@tangle-network/agent-eval 0.117.0 → 0.118.0

package/dist/benchmarks/index.js

@@ -17,24 +17,26 @@ import {
   runBenchmarkAdapter,
   summarizeBenchmarkCampaign
 } from "../chunk-JSDVRFAP.js";
-import "../chunk-ZUXV7UWZ.js";
-import "../chunk-HQPHZGL6.js";
+import "../chunk-Q442S5AS.js";
+import "../chunk-YLMUS4MM.js";
 import "../chunk-IDZTTFRR.js";
 import "../chunk-3YYRZDON.js";
-import "../chunk-MGEHEHSN.js";
+import "../chunk-OIIMMLRB.js";
 import "../chunk-ARU2PZFM.js";
-import "../chunk-PJQFMIOX.js";
-import "../chunk-4JLWXDYA.js";
-import "../chunk-GGE4NNQT.js";
-import "../chunk-S2F4J57L.js";
-import "../chunk-5UF54T55.js";
-import "../chunk-XJYR7XFV.js";
-import "../chunk-VSMTAMNK.js";
 import "../chunk-NJC7U437.js";
+import "../chunk-PJQFMIOX.js";
 import "../chunk-VCTY3W6J.js";
 import "../chunk-VI2UW6B6.js";
+import "../chunk-3EHHMC6E.js";
+import "../chunk-GGE4NNQT.js";
+import "../chunk-QKEGNI5B.js";
 import "../chunk-PC4UYEBM.js";
+import "../chunk-S3UZOQ5Y.js";
+import "../chunk-MA6HLL3S.js";
+import "../chunk-XJYR7XFV.js";
+import "../chunk-VSMTAMNK.js";
 import "../chunk-ONWEPEDO.js";
+import "../chunk-WDHBCA3M.js";
 import "../chunk-PZ5AY32C.js";
 export {
   BENCHMARK_SPLIT_SEED,

package/dist/builder-eval/index.d.ts

@@ -1,7 +1,452 @@
-import { S as SandboxDriver, H as HarnessConfig, a as SandboxHarnessResult, T as TestGradedScenario, b as TestGradedRunResult } from '../test-graded-scenario-B0ybnPY7.js';
-import { T as TraceEmitter } from '../emitter-CjD7vUwv.js';
-import { R as Run } from '../schema-B3Q3l9Z_.js';
-import { T as TraceStore } from '../store-DGqD0Pyo.js';
+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;
+    /** All generated tokens, including the reasoning subset when present. */
+    outputTokens?: number;
+    cachedTokens?: number;
+    cacheWriteTokens?: number;
+    /** Reasoning-token subset of `outputTokens`. */
+    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';
+
+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[]>;
+}
+
+/**
+ * TraceEmitter — hierarchical span builder that auto-parents using an
+ * internal stack. One emitter per Run; emitters do NOT share state.
+ *
+ * Convenience methods (`llm`, `tool`, `retrieval`, `judge`, `sandbox`)
+ * return a `SpanHandle` with `.end()` / `.fail()` so callers don't
+ * have to thread spanIds manually. For async workflows that can't use
+ * the stack (e.g. fan-out parallel calls), pass `parentSpanId`
+ * explicitly.
+ */
+
+interface SpanHandle<S extends Span = Span> {
+    span: S;
+    end(patch?: Partial<S>): Promise<void>;
+    fail(error: string | Error, patch?: Partial<S>): Promise<void>;
+}
+interface RunCompleteHookContext {
+    runId: string;
+    emitter: TraceEmitter;
+    store: TraceStore;
+    /** Outcome the caller passed to `endRun` (undefined for `abortRun`). */
+    outcome?: RunOutcome;
+    /** Final run status. */
+    status: 'completed' | 'failed' | 'aborted';
+}
+type RunCompleteHook = (ctx: RunCompleteHookContext) => Promise<void> | void;
+interface TraceEmitterOptions {
+    runId?: string;
+    /** Inject a clock for deterministic tests. */
+    now?: () => number;
+    /** Inject an id generator for deterministic tests. */
+    id?: () => string;
+    /**
+     * Hooks fired after `endRun` / `abortRun` writes the final run state.
+     * Designed for trace-analyst auto-execution, integrity assertions, and
+     * outbound notifications. Hooks run sequentially in the order supplied.
+     *
+     * By default a hook that throws is swallowed and logged as a `note` event
+     * on the run — auto-orchestration must not crash the underlying flow.
+     * Set `hookErrors: 'throw'` to propagate.
+     */
+    onRunComplete?: RunCompleteHook[];
+    /** `'swallow'` (default) | `'throw'`. */
+    hookErrors?: 'swallow' | 'throw';
+}
+declare class TraceEmitter {
+    private store;
+    private stack;
+    private _runId;
+    private now;
+    private id;
+    private hooks;
+    private hookErrors;
+    constructor(store: TraceStore, options?: TraceEmitterOptions);
+    get runId(): string;
+    get traceStore(): TraceStore;
+    /** Append a hook after construction (e.g. attach the trace analyst). */
+    addRunCompleteHook(hook: RunCompleteHook): void;
+    /**
+     * Begin a Run.
+     *
+     * `scenarioId` is required on the persisted Run shape — every Run downstream
+     * gets a non-empty scenarioId so filters and aggregations stay simple — but
+     * the INPUT here accepts it as optional. When omitted, startRun substitutes
+     * a sensible default (`run.layer ?? run.tags?.['kind'] ?? 'runtime'`) so
+     * runtime / operator / meta-eval runs that have no curated-scenario corpus
+     * to anchor to don't have to invent placeholder strings at the call site.
+     */
+    startRun(run: Omit<Run, 'runId' | 'scenarioId' | 'startedAt' | 'status'> & {
+        scenarioId?: string;
+    }): Promise<Run>;
+    endRun(outcome?: RunOutcome): Promise<void>;
+    abortRun(reason: string): Promise<void>;
+    private runHooks;
+    span<S extends Span = Span>(init: {
+        kind: SpanKind;
+        name: string;
+        parentSpanId?: string;
+        attributes?: Record<string, unknown>;
+    } & Partial<Omit<S, 'spanId' | 'runId' | 'startedAt' | 'kind' | 'name'>>): Promise<SpanHandle<S>>;
+    private handle;
+    private pop;
+    llm(init: Omit<LlmSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>): Promise<SpanHandle<LlmSpan>>;
+    tool(init: Omit<ToolSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>): Promise<SpanHandle<ToolSpan>>;
+    retrieval(init: Omit<RetrievalSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>): Promise<SpanHandle<RetrievalSpan>>;
+    recordJudge(verdict: Omit<JudgeSpan, 'spanId' | 'runId' | 'kind' | 'startedAt' | 'endedAt'>): Promise<JudgeSpan>;
+    sandbox(init: Omit<SandboxSpan, 'spanId' | 'runId' | 'kind' | 'startedAt'>): Promise<SpanHandle<SandboxSpan>>;
+    emit(event: {
+        kind: EventKind;
+        spanId?: string;
+        payload?: Record<string, unknown>;
+    }): Promise<TraceEvent>;
+    recordBudget(entry: Omit<BudgetLedgerEntry, 'runId' | 'timestamp'> & {
+        timestamp?: number;
+    }): Promise<BudgetLedgerEntry>;
+    recordArtifact(artifact: Omit<Artifact, 'artifactId' | 'runId'>): Promise<Artifact>;
+    /**
+     * Runs `fn` inside a span; auto-ends on success, auto-fails on throw.
+     * Returns the fn's return value. Use this for the 95% case.
+     */
+    within<T>(init: Parameters<TraceEmitter['span']>[0], fn: (handle: SpanHandle) => Promise<T>): Promise<T>;
+}
+
+/**
+ * SandboxHarness — executes a scenario in an isolated environment and
+ * emits a rich SandboxSpan into the trace.
+ *
+ * Two built-in drivers:
+ *   - `SubprocessSandboxDriver` — spawn in a local cwd with env vars.
+ *     Fast, no dependencies, fine for unit tests and most CI gates.
+ *   - `DockerSandboxDriver` — lifted from tangle-router's sandbox path;
+ *     shells out to `docker run`. Stronger isolation, slower startup.
+ *
+ * Consumers implement `SandboxDriver` for custom backends (Firecracker,
+ * Cloudflare sandbox product, etc.). The harness doesn't care which.
+ */
+
+interface HarnessConfig {
+    /** Setup command (e.g. "pnpm install"). Non-zero exit fails the run. */
+    setupCommand?: string;
+    /** Run command (e.g. "pnpm build"). */
+    runCommand?: string;
+    /** Test command (e.g. "pnpm test --run"). Drives the test count + pass count. */
+    testCommand?: string;
+    /** Absolute cwd for the subprocess driver. Ignored by docker driver. */
+    cwd?: string;
+    /** Max wall-clock per phase in ms. Default 10 minutes. */
+    timeoutMs?: number;
+    /**
+     * Cap on captured stdout+stderr bytes per phase. A runaway process can
+     * otherwise grow the in-memory buffer without bound. Once hit, further
+     * output is dropped and `outputTruncated` is set. Default 16 MiB.
+     */
+    maxOutputBytes?: number;
+    /** Image for the docker driver. */
+    image?: string;
+    /** Extra env vars (validated; shell-escaped). */
+    env?: Record<string, string>;
+    /** Parser for the test output — maps stdout/stderr/exit code → pass count. */
+    testParser?: TestOutputParser;
+}
+interface TestOutputParser {
+    id: string;
+    parse(stdout: string, stderr: string, exitCode: number): {
+        testsTotal: number;
+        testsPassed: number;
+    } | undefined;
+}
+interface SandboxResult {
+    phase: 'setup' | 'run' | 'test';
+    exitCode: number;
+    stdout: string;
+    stderr: string;
+    wallMs: number;
+    testsTotal?: number;
+    testsPassed?: number;
+    /**
+     * True when the process was killed because it exceeded `timeoutMs`. A
+     * SIGKILLed child can still close with exit code 0; callers MUST treat
+     * a timed-out phase as a hard failure regardless of `exitCode`, never
+     * as a pass. `undefined`/`false` means the process completed on its own.
+     */
+    killedByTimeout?: boolean;
+    /**
+     * True when captured stdout/stderr hit `maxOutputBytes` and further
+     * output was dropped. The result is still returned (the process was
+     * not killed for this), but downstream parsers see truncated text.
