per cell). Returns the registry-relative result paths.
+ */
+export declare function publishFile(file: string, registry: ArenaRegistry, ctx: PublishContext): Promise<string[]>;
+//#endregion
+//#region src/board.d.ts
+interface BoardRow {
+  readonly agent: string;
+  readonly model: string;
+  /** passed runs / total runs, 0..1 */
+  readonly successRate: number;
+  readonly medianTokens?: number;
+  readonly medianDurationMs: number;
+  readonly runs: number;
+  /** Success rate per day for the last `days` days — NaN on days with no runs. */
+  readonly trend: readonly number[];
+}
+interface BoardCohort {
+  readonly pack: string;
+  readonly task: string;
+  readonly sverkaVersion: string;
+  readonly promptHash: string;
+  readonly rows: readonly BoardRow[];
+}
+interface BuildBoardOptions {
+  /** Trend window length in days (default: 30). */
+  days?: number;
+  /** Anchor date for the trend window (default: the cohort's latest sample). */
+  now?: Date;
+}
+/**
+ * Aggregate published results into board cohorts. The cohort key is
+ * (pack, task, sverkaVersion, promptHash) — a board row is only ever a
+ * comparison inside one prompt+toolchain generation.
+ */
+export declare function buildBoard(results: readonly ArenaResultV1[], opts?: BuildBoardOptions): BoardCohort[];
+/** Sparkline over the trend array — NaN days render as a gap dot. */
+export declare function sparkline(trend: readonly number[]): string;
+/** Render the leaderboard as a text table — `sverka-arena board`. */
+export declare function renderBoard(cohorts: readonly BoardCohort[]): string;
+/**
+ * Render the leaderboard as a self-contained static HTML page — no JS,
+ * no server. Used by `sverka-arena board --format html` and the website
+ * arena page (generated at build time from the registry).
+ */
+export declare function renderBoardHtml(cohorts: readonly BoardCohort[], opts?: {
+  title?: string;
+  generatedAt?: string;
+}): string;
+//#endregion
+//#region src/pack.d.ts
+interface PackDefaults {
+  readonly timeoutMs?: number;
+  readonly repetitions?: number;
+  readonly outputDir?: string;
+}
+interface TaskPack {
+  readonly name: string;
+  readonly version?: string;
+  readonly description?: string;
+  /** Local directory the pack was loaded from. */
+  readonly dir: string;
+  readonly tasks: readonly Task[];
+  readonly defaults: PackDefaults;
+}
+interface PackLint {
+  readonly errors: readonly string[];
+  readonly warnings: readonly string[];
+}
+interface LoadOptions {
+  /** Where `repo:` fixtures are cloned (default: tmp cache). */
+  cacheDir?: string;
+}
+/** Load a pack directory into a {@link TaskPack}. */
+export declare function loadPack(dir: string, opts?: LoadOptions): Promise<TaskPack>;
+/**
+ * Validate a pack directory without loading it into a run — collects
+ * errors AND warnings instead of throwing on the first problem.