+     */
+    outputTruncated?: boolean;
+}
+interface SandboxDriver {
+    id: string;
+    exec(phase: SandboxResult['phase'], command: string, config: HarnessConfig): Promise<SandboxResult>;
+}
+interface SandboxHarnessResult {
+    passed: boolean;
+    setup?: SandboxResult;
+    run?: SandboxResult;
+    test?: SandboxResult;
+    totalWallMs: number;
+    /** Final score — 0 when no tests; otherwise testsPassed/testsTotal. */
+    score: number;
+}
+
+/**
+ * TestGradedScenario — a scenario whose score comes from a test suite.
+ *
+ * This is the SWE-bench pattern generalized. The scenario ships:
+ *   - fixture data (setup instructions)
+ *   - a test command the harness runs
+ *   - optional assertion overrides
+ *
+ * The runner emits a run, delegates to SandboxHarness, records the
+ * outcome, and returns a structured verdict. Consumers bind their own
+ * agent execution to this contract.
+ */
+
+interface TestGradedScenario {
+    id: string;
+    description?: string;
+    harness: HarnessConfig;
+    /** Optional pass threshold in 0..1 (default 1.0 = all tests must pass). */
+    passThreshold?: number;
+    /** Provenance for dataset tracking. */
+    datasetVersion?: string;
+    /** Free-form tags (difficulty, category, etc.). */
+    tags?: Record<string, string>;
+}
+interface TestGradedRunResult {
+    runId: string;
+    scenario: TestGradedScenario;
+    harness: SandboxHarnessResult;
+    pass: boolean;
+    score: number;
+    failureClass?: FailureClass;
+}
 
 /**
  * BuilderSession — ties a builder-of-builders workflow together.

package/dist/builder-eval/index.js

@@ -1,17 +1,18 @@
 import {
   SandboxHarness,
   runTestGradedScenario
-} from "../chunk-HZHNRYHK.js";
+} from "../chunk-K6N6XJJX.js";
 import {
   pearsonR,
   spearmanR
 } from "../chunk-PJQFMIOX.js";
 import {
   judgeSpans
-} from "../chunk-LQUTGLOZ.js";
+} from "../chunk-ZET2UAYW.js";
 import {
   TraceEmitter
-} from "../chunk-TVVP3ZZQ.js";
+} from "../chunk-VQMK5FMP.js";
+import "../chunk-MA6HLL3S.js";
 import "../chunk-ONWEPEDO.js";
 import "../chunk-PZ5AY32C.js";
 

package/dist/builder-eval/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/builder-eval/builder-session.ts","../../src/builder-eval/correlation.ts","../../src/builder-eval/project-registry.ts","../../src/builder-eval/three-layer-eval.ts"],"sourcesContent":["/**\n * BuilderSession — ties a builder-of-builders workflow together.\n *\n * Models agent-builder's shape: Project → Chat → Edit → Ship → App →\n * AppAgent. Each layer is a Run (linked via parentRunId). The\n * framework-enforced invariants:\n *\n *   - One Project → many Chats; chatId scopes runs within a project.\n *   - One Chat = one builder Run with `layer='builder'`.\n *   - One Ship = one child Run with `layer='app-build'` + SandboxHarness.\n *   - One AppScenario = one grandchild Run with `layer='app-runtime'`.\n *\n * Consumers obtain a BuilderSession, call `startChat`, drive the\n * builder agent (emitting spans), and call `ship` / `runAppScenario`\n * as the workflow progresses. The session reconstructs itself from\n * trace data via `resume(store, projectId)`.\n */\n\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from '../sandbox-harness'\nimport { SandboxHarness } from '../sandbox-harness'\nimport type { TestGradedRunResult, TestGradedScenario } from '../test-graded-scenario'\nimport { runTestGradedScenario } from '../test-graded-scenario'\nimport { TraceEmitter } from '../trace/emitter'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BuilderSessionInit {\n  projectId: string\n  chatId?: string\n  /** Free-form: user's task description, project name, etc. Stored on the builder Run. */\n  tags?: Record<string, string>\n}\n\nexport interface ShipOptions {\n  harness: HarnessConfig\n  driver?: SandboxDriver\n  /** scenarioId of this app-build run. Defaults to `${projectId}/build`. */\n  scenarioId?: string\n}\n\nexport interface RunAppScenarioOptions {\n  scenario: TestGradedScenario\n  /** Harness driver override; defaults to the one the session was created with. */\n  driver?: SandboxDriver\n}\n\nexport class BuilderSession {\n  private store: TraceStore\n  private builderEmitter: TraceEmitter\n  readonly projectId: string\n  readonly chatId: string\n  private builderRunId?: string\n  private lastBuildRunId?: string\n  private defaultDriver?: SandboxDriver\n\n  constructor(store: TraceStore, init: BuilderSessionInit, driver?: SandboxDriver) {\n    this.store = store\n    this.projectId = init.projectId\n    this.chatId = init.chatId ?? cryptoId()\n    this.defaultDriver = driver\n    this.builderEmitter = new TraceEmitter(store)\n  }\n\n  /** Start the builder (L0) run for this chat. Returns the runId. */\n  async startChat(scenarioId = `${this.projectId}/chat`): Promise<string> {\n    await this.builderEmitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'builder',\n    })\n    this.builderRunId = this.builderEmitter.runId\n    return this.builderRunId\n  }\n\n  /** The emitter for builder-level spans (edits, LLM calls, tool invocations). */\n  get emitter(): TraceEmitter {\n    if (!this.builderRunId) throw new Error('BuilderSession.emitter: call startChat() first')\n    return this.builderEmitter\n  }\n\n  /**\n   * Ship the project's generated app: run the sandbox harness as a child\n   * Run (`layer='app-build'`). Returns the build result + runId.\n   */\n  async ship(options: ShipOptions): Promise<{ runId: string; result: SandboxHarnessResult }> {\n    if (!this.builderRunId) throw new Error('BuilderSession.ship: call startChat() first')\n    const buildEmitter = new TraceEmitter(this.store)\n    await buildEmitter.startRun({\n      scenarioId: options.scenarioId ?? `${this.projectId}/build`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    const harness = new SandboxHarness(options.driver ?? this.defaultDriver)\n    const result = await harness.run(options.harness, buildEmitter)\n    await buildEmitter.endRun({\n      pass: result.passed,\n      score: result.score,\n      failureClass: result.passed ? 'success' : 'sandbox_failure',\n    })\n    this.lastBuildRunId = buildEmitter.runId\n    return { runId: buildEmitter.runId, result }\n  }\n\n  /**\n   * Run a domain scenario against the just-built app as a grandchild Run\n   * (`layer='app-runtime'`). The `ship` call must precede this so the\n   * parent is set correctly; if no build exists yet the session attaches\n   * directly to the builder run (useful for prototypes).\n   */\n  async runAppScenario(options: RunAppScenarioOptions): Promise<TestGradedRunResult> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error('BuilderSession.runAppScenario: call startChat() + ship() first')\n    const { scenario, driver } = options\n    const result = await runTestGradedScenario(scenario, this.store, {\n      driver: driver ?? this.defaultDriver,\n      provenance: { codeSha: undefined, promptSha: undefined, modelFingerprint: undefined },\n    })\n    // Attach to the parent chain by updating the stored Run in place.\n    await this.store.updateRun(result.runId, {\n      parentRunId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'app-runtime',\n    })\n    return result\n  }\n\n  /** Record an end-of-chat meta score (judge verdict on whether the builder\n   *  served the user's intent). Accepts a numeric score + optional rationale. */\n  async recordMetaScore(score: number, rationale?: string): Promise<void> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordMetaScore: call startChat() first')\n    await this.builderEmitter.recordJudge({\n      judgeId: 'builder-meta',\n      targetSpanId: this.builderRunId, // attach to the builder run itself\n      dimension: 'user_intent_satisfaction',\n      score,\n      rationale,\n      name: 'builder-meta',\n    })\n  }\n\n  /** Close the builder Run with a final outcome. */\n  async endChat(outcome: { pass: boolean; score?: number; notes?: string }): Promise<void> {\n    await this.builderEmitter.endRun({\n      pass: outcome.pass,\n      score: outcome.score,\n      notes: outcome.notes,\n    })\n  }\n\n  /**\n   * Inline app-runtime run — for cases where the \"scenario\" isn't a\n   * SWE-bench-style test suite but a live agent interaction (LLM chat,\n   * domain flow). Returns an emitter bound to a fresh Run in the\n   * `app-runtime` layer; caller emits spans inside and calls\n   * `.endRun()` with the final verdict.\n   */\n  async startAppRuntime(scenarioId: string): Promise<TraceEmitter> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error(\n        'BuilderSession.startAppRuntime: call startChat() + (optionally) ship() first',\n      )\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId,\n      layer: 'app-runtime',\n    })\n    return emitter\n  }\n\n  /**\n   * Lightweight \"ship marker\" — record an app-build Run with a caller-\n   * provided verdict. Use when there isn't a sandbox harness to run but\n   * you still want to mark the build state at publish time.\n   */\n  async recordShipMarker(args: {\n    pass: boolean\n    score: number\n    scenarioId?: string\n    notes?: string\n  }): Promise<string> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordShipMarker: call startChat() first')\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId: args.scenarioId ?? `${this.projectId}/ship`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    await emitter.endRun({\n      pass: args.pass,\n      score: args.score,\n      failureClass: args.pass ? 'success' : 'sandbox_failure',\n      notes: args.notes,\n    })\n    this.lastBuildRunId = emitter.runId\n    return emitter.runId\n  }\n\n  get lastBuildRunIdValue(): string | undefined {\n    return this.lastBuildRunId\n  }\n  get builderRunIdValue(): string | undefined {\n    return this.builderRunId\n  }\n}\n\n/**\n * Reconstruct the most recent BuilderSession state for a given project —\n * returns { builderRunId, lastBuildRunId, chatRuns }. For chat-first UIs\n * this is how a resumed session finds its place in the edit history.\n */\nexport async function resumeBuilderSession(\n  store: TraceStore,\n  projectId: string,\n): Promise<{\n  projectId: string\n  chatRuns: Run[]\n  lastBuilderRun?: Run\n  lastBuildRun?: Run\n  lastAppRuntimeRuns: Run[]\n}> {\n  const runs = await store.listRuns({ projectId })\n  const chatRuns = runs\n    .filter((r) => r.layer === 'builder')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const buildRuns = runs\n    .filter((r) => r.layer === 'app-build')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const appRuntimeRuns = runs\n    .filter((r) => r.layer === 'app-runtime')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  return {\n    projectId,\n    chatRuns,\n    lastBuilderRun: chatRuns[0],\n    lastBuildRun: buildRuns[0],\n    lastAppRuntimeRuns: appRuntimeRuns,\n  }\n}\n\nfunction cryptoId(): string {\n  if (typeof globalThis.crypto?.randomUUID === 'function') return globalThis.crypto.randomUUID()\n  return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`\n}\n","/**\n * Meta-eval correlation — the highest-leverage signal in the framework.\n *\n * Given a corpus of three-layer project reports, compute how well each\n * pair of layers correlates. The question we care about most:\n *\n *   Does `metaScore` (what the builder thinks it did) predict\n *   `runtimeScore` (what the user actually gets)?\n *\n * If r < ~0.4, the builder's self-scoring is broken — it's optimizing\n * for something other than real-world success. If r > 0.7, meta_score\n * is a usable proxy and can drive CI gates cheaply.\n *\n * Non-parametric rank correlation (Spearman) is also reported because\n * meta scores are often ordinal-ish.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport type { ThreeLayerProjectReport } from './three-layer-eval'\n\nexport interface LayerCorrelation {\n  n: number\n  pearson: number\n  spearman: number\n}\n\nexport interface CorrelationReport {\n  /** Pairs present in the corpus (layers with ≥ 2 matched data points). */\n  metaVsBuild?: LayerCorrelation\n  metaVsRuntime?: LayerCorrelation\n  buildVsRuntime?: LayerCorrelation\n  /** Number of complete projects (all 3 scores present). */\n  completeProjects: number\n}\n\nexport function correlateLayers(reports: ThreeLayerProjectReport[]): CorrelationReport {\n  const completeProjects = reports.filter((r) => r.complete).length\n  return {\n    metaVsBuild: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.buildScore,\n    ),\n    metaVsRuntime: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.runtimeScore,\n    ),\n    buildVsRuntime: pairwise(\n      reports,\n      (r) => r.buildScore,\n      (r) => r.runtimeScore,\n    ),\n    completeProjects,\n  }\n}\n\nfunction pairwise(\n  reports: ThreeLayerProjectReport[],\n  a: (r: ThreeLayerProjectReport) => number | null,\n  b: (r: ThreeLayerProjectReport) => number | null,\n): LayerCorrelation | undefined {\n  const xs: number[] = []\n  const ys: number[] = []\n  for (const r of reports) {\n    const x = a(r)\n    const y = b(r)\n    if (x !== null && y !== null && Number.isFinite(x) && Number.isFinite(y)) {\n      xs.push(x)\n      ys.push(y)\n    }\n  }\n  if (xs.length < 2) return undefined\n  return {\n    n: xs.length,\n    pearson: pearsonR(xs, ys),\n    spearman: spearmanR(xs, ys),\n  }\n}\n","/**\n * ProjectRegistry — project-level aggregation over the trace corpus.\n *\n * Thin reader over TraceStore that answers the questions a chat-first,\n * resumable UI needs:\n *   - listProjects() → project IDs with latest activity\n *   - projectTimeline(id) → chats + builds + runtime runs, chronological\n *   - projectChats(id) → chat-level summaries (turn count, outcome)\n *\n * All queries are pure reads; no state duplication.\n */\n\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ProjectSummary {\n  projectId: string\n  chatCount: number\n  buildCount: number\n  appRuntimeCount: number\n  lastActivityAt: number\n  latestChatId?: string\n  latestOutcome?: { pass: boolean; score?: number }\n}\n\nexport interface ChatSummary {\n  chatId: string\n  projectId: string\n  builderRunId: string\n  startedAt: number\n  endedAt?: number\n  status: Run['status']\n  outcome?: Run['outcome']\n  /** Counts of spans emitted during the chat. */\n  llmTurns?: number\n  toolCalls?: number\n  buildRunId?: string\n  appRuntimeRunIds: string[]\n}\n\nexport interface ProjectTimelineEntry {\n  run: Run\n  layerBucket: 'chat' | 'build' | 'runtime' | 'other'\n}\n\nexport class ProjectRegistry {\n  constructor(private store: TraceStore) {}\n\n  async listProjects(): Promise<ProjectSummary[]> {\n    const runs = await this.store.listRuns()\n    const byProject = new Map<string, Run[]>()\n    for (const r of runs) {\n      if (!r.projectId) continue\n      const arr = byProject.get(r.projectId) ?? []\n      arr.push(r)\n      byProject.set(r.projectId, arr)\n    }\n    const summaries: ProjectSummary[] = []\n    for (const [projectId, projectRuns] of byProject) {\n      const sorted = projectRuns.slice().sort((a, b) => b.startedAt - a.startedAt)\n      const chats = projectRuns.filter((r) => r.layer === 'builder')\n      const builds = projectRuns.filter((r) => r.layer === 'app-build')\n      const runtimes = projectRuns.filter((r) => r.layer === 'app-runtime')\n      const latest = sorted[0]\n      if (!latest) continue\n      summaries.push({\n        projectId,\n        chatCount: chats.length,\n        buildCount: builds.length,\n        appRuntimeCount: runtimes.length,\n        lastActivityAt: latest.startedAt,\n        latestChatId: chats[0]?.chatId,\n        latestOutcome: latest.outcome\n          ? { pass: latest.outcome.pass ?? false, score: latest.outcome.score }\n          : undefined,\n      })\n    }\n    return summaries.sort((a, b) => b.lastActivityAt - a.lastActivityAt)\n  }\n\n  async projectTimeline(projectId: string): Promise<ProjectTimelineEntry[]> {\n    const runs = await this.store.listRuns({ projectId })\n    const ordered = runs.slice().sort((a, b) => a.startedAt - b.startedAt)\n    return ordered.map((run) => ({\n      run,\n      layerBucket:\n        run.layer === 'builder'\n          ? 'chat'\n          : run.layer === 'app-build'\n            ? 'build'\n            : run.layer === 'app-runtime'\n              ? 