+ */
+export declare function lintPack(dir: string): Promise<PackLint>;
+/** Scaffold a new pack at `<dir>` — `sverka-arena pack init <name>`. */
+export declare function initPack(dir: string, name: string): Promise<void>;
+interface ResolvePackOptions extends LoadOptions {
+  /** Registry ref for bare pack names (`packs/<name>/` inside it). */
+  registry?: string;
+  /** Bearer token for https git registries. */
+  token?: string;
+}
+/**
+ * Resolve a pack reference to a loaded {@link TaskPack}:
+ *   <dir>          — a local pack directory
+ *   <git-url>      — cloned into a tmp cache, then loaded
+ *   <bare-name>    — packs/<name>/ inside --registry (file or git)
+ */
+export declare function resolvePack(ref: string, opts?: ResolvePackOptions): Promise<TaskPack>;
+//#endregion
+export type { ActionStep, AgentAdapter, AgentProcess, AgentSpawnConfig, AggregateMetrics, ArenaConfig, ArenaConfigFile, ArenaErrorCode, ArenaRegistry, ArenaResult, ArenaResultV1, BoardCohort, BoardRow, BuildBoardOptions, CaseAnalysis, CheckResult, ComboComparison, DeterministicCheck, GitRegistryConfig, JudgeConfig, JudgeVerdict, ListQuery, ModelConfig, Observation, PackDefaults, PackLint, PluginConfig, PublishContext, PublishOptions, ResolvePackOptions, RunMetrics, RunResult, S3ClientLike, S3RegistryConfig, Task, TaskPack, TaskResult, ToolCall, ToolCallMessage, ToolCallUpdateMessage, TraceData, TraceInput, TraceStep };
 //# sourceMappingURL=index.d.mts.map

package/dist/index.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.mts","names":[],"sources":["../src/types.ts","../src/adapters/devin.ts","../src/acp.ts","../src/runner.ts","../src/judge.ts","../src/config.ts","../src/report.ts","../src/trace-report.ts","../src/aggregate-report.ts"],"mappings":";;;;;;;;;UACiB;EACf;EACA,MAAM,QAAQ,mBAAmB;;UAGlB;EACf,OAAO;EACP;EACA,SAAS;EACT;EACA,MAAM;;UAGS;EACf,IAAI,gBAAgB,oBAAoB,QAAQ;EAChD;;;UAIe;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;;EAEA;;EAEA;;;;;EAKA;;;;EAIA;;;EAGA,SAAS;;;UAIM;EACf;EACA;EACA;;;UAIe;EACf,OAAO;EACP,OAAO;EACP,QAAQ;EACR,SAAS;;EAET;EACA;EACA;;EAEA,QAAQ;;;;;EAKR;;;UAIe;;EAEf,OAAO;;EAEP,OAAO;;EAEP;;EAEA;;EAEA;;;UAIe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA,SAAS;EACT,OAAO;;EAEP;EACA;;EAEA,cAAc;EACd;;EAEA,UAAU;;;UAIK;EACf;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA,OAAO;EACP;IACE;IACA;IACA;IACA;;;;UAKa;EACf;EACA;EACA;EACA;EACA,YAAY;EACZ,eAAe;EACf;;;UAIe;EACf;EACA,WAAW;EACX;;EAEA;;;UAIe;EACf;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA;;;UAIe;EACf;EACA;IACE;IACA;IACA;IACA;IACA;;EAEF,SAAS;EACT,YAAY;;EAEZ,UAAU;;;UAIK;EACf;EACA;EACA;;EAEA,aAAa;;EAEb;;;UAIe;;EAEf;EACA;;EAEA;EACA;EACA;EACA;EACA;;EAEA;;;;;UC/MQ;EACR;EACA;EACA;EACA;EACA;IACE;IACA;MACE;MACA;;;;;UAMI;EACR;EACA;EACA;EACA;;;UAIQ;EACR;EACA;EACA;IACE;IACA;IACA;;EAEF,eAAe;EACf,OAAO;;;wBAMO;;wBA+MA,cAAc,YAAY;;wBAmC1B,mBAAmB,OAAO;EACxC;EACA;EACA;EACA;;;;;;;;;;;wBA0DoB,eACpB,mBACA,SAAS,iBACR;;;;;;;;;;;;qBAsHU,wBAAwB;WAC1B;EAET,MAAM,QAAQ,mBAAmB;;;;;KCpevB,kBAAkB,QAC5B;EACE;;;KAIQ,wBAAwB,QAClC;EACE;;;;;;;;;;;;wBA2EY,YAAY,QAAQ,mBAAmB;;;;;;;;wBAgBvC,2BAA2B,IAAI,qBAC7C,IAAI,0BACJ,IAAI;;;;;;;;;;;;;;;wBAsFgB,cACpB,MAAM,cACN,mBACA,gBACA,mBACA,cAAc,QAAQ,iBAAiB,6BACvC,gBAAgB,QAAQ,uBAAuB,6BAC/C,aAAa,6BACZ,QAAQ;;;;;;;;;;;;;wBC3KK,mBAAmB,SAAS,iBAAiB;;;;;;;;;wBA0BvC,SAAS,QAAQ,cAAc,QAAQ;;;;;;;;wBA2mB7C,iBACd,SAAS,aACT,SAAS,GAAG,uBACZ,iBACC;;;;;;;;;;wBAoIa,gBACd,SAAS,aACT,OAAO,SACN;;;;qBC5yBU;;;;;;;;;wBAkBG,iBACd,MAAM,MACN,mBACA,QAAQ,aACR;;;;;;;;wBA0Cc,mBACd,kBACA,kBACA,KAAK,YACJ;;;;;wBAoDmB,SACpB,KAAK,WACL,MAAM,MACN,QAAQ,aACR,mBACC,QAAQ;;;;;;;wBAmCW,aACpB,SAAS,aACT,OAAO,QACP,QAAQ,cACP,QAAQ;;;;;;;;;;;;wBAoCK,cACd,UAAU,WACV,WAAW,YACV;;;qBCnNU,mBAAmB;WAGnB;EADT,YAAA,iBACS;;wBAaG,eAAe,eAAe;cA4CxC,kBAAgB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkBpB,EAAA,KAAA;;KAGU,kBAAkB,EAAE,aAAa;;wBAG7B,aAAa,QAAQ,kBAAkB;;;;;wBAUjC,gBACpB,qBACC,QAAQ;;;;wBC1DK,aAAa,QAAQ;;;;KClChC;;UASY;EACf;EACA,MAAM;EACN;EACA;;EAEA;;EAEA;;;;;;;wBAuCc,eAAe,OAAO,YAAY;;;;;;;;wBA6FlC,iBACd,QAAQ,WACR;EAAS;EAAgB;IACxB;;;;;wBAgEa,WAAW,kBAAkB,eAAe;;wBAsB5C,iBACd,QAAQ,WACR,oBACA,UAAU;;;;wBC9OI,UAAU,QAAQ,WAAW;;;;;wBAiF7B,iBAAiB,QAAQ,cAAc;;;;;;;wBA4BvC,oBAAoB,QAAQ,cAAc;;wBA4E1C,qBACd,QAAQ,aACR,oBACA,UAAU"}
+{"version":3,"file":"index.d.mts","names":[],"sources":["../src/types.ts","../src/adapters/devin.ts","../src/acp.ts","../src/runner.ts","../src/judge.ts","../src/config.ts","../src/report.ts","../src/trace-report.ts","../src/aggregate-report.ts","../src/registry.ts","../src/publish.ts","../src/board.ts","../src/pack.ts"],"mappings":";;;;;;;;;UACiB;EACf;EACA,MAAM,QAAQ,mBAAmB;;UAGlB;EACf,OAAO;EACP;EACA,SAAS;EACT;EACA,MAAM;;UAGS;EACf,IAAI,gBAAgB,oBAAoB,QAAQ;EAChD;;;UAIe;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;;EAEA;;EAEA;;;;;EAKA;;;;EAIA;;;EAGA,SAAS;;;UAIM;EACf;EACA;EACA;;;UAIe;EACf,OAAO;EACP,OAAO;EACP,QAAQ;EACR,SAAS;;EAET;EACA;EACA;;EAEA,QAAQ;;;;;EAKR;;;UAIe;;EAEf,OAAO;;EAEP,OAAO;;EAEP;;EAEA;;EAEA;;;UAIe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA,SAAS;EACT,OAAO;;EAEP;EACA;;EAEA,cAAc;EACd;;EAEA,UAAU;;;UAIK;EACf;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;UAIe;EACf;EACA;EACA,OAAO;EACP;IACE;IACA;IACA;IACA;;;;UAKa;EACf;EACA;EACA;EACA;EACA,YAAY;EACZ,eAAe;EACf;;;UAIe;EACf;EACA,WAAW;EACX;;EAEA;;;UAIe;EACf;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA;;;UAIe;EACf;EACA;IACE;IACA;IACA;IACA;IACA;;EAEF,SAAS;EACT,YAAY;;EAEZ,UAAU;;;UAIK;EACf;EACA;EACA;;EAEA,aAAa;;EAEb;;;UAIe;;EAEf;EACA;;EAEA;EACA;EACA;EACA;EACA;;EAEA;;;;;UC/MQ;EACR;EACA;EACA;EACA;EACA;IACE;IACA;MACE;MACA;;;;;UAMI;EACR;EACA;EACA;EACA;;;UAIQ;EACR;EACA;EACA;IACE;IACA;IACA;;EAEF,eAAe;EACf,OAAO;;;wBAMO;;wBA+MA,cAAc,YAAY;;wBAmC1B,mBAAmB,OAAO;EACxC;EACA;EACA;EACA;;;;;;;;;;;wBA0DoB,eACpB,mBACA,SAAS,iBACR;;;;;;;;;;;;qBAsHU,wBAAwB;WAC1B;EAET,MAAM,QAAQ,mBAAmB;;;;;KCpevB,kBAAkB,QAC5B;EACE;;;KAIQ,wBAAwB,QAClC;EACE;;;;;;;;;;;;wBA2EY,YAAY,QAAQ,mBAAmB;;;;;;;;wBAgBvC,2BAA2B,IAAI,qBAC7C,IAAI,0BACJ,IAAI;;;;;;;;;;;;;;;wBAsFgB,cACpB,MAAM,cACN,mBACA,gBACA,mBACA,cAAc,QAAQ,iBAAiB,6BACvC,gBAAgB,QAAQ,uBAAuB,6BAC/C,aAAa,6BACZ,QAAQ;;;;;;;;;;;;;wBC3KK,mBAAmB,SAAS,iBAAiB;;;;;;;;;wBA0BvC,SAAS,QAAQ,cAAc,QAAQ;;;;;;;;wBA2mB7C,iBACd,SAAS,aACT,SAAS,GAAG,uBACZ,iBACC;;;;;;;;;;wBAoIa,gBACd,SAAS,aACT,OAAO,SACN;;;;qBC5yBU;;;;;;;;;wBAkBG,iBACd,MAAM,MACN,mBACA,QAAQ,aACR;;;;;;;;wBA0Cc,mBACd,kBACA,kBACA,KAAK,YACJ;;;;;wBAoDmB,SACpB,KAAK,WACL,MAAM,MACN,QAAQ,aACR,mBACC,QAAQ;;;;;;;wBAmCW,aACpB,SAAS,aACT,OAAO,QACP,QAAQ,cACP,QAAQ;;;;;;;;;;;;wBAoCK,cACd,UAAU,WACV,WAAW,YACV;;;KCnNS;qBAUC,mBAAmB;WAKnB,MAAM;WAJC;EAGhB,YAAA,iBACS,MAAM,gBACf;;wBAgBY,eAAe,eAAe;cA4CxC,kBAAgB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkBpB,EAAA,KAAA;;KAGU,kBAAkB,EAAE,aAAa;;wBAG7B,aAAa,QAAQ,kBAAkB;;;;;wBAUjC,gBACpB,qBACC,QAAQ;;;;wBC1EK,aAAa,QAAQ;;;;KClChC;;UASY;EACf;EACA,MAAM;EACN;EACA;;EAEA;;EAEA;;;;;;;wBAuCc,eAAe,OAAO,YAAY;;;;;;;;wBA6FlC,iBACd,QAAQ,WACR;EAAS;EAAgB;IACxB;;;;;wBAgEa,WAAW,kBAAkB,eAAe;;wBAsB5C,iBACd,QAAQ,WACR,oBACA,UAAU;;;;wBC9OI,UAAU,QAAQ,WAAW;;;;;wBAiF7B,iBAAiB,QAAQ,cAAc;;;;;;;wBA4BvC,oBAAoB,QAAQ,cAAc;;wBA4E1C,qBACd,QAAQ,aACR,oBACA,UAAU;;;UCnLK;WACN;;WAEA;WACA;aAAkB;aAA0B;;WAC5C;aACE;aACA;aACA;aACA;;WAEF;;UAGM;WACN;WACA;WACA;WACA;WACA;WACA;WACA;WACA;WACA,gBAAgB;;qBAmBd,qBAAmB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;GAU9B,EAAA,KAAA;;wBAGc,mBAAmB,eAAe;;wBA6BlC,WAAW,KAAK,eAAe;;wBAiB/B,WAAW;;KAOf;EAAwB;EAAc;;UAEjC;;EAEf,kBAAkB;;EAElB;;UAGe;EACf;EACA;;EAEA;;UAGe;EACf,QAAQ,QAAQ,eAAe,OAAO,iBAAiB;EACvD,KAAK,QAAQ,YAAY,QAAQ;;UA+ElB;;EAEf;;EAEA;;EAEA;;EAEA;;wBA6Sc,kBAAkB,KAAK,oBAAoB;;UAU1C;EACf,UAAU;IACR;IACA;IACA,eAAe;MACb;EACJ,WAAW;IAAS;IAAgB;MAAgB;IAClD;;EAEF,eAAe;IACb;IACA;IACA;MACE;IACF;MAAa;;IACb;IACA;;;UAIa;EACf;;EAEA;;;EAGA,SAAS;;wBAgJK,iBAAiB,KAAK,mBAAmB;;wBAUzC,mBAAmB,cAAc;;;;;wBAuI3B,aACpB,UAAU,eACV;EAAiB,KAAK;EAAe,OAAO;KAC5C,kBACC;;;;;wBAoGmB,gBACpB,UAAU,gBACT;EAAU;;;;;;;;wBAsBG,aACd,aACA;EAAQ;EAAgB;EAAc,SAAS;IAC9C;;;;;;wBAkCmB,mBACpB,aACA;EAAQ;EAAgB;IACvB;;;UCz/Bc;;EAEf;;EAEA;;EAEA;;EAEA;;;EAGA,UAAU;;EAEV;;EAEA,SAAS;IAAQ;IAAe;KAAqB;;;wBAIvC;;;;;wBA2GA,cACd,QAAQ,aACR,KAAK,iBACJ;;;;;wBAsCmB,cACpB,QAAQ,aACR,UAAU,eACV,KAAK,iBACJ;;;;;;wBAkBmB,YACpB,cACA,UAAU,eACV,KAAK,iBACJ;;;UC3Mc;WACN;WACA;;WAEA;WACA;WACA;WACA;;WAEA;;UAGM;WACN;WACA;WACA;WACA;WACA,eAAe;;UAuBT;;EAEf;;EAEA,MAAM;;;;;;;wBAQQ,WACd,kBAAkB,iBAClB,OAAM,oBACL;;wBAiHa,UAAU;;wBA6BV,YAAY,kBAAkB;;;;;;wBA6D9B,gBACd,kBAAkB,eAClB;EAAQ;EAAgB;;;;UCjNT;WACN;WACA;WACA;;UAGM;WACN;WACA;WACA;;WAEA;WACA,gBAAgB;WAChB,UAAU;;UAGJ;WACN;WACA;;UAKD;;EAER;;;wBAoFoB,SACpB,aACA,OAAM,cACL,QAAQ;;;;;wBAkFW,SAAS,cAAc,QAAQ;;wBAwF/B,SAAS,aAAa,eAAe;UAkH1C,2BAA2B;;EAE1C;;EAEA;;;;;;;;wBASoB,YACpB,aACA,OAAM,qBACL,QAAQ"}

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { C as installPlugins, D as sanitizeEnv, E as runAcpSession, S as extractUsageTokens, T as createPermissionHandler, _ as judgeAllRuns, a as renderReport, b as DevinAdapter, c as loadArenaConfig, d as computeAnalysis, f as pluginCombinations, g as compareCombos, h as buildJudgePrompt, i as writeAggregateReport, l as resolveAdapter, m as DEFAULT_JUDGE_PROMPT, n as arenaResultToEvents, o as ArenaError, p as runArena, r as runStepId, s as defineConfig, t as arenaResultGraph, u as aggregateResults, v as judgeRun, w as transcriptDir, x as countLlmCalls, y as parseJudgeResponse } from "./aggregate-report-BkLMmG16.mjs";