'runtime'\n              : 'other',\n    }))\n  }\n\n  async projectChats(projectId: string): Promise<ChatSummary[]> {\n    const builderRuns = (await this.store.listRuns({ projectId, layer: 'builder' })).sort(\n      (a, b) => b.startedAt - a.startedAt,\n    )\n    const childrenFor = async (runId: string) => this.store.listRuns({ parentRunId: runId })\n    const out: ChatSummary[] = []\n    for (const run of builderRuns) {\n      const spans = await this.store.spans({ runId: run.runId })\n      const children = await childrenFor(run.runId)\n      const build = children.find((c) => c.layer === 'app-build')\n      const runtime: string[] = []\n      // Runtime runs can be grandchildren (attached to the build) or siblings\n      // when shipped skipped.\n      if (build) {\n        const grands = await childrenFor(build.runId)\n        for (const g of grands) if (g.layer === 'app-runtime') runtime.push(g.runId)\n      }\n      for (const c of children) if (c.layer === 'app-runtime') runtime.push(c.runId)\n      out.push({\n        chatId: run.chatId ?? run.runId,\n        projectId,\n        builderRunId: run.runId,\n        startedAt: run.startedAt,\n        endedAt: run.endedAt,\n        status: run.status,\n        outcome: run.outcome,\n        llmTurns: spans.filter((s) => s.kind === 'llm').length,\n        toolCalls: spans.filter((s) => s.kind === 'tool').length,\n        buildRunId: build?.runId,\n        appRuntimeRunIds: runtime,\n      })\n    }\n    return out\n  }\n}\n","/**\n * Three-layer evaluation — the canonical scoring breakdown for\n * builder-of-builders workflows.\n *\n *   meta_score:    did the builder understand + satisfy user intent?\n *                  (judge verdict attached to the builder run)\n *   build_score:   did the generated scaffold build + pass its own tests?\n *                  (outcome.score on the app-build child run)\n *   runtime_score: did the generated agent pass its domain scenarios?\n *                  (mean outcome.score over app-runtime grandchild runs)\n *\n * Returns a structured report per project. The cross-layer correlation\n * is the highest-leverage signal the framework computes — if\n * meta_score doesn't predict runtime_score, the builder's self-scoring\n * is broken.\n *\n * Scaffold-only mode: when a project has no `app-runtime` runs (e.g. a\n * scaffold-builder eval that grades compose + build without driving a\n * runtime scenario), `kind` is `'scaffold-only'` and `complete` measures\n * meta + build only. Consumers can tell the two apart without having to\n * interpret null-runtime as either \"not yet computed\" or \"N/A for this\n * project shape\".\n */\n\nimport { judgeSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport type ProjectKind = 'full' | 'scaffold-only'\n\nexport interface ThreeLayerProjectReport {\n  projectId: string\n  /**\n   * `'full'` when the project has at least one `app-runtime` run;\n   * `'scaffold-only'` when it only has meta + build layers. Lets\n   * downstream consumers treat a null runtime score as expected\n   * (scaffold-only) vs. missing (full, pipeline broke).\n   */\n  kind: ProjectKind\n  builderRunId?: string\n  /** Judge-verdict score on the builder run (0..1 after normalization). */\n  metaScore: number | null\n  buildRunId?: string\n  /** 0..1 from the sandbox harness (testsPassed / testsTotal). */\n  buildScore: number | null\n  appRuntimeRunIds: string[]\n  /** Mean of outcome.score over app-runtime runs, 0..1. Always null in scaffold-only mode. */\n  runtimeScore: number | null\n  runtimePassRate: number | null\n  /**\n   * Layer-aware completeness:\n   *   - `kind='full'`: all three layers scored\n   *   - `kind='scaffold-only'`: meta + build scored (runtime not applicable)\n   */\n  complete: boolean\n}\n\nexport async function scoreProject(\n  store: TraceStore,\n  projectId: string,\n): Promise<ThreeLayerProjectReport> {\n  const allRuns = await store.listRuns({ projectId })\n  const builder = latestByLayer(allRuns, 'builder')\n  const build = latestByLayer(allRuns, 'app-build')\n  const runtime = allRuns.filter((r) => r.layer === 'app-runtime')\n\n  const metaScore = builder ? await extractMetaScore(store, builder.runId) : null\n  const buildScore = build?.outcome?.score ?? null\n  const runtimeScores = runtime\n    .map((r) => r.outcome?.score)\n    .filter((s): s is number => typeof s === 'number')\n  const runtimeScore =\n    runtimeScores.length > 0\n      ? runtimeScores.reduce((a, b) => a + b, 0) / runtimeScores.length\n      : null\n  const runtimePassed = runtime.filter((r) => r.outcome?.pass === true).length\n  const runtimePassRate = runtime.length > 0 ? runtimePassed / runtime.length : null\n\n  const kind: ProjectKind = runtime.length === 0 ? 'scaffold-only' : 'full'\n  const complete =\n    kind === 'scaffold-only'\n      ? metaScore !== null && buildScore !== null\n      : metaScore !== null && buildScore !== null && runtimeScore !== null\n\n  return {\n    projectId,\n    kind,\n    builderRunId: builder?.runId,\n    metaScore,\n    buildRunId: build?.runId,\n    buildScore,\n    appRuntimeRunIds: runtime.map((r) => r.runId),\n    runtimeScore,\n    runtimePassRate,\n    complete,\n  }\n}\n\n/** Aggregate scoring across every project in a corpus. */\nexport async function scoreAllProjects(store: TraceStore): Promise<ThreeLayerProjectReport[]> {\n  const runs = await store.listRuns()\n  const projectIds = [...new Set(runs.map((r) => r.projectId).filter((p): p is string => !!p))]\n  return Promise.all(projectIds.map((p) => scoreProject(store, p)))\n}\n\nfunction latestByLayer(runs: Run[], layer: Run['layer']): Run | undefined {\n  const filtered = runs.filter((r) => r.layer === layer).sort((a, b) => b.startedAt - a.startedAt)\n  return filtered[0]\n}\n\nasync function extractMetaScore(store: TraceStore, builderRunId: string): Promise<number | null> {\n  const js = await judgeSpans(store, builderRunId)\n  const meta = js.find(\n    (s) => s.judgeId === 'builder-meta' && s.dimension === 'user_intent_satisfaction',\n  )\n  if (!meta) return null\n  // Normalize score to 0..1. Accept 0-1 natively; 0-10 scale is also common.\n  if (meta.score >= 0 && meta.score <= 1) return meta.score\n  if (meta.score >= 0 && meta.score <= 10) return meta.score / 10\n  return null\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AA8CO,IAAM,iBAAN,MAAqB;AAAA,EAClB;AAAA,EACA;AAAA,EACC;AAAA,EACA;AAAA,EACD;AAAA,EACA;AAAA,EACA;AAAA,EAER,YAAY,OAAmB,MAA0B,QAAwB;AAC/E,SAAK,QAAQ;AACb,SAAK,YAAY,KAAK;AACtB,SAAK,SAAS,KAAK,UAAU,SAAS;AACtC,SAAK,gBAAgB;AACrB,SAAK,iBAAiB,IAAI,aAAa,KAAK;AAAA,EAC9C;AAAA;AAAA,EAGA,MAAM,UAAU,aAAa,GAAG,KAAK,SAAS,SAA0B;AACtE,UAAM,KAAK,eAAe,SAAS;AAAA,MACjC;AAAA,MACA,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,OAAO;AAAA,IACT,CAAC;AACD,SAAK,eAAe,KAAK,eAAe;AACxC,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,UAAwB;AAC1B,QAAI,CAAC,KAAK,aAAc,OAAM,IAAI,MAAM,gDAAgD;AACxF,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,KAAK,SAAgF;AACzF,QAAI,CAAC,KAAK,aAAc,OAAM,IAAI,MAAM,6CAA6C;AACrF,UAAM,eAAe,IAAI,aAAa,KAAK,KAAK;AAChD,UAAM,aAAa,SAAS;AAAA,MAC1B,YAAY,QAAQ,cAAc,GAAG,KAAK,SAAS;AAAA,MACnD,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,aAAa,KAAK;AAAA,MAClB,OAAO;AAAA,IACT,CAAC;AACD,UAAM,UAAU,IAAI,eAAe,QAAQ,UAAU,KAAK,aAAa;AACvE,UAAM,SAAS,MAAM,QAAQ,IAAI,QAAQ,SAAS,YAAY;AAC9D,UAAM,aAAa,OAAO;AAAA,MACxB,MAAM,OAAO;AAAA,MACb,OAAO,OAAO;AAAA,MACd,cAAc,OAAO,SAAS,YAAY;AAAA,IAC5C,CAAC;AACD,SAAK,iBAAiB,aAAa;AACnC,WAAO,EAAE,OAAO,aAAa,OAAO,OAAO;AAAA,EAC7C;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,eAAe,SAA8D;AACjF,UAAM,cAAc,KAAK,kBAAkB,KAAK;AAChD,QAAI,CAAC;AACH,YAAM,IAAI,MAAM,gEAAgE;AAClF,UAAM,EAAE,UAAU,OAAO,IAAI;AAC7B,UAAM,SAAS,MAAM,sBAAsB,UAAU,KAAK,OAAO;AAAA,MAC/D,QAAQ,UAAU,KAAK;AAAA,MACvB,YAAY,EAAE,SAAS,QAAW,WAAW,QAAW,kBAAkB,OAAU;AAAA,IACtF,CAAC;AAED,UAAM,KAAK,MAAM,UAAU,OAAO,OAAO;AAAA,MACvC;AAAA,MACA,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,OAAO;AAAA,IACT,CAAC;AACD,WAAO;AAAA,EACT;AAAA;AAAA;AAAA,EAIA,MAAM,gBAAgB,OAAe,WAAmC;AACtE,QAAI,CAAC,KAAK;AACR,YAAM,IAAI,MAAM,wDAAwD;AAC1E,UAAM,KAAK,eAAe,YAAY;AAAA,MACpC,SAAS;AAAA,MACT,cAAc,KAAK;AAAA;AAAA,MACnB,WAAW;AAAA,MACX;AAAA,MACA;AAAA,MACA,MAAM;AAAA,IACR,CAAC;AAAA,EACH;AAAA;AAAA,EAGA,MAAM,QAAQ,SAA2E;AACvF,UAAM,KAAK,eAAe,OAAO;AAAA,MAC/B,MAAM,QAAQ;AAAA,MACd,OAAO,QAAQ;AAAA,MACf,OAAO,QAAQ;AAAA,IACjB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,gBAAgB,YAA2C;AAC/D,UAAM,cAAc,KAAK,kBAAkB,KAAK;AAChD,QAAI,CAAC;AACH,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AACF,UAAM,UAAU,IAAI,aAAa,KAAK,KAAK;AAC3C,UAAM,QAAQ,SAAS;AAAA,MACrB;AAAA,MACA,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA,OAAO;AAAA,IACT,CAAC;AACD,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,iBAAiB,MAKH;AAClB,QAAI,CAAC,KAAK;AACR,YAAM,IAAI,MAAM,yDAAyD;AAC3E,UAAM,UAAU,IAAI,aAAa,KAAK,KAAK;AAC3C,UAAM,QAAQ,SAAS;AAAA,MACrB,YAAY,KAAK,cAAc,GAAG,KAAK,SAAS;AAAA,MAChD,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,aAAa,KAAK;AAAA,MAClB,OAAO;AAAA,IACT,CAAC;AACD,UAAM,QAAQ,OAAO;AAAA,MACnB,MAAM,KAAK;AAAA,MACX,OAAO,KAAK;AAAA,MACZ,cAAc,KAAK,OAAO,YAAY;AAAA,MACtC,OAAO,KAAK;AAAA,IACd,CAAC;AACD,SAAK,iBAAiB,QAAQ;AAC9B,WAAO,QAAQ;AAAA,EACjB;AAAA,EAEA,IAAI,sBAA0C;AAC5C,WAAO,KAAK;AAAA,EACd;AAAA,EACA,IAAI,oBAAwC;AAC1C,WAAO,KAAK;AAAA,EACd;AACF;AAOA,eAAsB,qBACpB,OACA,WAOC;AACD,QAAM,OAAO,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;AAC/C,QAAM,WAAW,KACd,OAAO,CAAC,MAAM,EAAE,UAAU,SAAS,EACnC,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3C,QAAM,YAAY,KACf,OAAO,CAAC,MAAM,EAAE,UAAU,WAAW,EACrC,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3C,QAAM,iBAAiB,KACpB,OAAO,CAAC,MAAM,EAAE,UAAU,aAAa,EACvC,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3C,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,gBAAgB,SAAS,CAAC;AAAA,IAC1B,cAAc,UAAU,CAAC;AAAA,IACzB,oBAAoB;AAAA,EACtB;AACF;AAEA,SAAS,WAAmB;AAC1B,MAAI,OAAO,WAAW,QAAQ,eAAe,WAAY,QAAO,WAAW,OAAO,WAAW;AAC7F,SAAO,GAAG,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,EAAE,CAAC;AAC9E;;;AC5NO,SAAS,gBAAgB,SAAuD;AACrF,QAAM,mBAAmB,QAAQ,OAAO,CAAC,MAAM,EAAE,QAAQ,EAAE;AAC3D,SAAO;AAAA,IACL,aAAa;AAAA,MACX;AAAA,MACA,CAAC,MAAM,EAAE;AAAA,MACT,CAAC,MAAM,EAAE;AAAA,IACX;AAAA,IACA,eAAe;AAAA,MACb;AAAA,MACA,CAAC,MAAM,EAAE;AAAA,MACT,CAAC,MAAM,EAAE;AAAA,IACX;AAAA,IACA,gBAAgB;AAAA,MACd;AAAA,MACA,CAAC,MAAM,EAAE;AAAA,MACT,CAAC,MAAM,EAAE;AAAA,IACX;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,SACP,SACA,GACA,GAC8B;AAC9B,QAAM,KAAe,CAAC;AACtB,QAAM,KAAe,CAAC;AACtB,aAAW,KAAK,SAAS;AACvB,UAAM,IAAI,EAAE,CAAC;AACb,UAAM,IAAI,EAAE,CAAC;AACb,QAAI,MAAM,QAAQ,MAAM,QAAQ,OAAO,SAAS,CAAC,KAAK,OAAO,SAAS,CAAC,GAAG;AACxE,SAAG,KAAK,CAAC;AACT,SAAG,KAAK,CAAC;AAAA,IACX;AAAA,EACF;AACA,MAAI,GAAG,SAAS,EAAG,QAAO;AAC1B,SAAO;AAAA,IACL,GAAG,GAAG;AAAA,IACN,SAAS,SAAS,IAAI,EAAE;AAAA,IACxB,UAAU,UAAU,IAAI,EAAE;AAAA,EAC5B;AACF;;;ACjCO,IAAM,kBAAN,MAAsB;AAAA,EAC3B,YAAoB,OAAmB;AAAnB;AAAA,EAAoB;AAAA,EAApB;AAAA,EAEpB,MAAM,eAA0C;AAC9C,UAAM,OAAO,MAAM,KAAK,MAAM,SAAS;AACvC,UAAM,YAAY,oBAAI,IAAmB;AACzC,eAAW,KAAK,MAAM;AACpB,UAAI,CAAC,EAAE,UAAW;AAClB,YAAM,MAAM,UAAU,IAAI,EAAE,SAAS,KAAK,CAAC;AAC3C,UAAI,KAAK,CAAC;AACV,gBAAU,IAAI,EAAE,WAAW,GAAG;AAAA,IAChC;AACA,UAAM,YAA8B,CAAC;AACrC,eAAW,CAAC,WAAW,WAAW,KAAK,WAAW;AAChD,YAAM,SAAS,YAAY,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3E,YAAM,QAAQ,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,SAAS;AAC7D,YAAM,SAAS,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,WAAW;AAChE,YAAM,WAAW,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,aAAa;AACpE,YAAM,SAAS,OAAO,CAAC;AACvB,UAAI,CAAC,OAAQ;AACb,gBAAU,KAAK;AAAA,QACb;AAAA,QACA,WAAW,MAAM;AAAA,QACjB,YAAY,OAAO;AAAA,QACnB,iBAAiB,SAAS;AAAA,QAC1B,gBAAgB,OAAO;AAAA,QACvB,cAAc,MAAM,CAAC,GAAG;AAAA,QACxB,eAAe,OAAO,UAClB,EAAE,MAAM,OAAO,QAAQ,QAAQ,OAAO,OAAO,OAAO,QAAQ,MAAM,IAClE;AAAA,MACN,CAAC;AAAA,IACH;AACA,WAAO,UAAU,KAAK,CAAC,GAAG,MAAM,EAAE,iBAAiB,EAAE,cAAc;AAAA,EACrE;AAAA,EAEA,MAAM,gBAAgB,WAAoD;AACxE,UAAM,OAAO,MAAM,KAAK,MAAM,SAAS,EAAE,UAAU,CAAC;AACpD,UAAM,UAAU,KAAK,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACrE,WAAO,QAAQ,IAAI,CAAC,SAAS;AAAA,MAC3B;AAAA,MACA,aACE,IAAI,UAAU,YACV,SACA,IAAI,UAAU,cACZ,UACA,IAAI,UAAU,gBACZ,YACA;AAAA,IACZ,EAAE;AAAA,EACJ;AAAA,EAEA,MAAM,aAAa,WAA2C;AAC5D,UAAM,eAAe,MAAM,KAAK,MAAM,SAAS,EAAE,WAAW,OAAO,UAAU,CAAC,GAAG;AAAA,MAC/E,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE;AAAA,IAC5B;AACA,UAAM,cAAc,OAAO,UAAkB,KAAK,MAAM,SAAS,EAAE,aAAa,MAAM,CAAC;AACvF,UAAM,MAAqB,CAAC;AAC5B,eAAW,OAAO,aAAa;AAC7B,YAAM,QAAQ,MAAM,KAAK,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;AACzD,YAAM,WAAW,MAAM,YAAY,IAAI,KAAK;AAC5C,YAAM,QAAQ,SAAS,KAAK,CAAC,MAAM,EAAE,UAAU,WAAW;AAC1D,YAAM,UAAoB,CAAC;AAG3B,UAAI,OAAO;AACT,cAAM,SAAS,MAAM,YAAY,MAAM,KAAK;AAC5C,mBAAW,KAAK,OAAQ,KAAI,EAAE,UAAU,cAAe,SAAQ,KAAK,EAAE,KAAK;AAAA,MAC7E;AACA,iBAAW,KAAK,SAAU,KAAI,EAAE,UAAU,cAAe,SAAQ,KAAK,EAAE,KAAK;AAC7E,UAAI,KAAK;AAAA,QACP,QAAQ,IAAI,UAAU,IAAI;AAAA,QAC1B;AAAA,QACA,cAAc,IAAI;AAAA,QAClB,WAAW,IAAI;AAAA,QACf,SAAS,IAAI;AAAA,QACb,QAAQ,IAAI;AAAA,QACZ,SAAS,IAAI;AAAA,QACb,UAAU,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,EAAE;AAAA,QAChD,WAAW,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE;AAAA,QAClD,YAAY,OAAO;AAAA,QACnB,kBAAkB;AAAA,MACpB,CAAC;AAAA,IACH;AACA,WAAO;AAAA,EACT;AACF;;;ACzEA,eAAsB,aACpB,OACA,WACkC;AAClC,QAAM,UAAU,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;AAClD,QAAM,UAAU,cAAc,SAAS,SAAS;AAChD,QAAM,QAAQ,cAAc,SAAS,WAAW;AAChD,QAAM,UAAU,QAAQ,OAAO,CAAC,MAAM,EAAE,UAAU,aAAa;AAE/D,QAAM,YAAY,UAAU,MAAM,iBAAiB,OAAO,QAAQ,KAAK,IAAI;AAC3E,QAAM,aAAa,OAAO,SAAS,SAAS;AAC5C,QAAM,gBAAgB,QACnB,IAAI,CAAC,MAAM,EAAE,SAAS,KAAK,EAC3B,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACnD,QAAM,eACJ,cAAc,SAAS,IACnB,cAAc,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,cAAc,SACzD;AACN,QAAM,gBAAgB,QAAQ,OAAO,CAAC,MAAM,EAAE,SAAS,SAAS,IAAI,EAAE;AACtE,QAAM,kBAAkB,QAAQ,SAAS,IAAI,gBAAgB,QAAQ,SAAS;AAE9E,QAAM,OAAoB,QAAQ,WAAW,IAAI,kBAAkB;AACnE,QAAM,WACJ,SAAS,kBACL,cAAc,QAAQ,eAAe,OACrC,cAAc,QAAQ,eAAe,QAAQ,iBAAiB;AAEpE,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,cAAc,SAAS;AAAA,IACvB;AAAA,IACA,YAAY,OAAO;AAAA,IACnB;AAAA,IACA,kBAAkB,QAAQ,IAAI,CAAC,MAAM,EAAE,KAAK;AAAA,IAC5C;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAGA,eAAsB,iBAAiB,OAAuD;AAC5F,QAAM,OAAO,MAAM,MAAM,SAAS;AAClC,QAAM,aAAa,CAAC,GAAG,IAAI,IAAI,KAAK,IAAI,CAAC,MAAM,EAAE,SAAS,EAAE,OAAO,CAAC,MAAmB,CAAC,CAAC,CAAC,CAAC,CAAC;AAC5F,SAAO,QAAQ,IAAI,WAAW,IAAI,CAAC,MAAM,aAAa,OAAO,CAAC,CAAC,CAAC;AAClE;AAEA,SAAS,cAAc,MAAa,OAAsC;AACxE,QAAM,WAAW,KAAK,OAAO,CAAC,MAAM,EAAE,UAAU,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC/F,SAAO,SAAS,CAAC;AACnB;AAEA,eAAe,iBAAiB,OAAmB,cAA8C;AAC/F,QAAM,KAAK,MAAM,WAAW,OAAO,YAAY;AAC/C,QAAM,OAAO,GAAG;AAAA,IACd,CAAC,MAAM,EAAE,YAAY,kBAAkB,EAAE,cAAc;AAAA,EACzD;AACA,MAAI,CAAC,KAAM,QAAO;AAElB,MAAI,KAAK,SAAS,KAAK,KAAK,SAAS,EAAG,QAAO,KAAK;AACpD,MAAI,KAAK,SAAS,KAAK,KAAK,SAAS,GAAI,QAAO,KAAK,QAAQ;AAC7D,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../../src/builder-eval/builder-session.ts","../../src/builder-eval/correlation.ts","../../src/builder-eval/project-registry.ts","../../src/builder-eval/three-layer-eval.ts"],"sourcesContent":["/**\n * BuilderSession — ties a builder-of-builders workflow together.\n *\n * Models agent-builder's shape: Project → Chat → Edit → Ship → App →\n * AppAgent. Each layer is a Run (linked via parentRunId). The\n * framework-enforced invariants:\n *\n *   - One Project → many Chats; chatId scopes runs within a project.\n *   - One Chat = one builder Run with `layer='builder'`.\n *   - One Ship = one child Run with `layer='app-build'` + SandboxHarness.\n *   - One AppScenario = one grandchild Run with `layer='app-runtime'`.\n *\n * Consumers obtain a BuilderSession, call `startChat`, drive the\n * builder agent (emitting spans), and call `ship` / `runAppScenario`\n * as the workflow progresses. The session reconstructs itself from\n * trace data via `resume(store, projectId)`.\n */\n\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from '../sandbox-harness'\nimport { SandboxHarness } from '../sandbox-harness'\nimport type { TestGradedRunResult, TestGradedScenario } from '../test-graded-scenario'\nimport { runTestGradedScenario } from '../test-graded-scenario'\nimport { TraceEmitter } from '../trace/emitter'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface BuilderSessionInit {\n  projectId: string\n  chatId?: string\n  /** Free-form: user's task description, project name, etc. Stored on the builder Run. */\n  tags?: Record<string, string>\n}\n\nexport interface ShipOptions {\n  harness: HarnessConfig\n  driver?: SandboxDriver\n  /** scenarioId of this app-build run. Defaults to `${projectId}/build`. */\n  scenarioId?: string\n}\n\nexport interface RunAppScenarioOptions {\n  scenario: TestGradedScenario\n  /** Harness driver override; defaults to the one the session was created with. */\n  driver?: SandboxDriver\n}\n\nexport class BuilderSession {\n  private store: TraceStore\n  private builderEmitter: TraceEmitter\n  readonly projectId: string\n  readonly chatId: string\n  private builderRunId?: string\n  private lastBuildRunId?: string\n  private defaultDriver?: SandboxDriver\n\n  constructor(store: TraceStore, init: BuilderSessionInit, driver?: SandboxDriver) {\n    this.store = store\n    this.projectId = init.projectId\n    this.chatId = init.chatId ?? cryptoId()\n    this.defaultDriver = driver\n    this.builderEmitter = new TraceEmitter(store)\n  }\n\n  /** Start the builder (L0) run for this chat. Returns the runId. */\n  async startChat(scenarioId = `${this.projectId}/chat`): Promise<string> {\n    await this.builderEmitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'builder',\n    })\n    this.builderRunId = this.builderEmitter.runId\n    return this.builderRunId\n  }\n\n  /** The emitter for builder-level spans (edits, LLM calls, tool invocations). */\n  get emitter(): TraceEmitter {\n    if (!this.builderRunId) throw new Error('BuilderSession.emitter: call startChat() first')\n    return this.builderEmitter\n  }\n\n  /**\n   * Ship the project's generated app: run the sandbox harness as a child\n   * Run (`layer='app-build'`). Returns the build result + runId.