+import { A as defineConfig, B as judgeAllRuns, C as resultPath, D as writeAggregateReport, E as runStepId, F as pluginCombinations, G as extractUsageTokens, H as parseJudgeResponse, I as runArena, J as createPermissionHandler, K as installPlugins, L as DEFAULT_JUDGE_PROMPT, M as resolveAdapter, N as aggregateResults, O as renderReport, P as computeAnalysis, R as buildJudgePrompt, S as resolveRegistryDir, T as arenaResultToEvents, U as DevinAdapter, V as judgeRun, W as countLlmCalls, X as sanitizeEnv, Y as runAcpSession, _ as openRegistry, a as buildBoard, b as publishBatch, c as sparkline, d as publishFile, f as publishResult, g as createS3Registry, h as createGitRegistry, i as resolvePack, j as loadArenaConfig, k as ArenaError, l as explodeResult, m as createFileRegistry, n as lintPack, o as renderBoard, p as arenaResultV1Schema, q as transcriptDir, r as loadPack, s as renderBoardHtml, t as initPack, u as newRunId, v as parseArenaResultV1, w as arenaResultGraph, x as reindexRegistry, y as promptHash, z as compareCombos } from "./pack-DoL1QL1y.mjs";
 import { createHtmlRenderer } from "@sverka/reporter";
 //#region src/trace-report.ts
 function weight(kind) {
@@ -219,6 +219,6 @@ function writeTraceReport(result, outputPath, context) {
 	renderer.flush();
 }
 //#endregion
-export { ArenaError, DEFAULT_JUDGE_PROMPT, DevinAdapter, aggregateResults, arenaResultGraph, arenaResultToEvents, buildJudgePrompt, compareCombos, computeAnalysis, countLlmCalls, createPermissionHandler, defineConfig, extractUsageTokens, installPlugins, judgeAllRuns, judgeRun, loadArenaConfig, parseJudgeResponse, pluginCombinations, renderReport, resolveAdapter, runAcpSession, runArena, runStepId, sanitizeEnv, traceGraph, traceToActions, traceToRunEvents, transcriptDir, writeAggregateReport, writeTraceReport };
+export { ArenaError, DEFAULT_JUDGE_PROMPT, DevinAdapter, aggregateResults, arenaResultGraph, arenaResultToEvents, arenaResultV1Schema, buildBoard, buildJudgePrompt, compareCombos, computeAnalysis, countLlmCalls, createFileRegistry, createGitRegistry, createPermissionHandler, createS3Registry, defineConfig, explodeResult, extractUsageTokens, initPack, installPlugins, judgeAllRuns, judgeRun, lintPack, loadArenaConfig, loadPack, newRunId, openRegistry, parseArenaResultV1, parseJudgeResponse, pluginCombinations, promptHash, publishBatch, publishFile, publishResult, reindexRegistry, renderBoard, renderBoardHtml, renderReport, resolveAdapter, resolvePack, resolveRegistryDir, resultPath, runAcpSession, runArena, runStepId, sanitizeEnv, sparkline, traceGraph, traceToActions, traceToRunEvents, transcriptDir, writeAggregateReport, writeTraceReport };
 
 //# sourceMappingURL=index.mjs.map