\n   */\n  async ship(options: ShipOptions): Promise<{ runId: string; result: SandboxHarnessResult }> {\n    if (!this.builderRunId) throw new Error('BuilderSession.ship: call startChat() first')\n    const buildEmitter = new TraceEmitter(this.store)\n    await buildEmitter.startRun({\n      scenarioId: options.scenarioId ?? `${this.projectId}/build`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    const harness = new SandboxHarness(options.driver ?? this.defaultDriver)\n    const result = await harness.run(options.harness, buildEmitter)\n    await buildEmitter.endRun({\n      pass: result.passed,\n      score: result.score,\n      failureClass: result.passed ? 'success' : 'sandbox_failure',\n    })\n    this.lastBuildRunId = buildEmitter.runId\n    return { runId: buildEmitter.runId, result }\n  }\n\n  /**\n   * Run a domain scenario against the just-built app as a grandchild Run\n   * (`layer='app-runtime'`). The `ship` call must precede this so the\n   * parent is set correctly; if no build exists yet the session attaches\n   * directly to the builder run (useful for prototypes).\n   */\n  async runAppScenario(options: RunAppScenarioOptions): Promise<TestGradedRunResult> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error('BuilderSession.runAppScenario: call startChat() + ship() first')\n    const { scenario, driver } = options\n    const result = await runTestGradedScenario(scenario, this.store, {\n      driver: driver ?? this.defaultDriver,\n      provenance: { codeSha: undefined, promptSha: undefined, modelFingerprint: undefined },\n    })\n    // Attach to the parent chain by updating the stored Run in place.\n    await this.store.updateRun(result.runId, {\n      parentRunId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      layer: 'app-runtime',\n    })\n    return result\n  }\n\n  /** Record an end-of-chat meta score (judge verdict on whether the builder\n   *  served the user's intent). Accepts a numeric score + optional rationale. */\n  async recordMetaScore(score: number, rationale?: string): Promise<void> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordMetaScore: call startChat() first')\n    await this.builderEmitter.recordJudge({\n      judgeId: 'builder-meta',\n      targetSpanId: this.builderRunId, // attach to the builder run itself\n      dimension: 'user_intent_satisfaction',\n      score,\n      rationale,\n      name: 'builder-meta',\n    })\n  }\n\n  /** Close the builder Run with a final outcome. */\n  async endChat(outcome: { pass: boolean; score?: number; notes?: string }): Promise<void> {\n    await this.builderEmitter.endRun({\n      pass: outcome.pass,\n      score: outcome.score,\n      notes: outcome.notes,\n    })\n  }\n\n  /**\n   * Inline app-runtime run — for cases where the \"scenario\" isn't a\n   * SWE-bench-style test suite but a live agent interaction (LLM chat,\n   * domain flow). Returns an emitter bound to a fresh Run in the\n   * `app-runtime` layer; caller emits spans inside and calls\n   * `.endRun()` with the final verdict.\n   */\n  async startAppRuntime(scenarioId: string): Promise<TraceEmitter> {\n    const parentRunId = this.lastBuildRunId ?? this.builderRunId\n    if (!parentRunId)\n      throw new Error(\n        'BuilderSession.startAppRuntime: call startChat() + (optionally) ship() first',\n      )\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId,\n      layer: 'app-runtime',\n    })\n    return emitter\n  }\n\n  /**\n   * Lightweight \"ship marker\" — record an app-build Run with a caller-\n   * provided verdict. Use when there isn't a sandbox harness to run but\n   * you still want to mark the build state at publish time.\n   */\n  async recordShipMarker(args: {\n    pass: boolean\n    score: number\n    scenarioId?: string\n    notes?: string\n  }): Promise<string> {\n    if (!this.builderRunId)\n      throw new Error('BuilderSession.recordShipMarker: call startChat() first')\n    const emitter = new TraceEmitter(this.store)\n    await emitter.startRun({\n      scenarioId: args.scenarioId ?? `${this.projectId}/ship`,\n      projectId: this.projectId,\n      chatId: this.chatId,\n      parentRunId: this.builderRunId,\n      layer: 'app-build',\n    })\n    await emitter.endRun({\n      pass: args.pass,\n      score: args.score,\n      failureClass: args.pass ? 'success' : 'sandbox_failure',\n      notes: args.notes,\n    })\n    this.lastBuildRunId = emitter.runId\n    return emitter.runId\n  }\n\n  get lastBuildRunIdValue(): string | undefined {\n    return this.lastBuildRunId\n  }\n  get builderRunIdValue(): string | undefined {\n    return this.builderRunId\n  }\n}\n\n/**\n * Reconstruct the most recent BuilderSession state for a given project —\n * returns { builderRunId, lastBuildRunId, chatRuns }. For chat-first UIs\n * this is how a resumed session finds its place in the edit history.\n */\nexport async function resumeBuilderSession(\n  store: TraceStore,\n  projectId: string,\n): Promise<{\n  projectId: string\n  chatRuns: Run[]\n  lastBuilderRun?: Run\n  lastBuildRun?: Run\n  lastAppRuntimeRuns: Run[]\n}> {\n  const runs = await store.listRuns({ projectId })\n  const chatRuns = runs\n    .filter((r) => r.layer === 'builder')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const buildRuns = runs\n    .filter((r) => r.layer === 'app-build')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  const appRuntimeRuns = runs\n    .filter((r) => r.layer === 'app-runtime')\n    .sort((a, b) => b.startedAt - a.startedAt)\n  return {\n    projectId,\n    chatRuns,\n    lastBuilderRun: chatRuns[0],\n    lastBuildRun: buildRuns[0],\n    lastAppRuntimeRuns: appRuntimeRuns,\n  }\n}\n\nfunction cryptoId(): string {\n  if (typeof globalThis.crypto?.randomUUID === 'function') return globalThis.crypto.randomUUID()\n  return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`\n}\n","/**\n * Meta-eval correlation — the highest-leverage signal in the framework.\n *\n * Given a corpus of three-layer project reports, compute how well each\n * pair of layers correlates. The question we care about most:\n *\n *   Does `metaScore` (what the builder thinks it did) predict\n *   `runtimeScore` (what the user actually gets)?\n *\n * If r < ~0.4, the builder's self-scoring is broken — it's optimizing\n * for something other than real-world success. If r > 0.7, meta_score\n * is a usable proxy and can drive CI gates cheaply.\n *\n * Non-parametric rank correlation (Spearman) is also reported because\n * meta scores are often ordinal-ish.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport type { ThreeLayerProjectReport } from './three-layer-eval'\n\nexport interface LayerCorrelation {\n  n: number\n  pearson: number\n  spearman: number\n}\n\nexport interface CorrelationReport {\n  /** Pairs present in the corpus (layers with ≥ 2 matched data points). */\n  metaVsBuild?: LayerCorrelation\n  metaVsRuntime?: LayerCorrelation\n  buildVsRuntime?: LayerCorrelation\n  /** Number of complete projects (all 3 scores present). */\n  completeProjects: number\n}\n\nexport function correlateLayers(reports: ThreeLayerProjectReport[]): CorrelationReport {\n  const completeProjects = reports.filter((r) => r.complete).length\n  return {\n    metaVsBuild: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.buildScore,\n    ),\n    metaVsRuntime: pairwise(\n      reports,\n      (r) => r.metaScore,\n      (r) => r.runtimeScore,\n    ),\n    buildVsRuntime: pairwise(\n      reports,\n      (r) => r.buildScore,\n      (r) => r.runtimeScore,\n    ),\n    completeProjects,\n  }\n}\n\nfunction pairwise(\n  reports: ThreeLayerProjectReport[],\n  a: (r: ThreeLayerProjectReport) => number | null,\n  b: (r: ThreeLayerProjectReport) => number | null,\n): LayerCorrelation | undefined {\n  const xs: number[] = []\n  const ys: number[] = []\n  for (const r of reports) {\n    const x = a(r)\n    const y = b(r)\n    if (x !== null && y !== null && Number.isFinite(x) && Number.isFinite(y)) {\n      xs.push(x)\n      ys.push(y)\n    }\n  }\n  if (xs.length < 2) return undefined\n  return {\n    n: xs.length,\n    pearson: pearsonR(xs, ys),\n    spearman: spearmanR(xs, ys),\n  }\n}\n","/**\n * ProjectRegistry — project-level aggregation over the trace corpus.\n *\n * Thin reader over TraceStore that answers the questions a chat-first,\n * resumable UI needs:\n *   - listProjects() → project IDs with latest activity\n *   - projectTimeline(id) → chats + builds + runtime runs, chronological\n *   - projectChats(id) → chat-level summaries (turn count, outcome)\n *\n * All queries are pure reads; no state duplication.\n */\n\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport interface ProjectSummary {\n  projectId: string\n  chatCount: number\n  buildCount: number\n  appRuntimeCount: number\n  lastActivityAt: number\n  latestChatId?: string\n  latestOutcome?: { pass: boolean; score?: number }\n}\n\nexport interface ChatSummary {\n  chatId: string\n  projectId: string\n  builderRunId: string\n  startedAt: number\n  endedAt?: number\n  status: Run['status']\n  outcome?: Run['outcome']\n  /** Counts of spans emitted during the chat. */\n  llmTurns?: number\n  toolCalls?: number\n  buildRunId?: string\n  appRuntimeRunIds: string[]\n}\n\nexport interface ProjectTimelineEntry {\n  run: Run\n  layerBucket: 'chat' | 'build' | 'runtime' | 'other'\n}\n\nexport class ProjectRegistry {\n  constructor(private store: TraceStore) {}\n\n  async listProjects(): Promise<ProjectSummary[]> {\n    const runs = await this.store.listRuns()\n    const byProject = new Map<string, Run[]>()\n    for (const r of runs) {\n      if (!r.projectId) continue\n      const arr = byProject.get(r.projectId) ?? []\n      arr.push(r)\n      byProject.set(r.projectId, arr)\n    }\n    const summaries: ProjectSummary[] = []\n    for (const [projectId, projectRuns] of byProject) {\n      const sorted = projectRuns.slice().sort((a, b) => b.startedAt - a.startedAt)\n      const chats = projectRuns.filter((r) => r.layer === 'builder')\n      const builds = projectRuns.filter((r) => r.layer === 'app-build')\n      const runtimes = projectRuns.filter((r) => r.layer === 'app-runtime')\n      const latest = sorted[0]\n      if (!latest) continue\n      summaries.push({\n        projectId,\n        chatCount: chats.length,\n        buildCount: builds.length,\n        appRuntimeCount: runtimes.length,\n        lastActivityAt: latest.startedAt,\n        latestChatId: chats[0]?.chatId,\n        latestOutcome: latest.outcome\n          ? { pass: latest.outcome.pass ?? false, score: latest.outcome.score }\n          : undefined,\n      })\n    }\n    return summaries.sort((a, b) => b.lastActivityAt - a.lastActivityAt)\n  }\n\n  async projectTimeline(projectId: string): Promise<ProjectTimelineEntry[]> {\n    const runs = await this.store.listRuns({ projectId })\n    const ordered = runs.slice().sort((a, b) => a.startedAt - b.startedAt)\n    return ordered.map((run) => ({\n      run,\n      layerBucket:\n        run.layer === 'builder'\n          ? 'chat'\n          : run.layer === 'app-build'\n            ? 'build'\n            : run.layer === 'app-runtime'\n              ? 'runtime'\n              : 'other',\n    }))\n  }\n\n  async projectChats(projectId: string): Promise<ChatSummary[]> {\n    const builderRuns = (await this.store.listRuns({ projectId, layer: 'builder' })).sort(\n      (a, b) => b.startedAt - a.startedAt,\n    )\n    const childrenFor = async (runId: string) => this.store.listRuns({ parentRunId: runId })\n    const out: ChatSummary[] = []\n    for (const run of builderRuns) {\n      const spans = await this.store.spans({ runId: run.runId })\n      const children = await childrenFor(run.runId)\n      const build = children.find((c) => c.layer === 'app-build')\n      const runtime: string[] = []\n      // Runtime runs can be grandchildren (attached to the build) or siblings\n      // when shipped skipped.\n      if (build) {\n        const grands = await childrenFor(build.runId)\n        for (const g of grands) if (g.layer === 'app-runtime') runtime.push(g.runId)\n      }\n      for (const c of children) if (c.layer === 'app-runtime') runtime.push(c.runId)\n      out.push({\n        chatId: run.chatId ?? run.runId,\n        projectId,\n        builderRunId: run.runId,\n        startedAt: run.startedAt,\n        endedAt: run.endedAt,\n        status: run.status,\n        outcome: run.outcome,\n        llmTurns: spans.filter((s) => s.kind === 'llm').length,\n        toolCalls: spans.filter((s) => s.kind === 'tool').length,\n        buildRunId: build?.runId,\n        appRuntimeRunIds: runtime,\n      })\n    }\n    return out\n  }\n}\n","/**\n * Three-layer evaluation — the canonical scoring breakdown for\n * builder-of-builders workflows.\n *\n *   meta_score:    did the builder understand + satisfy user intent?\n *                  (judge verdict attached to the builder run)\n *   build_score:   did the generated scaffold build + pass its own tests?\n *                  (outcome.score on the app-build child run)\n *   runtime_score: did the generated agent pass its domain scenarios?\n *                  (mean outcome.score over app-runtime grandchild runs)\n *\n * Returns a structured report per project. The cross-layer correlation\n * is the highest-leverage signal the framework computes — if\n * meta_score doesn't predict runtime_score, the builder's self-scoring\n * is broken.\n *\n * Scaffold-only mode: when a project has no `app-runtime` runs (e.g. a\n * scaffold-builder eval that grades compose + build without driving a\n * runtime scenario), `kind` is `'scaffold-only'` and `complete` measures\n * meta + build only. Consumers can tell the two apart without having to\n * interpret null-runtime as either \"not yet computed\" or \"N/A for this\n * project shape\".\n */\n\nimport { judgeSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\n\nexport type ProjectKind = 'full' | 'scaffold-only'\n\nexport interface ThreeLayerProjectReport {\n  projectId: string\n  /**\n   * `'full'` when the project has at least one `app-runtime` run;\n   * `'scaffold-only'` when it only has meta + build layers. Lets\n   * downstream consumers treat a null runtime score as expected\n   * (scaffold-only) vs. missing (full, pipeline broke).\n   */\n  kind: ProjectKind\n  builderRunId?: string\n  /** Judge-verdict score on the builder run (0..1 after normalization). */\n  metaScore: number | null\n  buildRunId?: string\n  /** 0..1 from the sandbox harness (testsPassed / testsTotal). */\n  buildScore: number | null\n  appRuntimeRunIds: string[]\n  /** Mean of outcome.score over app-runtime runs, 0..1. Always null in scaffold-only mode. */\n  runtimeScore: number | null\n  runtimePassRate: number | null\n  /**\n   * Layer-aware completeness:\n   *   - `kind='full'`: all three layers scored\n   *   - `kind='scaffold-only'`: meta + build scored (runtime not applicable)\n   */\n  complete: boolean\n}\n\nexport async function scoreProject(\n  store: TraceStore,\n  projectId: string,\n): Promise<ThreeLayerProjectReport> {\n  const allRuns = await store.listRuns({ projectId })\n  const builder = latestByLayer(allRuns, 'builder')\n  const build = latestByLayer(allRuns, 'app-build')\n  const runtime = allRuns.filter((r) => r.layer === 'app-runtime')\n\n  const metaScore = builder ? await extractMetaScore(store, builder.runId) : null\n  const buildScore = build?.outcome?.score ?? null\n  const runtimeScores = runtime\n    .map((r) => r.outcome?.score)\n    .filter((s): s is number => typeof s === 'number')\n  const runtimeScore =\n    runtimeScores.length > 0\n      ? runtimeScores.reduce((a, b) => a + b, 0) / runtimeScores.length\n      : null\n  const runtimePassed = runtime.filter((r) => r.outcome?.pass === true).length\n  const runtimePassRate = runtime.length > 0 ? runtimePassed / runtime.length : null\n\n  const kind: ProjectKind = runtime.length === 0 ? 'scaffold-only' : 'full'\n  const complete =\n    kind === 'scaffold-only'\n      ? metaScore !== null && buildScore !== null\n      : metaScore !== null && buildScore !== null && runtimeScore !== null\n\n  return {\n    projectId,\n    kind,\n    builderRunId: builder?.runId,\n    metaScore,\n    buildRunId: build?.runId,\n    buildScore,\n    appRuntimeRunIds: runtime.map((r) => r.runId),\n    runtimeScore,\n    runtimePassRate,\n    complete,\n  }\n}\n\n/** Aggregate scoring across every project in a corpus. */\nexport async function scoreAllProjects(store: TraceStore): Promise<ThreeLayerProjectReport[]> {\n  const runs = await store.listRuns()\n  const projectIds = [...new Set(runs.map((r) => r.projectId).filter((p): p is string => !!p))]\n  return Promise.all(projectIds.map((p) => scoreProject(store, p)))\n}\n\nfunction latestByLayer(runs: Run[], layer: Run['layer']): Run | undefined {\n  const filtered = runs.filter((r) => r.layer === layer).sort((a, b) => b.startedAt - a.startedAt)\n  return filtered[0]\n}\n\nasync function extractMetaScore(store: TraceStore, builderRunId: string): Promise<number | null> {\n  const js = await judgeSpans(store, builderRunId)\n  const meta = js.find(\n    (s) => s.judgeId === 'builder-meta' && s.dimension === 'user_intent_satisfaction',\n  )\n  if (!meta) return null\n  // Normalize score to 0..1. Accept 0-1 natively; 0-10 scale is also common.\n  if (meta.score >= 0 && meta.score <= 1) return meta.score\n  if (meta.score >= 0 && meta.score <= 10) return meta.score / 10\n  return null\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;AA8CO,IAAM,iBAAN,MAAqB;AAAA,EAClB;AAAA,EACA;AAAA,EACC;AAAA,EACA;AAAA,EACD;AAAA,EACA;AAAA,EACA;AAAA,EAER,YAAY,OAAmB,MAA0B,QAAwB;AAC/E,SAAK,QAAQ;AACb,SAAK,YAAY,KAAK;AACtB,SAAK,SAAS,KAAK,UAAU,SAAS;AACtC,SAAK,gBAAgB;AACrB,SAAK,iBAAiB,IAAI,aAAa,KAAK;AAAA,EAC9C;AAAA;AAAA,EAGA,MAAM,UAAU,aAAa,GAAG,KAAK,SAAS,SAA0B;AACtE,UAAM,KAAK,eAAe,SAAS;AAAA,MACjC;AAAA,MACA,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,OAAO;AAAA,IACT,CAAC;AACD,SAAK,eAAe,KAAK,eAAe;AACxC,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,UAAwB;AAC1B,QAAI,CAAC,KAAK,aAAc,OAAM,IAAI,MAAM,gDAAgD;AACxF,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,KAAK,SAAgF;AACzF,QAAI,CAAC,KAAK,aAAc,OAAM,IAAI,MAAM,6CAA6C;AACrF,UAAM,eAAe,IAAI,aAAa,KAAK,KAAK;AAChD,UAAM,aAAa,SAAS;AAAA,MAC1B,YAAY,QAAQ,cAAc,GAAG,KAAK,SAAS;AAAA,MACnD,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,aAAa,KAAK;AAAA,MAClB,OAAO;AAAA,IACT,CAAC;AACD,UAAM,UAAU,IAAI,eAAe,QAAQ,UAAU,KAAK,aAAa;AACvE,UAAM,SAAS,MAAM,QAAQ,IAAI,QAAQ,SAAS,YAAY;AAC9D,UAAM,aAAa,OAAO;AAAA,MACxB,MAAM,OAAO;AAAA,MACb,OAAO,OAAO;AAAA,MACd,cAAc,OAAO,SAAS,YAAY;AAAA,IAC5C,CAAC;AACD,SAAK,iBAAiB,aAAa;AACnC,WAAO,EAAE,OAAO,aAAa,OAAO,OAAO;AAAA,EAC7C;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,eAAe,SAA8D;AACjF,UAAM,cAAc,KAAK,kBAAkB,KAAK;AAChD,QAAI,CAAC;AACH,YAAM,IAAI,MAAM,gEAAgE;AAClF,UAAM,EAAE,UAAU,OAAO,IAAI;AAC7B,UAAM,SAAS,MAAM,sBAAsB,UAAU,KAAK,OAAO;AAAA,MAC/D,QAAQ,UAAU,KAAK;AAAA,MACvB,YAAY,EAAE,SAAS,QAAW,WAAW,QAAW,kBAAkB,OAAU;AAAA,IACtF,CAAC;AAED,UAAM,KAAK,MAAM,UAAU,OAAO,OAAO;AAAA,MACvC;AAAA,MACA,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,OAAO;AAAA,IACT,CAAC;AACD,WAAO;AAAA,EACT;AAAA;AAAA;AAAA,EAIA,MAAM,gBAAgB,OAAe,WAAmC;AACtE,QAAI,CAAC,KAAK;AACR,YAAM,IAAI,MAAM,wDAAwD;AAC1E,UAAM,KAAK,eAAe,YAAY;AAAA,MACpC,SAAS;AAAA,MACT,cAAc,KAAK;AAAA;AAAA,MACnB,WAAW;AAAA,MACX;AAAA,MACA;AAAA,MACA,MAAM;AAAA,IACR,CAAC;AAAA,EACH;AAAA;AAAA,EAGA,MAAM,QAAQ,SAA2E;AACvF,UAAM,KAAK,eAAe,OAAO;AAAA,MAC/B,MAAM,QAAQ;AAAA,MACd,OAAO,QAAQ;AAAA,MACf,OAAO,QAAQ;AAAA,IACjB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,gBAAgB,YAA2C;AAC/D,UAAM,cAAc,KAAK,kBAAkB,KAAK;AAChD,QAAI,CAAC;AACH,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AACF,UAAM,UAAU,IAAI,aAAa,KAAK,KAAK;AAC3C,UAAM,QAAQ,SAAS;AAAA,MACrB;AAAA,MACA,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA,OAAO;AAAA,IACT,CAAC;AACD,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,iBAAiB,MAKH;AAClB,QAAI,CAAC,KAAK;AACR,YAAM,IAAI,MAAM,yDAAyD;AAC3E,UAAM,UAAU,IAAI,aAAa,KAAK,KAAK;AAC3C,UAAM,QAAQ,SAAS;AAAA,MACrB,YAAY,KAAK,cAAc,GAAG,KAAK,SAAS;AAAA,MAChD,WAAW,KAAK;AAAA,MAChB,QAAQ,KAAK;AAAA,MACb,aAAa,KAAK;AAAA,MAClB,OAAO;AAAA,IACT,CAAC;AACD,UAAM,QAAQ,OAAO;AAAA,MACnB,MAAM,KAAK;AAAA,MACX,OAAO,KAAK;AAAA,MACZ,cAAc,KAAK,OAAO,YAAY;AAAA,MACtC,OAAO,KAAK;AAAA,IACd,CAAC;AACD,SAAK,iBAAiB,QAAQ;AAC9B,WAAO,QAAQ;AAAA,EACjB;AAAA,EAEA,IAAI,sBAA0C;AAC5C,WAAO,KAAK;AAAA,EACd;AAAA,EACA,IAAI,oBAAwC;AAC1C,WAAO,KAAK;AAAA,EACd;AACF;AAOA,eAAsB,qBACpB,OACA,WAOC;AACD,QAAM,OAAO,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;AAC/C,QAAM,WAAW,KACd,OAAO,CAAC,MAAM,EAAE,UAAU,SAAS,EACnC,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3C,QAAM,YAAY,KACf,OAAO,CAAC,MAAM,EAAE,UAAU,WAAW,EACrC,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3C,QAAM,iBAAiB,KACpB,OAAO,CAAC,MAAM,EAAE,UAAU,aAAa,EACvC,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3C,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,gBAAgB,SAAS,CAAC;AAAA,IAC1B,cAAc,UAAU,CAAC;AAAA,IACzB,oBAAoB;AAAA,EACtB;AACF;AAEA,SAAS,WAAmB;AAC1B,MAAI,OAAO,WAAW,QAAQ,eAAe,WAAY,QAAO,WAAW,OAAO,WAAW;AAC7F,SAAO,GAAG,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,EAAE,CAAC;AAC9E;;;AC5NO,SAAS,gBAAgB,SAAuD;AACrF,QAAM,mBAAmB,QAAQ,OAAO,CAAC,MAAM,EAAE,QAAQ,EAAE;AAC3D,SAAO;AAAA,IACL,aAAa;AAAA,MACX;AAAA,MACA,CAAC,MAAM,EAAE;AAAA,MACT,CAAC,MAAM,EAAE;AAAA,IACX;AAAA,IACA,eAAe;AAAA,MACb;AAAA,MACA,CAAC,MAAM,EAAE;AAAA,MACT,CAAC,MAAM,EAAE;AAAA,IACX;AAAA,IACA,gBAAgB;AAAA,MACd;AAAA,MACA,CAAC,MAAM,EAAE;AAAA,MACT,CAAC,MAAM,EAAE;AAAA,IACX;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,SACP,SACA,GACA,GAC8B;AAC9B,QAAM,KAAe,CAAC;AACtB,QAAM,KAAe,CAAC;AACtB,aAAW,KAAK,SAAS;AACvB,UAAM,IAAI,EAAE,CAAC;AACb,UAAM,IAAI,EAAE,CAAC;AACb,QAAI,MAAM,QAAQ,MAAM,QAAQ,OAAO,SAAS,CAAC,KAAK,OAAO,SAAS,CAAC,GAAG;AACxE,SAAG,KAAK,CAAC;AACT,SAAG,KAAK,CAAC;AAAA,IACX;AAAA,EACF;AACA,MAAI,GAAG,SAAS,EAAG,QAAO;AAC1B,SAAO;AAAA,IACL,GAAG,GAAG;AAAA,IACN,SAAS,SAAS,IAAI,EAAE;AAAA,IACxB,UAAU,UAAU,IAAI,EAAE;AAAA,EAC5B;AACF;;;ACjCO,IAAM,kBAAN,MAAsB;AAAA,EAC3B,YAAoB,OAAmB;AAAnB;AAAA,EAAoB;AAAA,EAApB;AAAA,EAEpB,MAAM,eAA0C;AAC9C,UAAM,OAAO,MAAM,KAAK,MAAM,SAAS;AACvC,UAAM,YAAY,oBAAI,IAAmB;AACzC,eAAW,KAAK,MAAM;AACpB,UAAI,CAAC,EAAE,UAAW;AAClB,YAAM,MAAM,UAAU,IAAI,EAAE,SAAS,KAAK,CAAC;AAC3C,UAAI,KAAK,CAAC;AACV,gBAAU,IAAI,EAAE,WAAW,GAAG;AAAA,IAChC;AACA,UAAM,YAA8B,CAAC;AACrC,eAAW,CAAC,WAAW,WAAW,KAAK,WAAW;AAChD,YAAM,SAAS,YAAY,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC3E,YAAM,QAAQ,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,SAAS;AAC7D,YAAM,SAAS,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,WAAW;AAChE,YAAM,WAAW,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,aAAa;AACpE,YAAM,SAAS,OAAO,CAAC;AACvB,UAAI,CAAC,OAAQ;AACb,gBAAU,KAAK;AAAA,QACb;AAAA,QACA,WAAW,MAAM;AAAA,QACjB,YAAY,OAAO;AAAA,QACnB,iBAAiB,SAAS;AAAA,QAC1B,gBAAgB,OAAO;AAAA,QACvB,cAAc,MAAM,CAAC,GAAG;AAAA,QACxB,eAAe,OAAO,UAClB,EAAE,MAAM,OAAO,QAAQ,QAAQ,OAAO,OAAO,OAAO,QAAQ,MAAM,IAClE;AAAA,MACN,CAAC;AAAA,IACH;AACA,WAAO,UAAU,KAAK,CAAC,GAAG,MAAM,EAAE,iBAAiB,EAAE,cAAc;AAAA,EACrE;AAAA,EAEA,MAAM,gBAAgB,WAAoD;AACxE,UAAM,OAAO,MAAM,KAAK,MAAM,SAAS,EAAE,UAAU,CAAC;AACpD,UAAM,UAAU,KAAK,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACrE,WAAO,QAAQ,IAAI,CAAC,SAAS;AAAA,MAC3B;AAAA,MACA,aACE,IAAI,UAAU,YACV,SACA,IAAI,UAAU,cACZ,UACA,IAAI,UAAU,gBACZ,YACA;AAAA,IACZ,EAAE;AAAA,EACJ;AAAA,EAEA,MAAM,aAAa,WAA2C;AAC5D,UAAM,eAAe,MAAM,KAAK,MAAM,SAAS,EAAE,WAAW,OAAO,UAAU,CAAC,GAAG;AAAA,MAC/E,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE;AAAA,IAC5B;AACA,UAAM,cAAc,OAAO,UAAkB,KAAK,MAAM,SAAS,EAAE,aAAa,MAAM,CAAC;AACvF,UAAM,MAAqB,CAAC;AAC5B,eAAW,OAAO,aAAa;AAC7B,YAAM,QAAQ,MAAM,KAAK,MAAM,MAAM,EAAE,OAAO,IAAI,MAAM,CAAC;AACzD,YAAM,WAAW,MAAM,YAAY,IAAI,KAAK;AAC5C,YAAM,QAAQ,SAAS,KAAK,CAAC,MAAM,EAAE,UAAU,WAAW;AAC1D,YAAM,UAAoB,CAAC;AAG3B,UAAI,OAAO;AACT,cAAM,SAAS,MAAM,YAAY,MAAM,KAAK;AAC5C,mBAAW,KAAK,OAAQ,KAAI,EAAE,UAAU,cAAe,SAAQ,KAAK,EAAE,KAAK;AAAA,MAC7E;AACA,iBAAW,KAAK,SAAU,KAAI,EAAE,UAAU,cAAe,SAAQ,KAAK,EAAE,KAAK;AAC7E,UAAI,KAAK;AAAA,QACP,QAAQ,IAAI,UAAU,IAAI;AAAA,QAC1B;AAAA,QACA,cAAc,IAAI;AAAA,QAClB,WAAW,IAAI;AAAA,QACf,SAAS,IAAI;AAAA,QACb,QAAQ,IAAI;AAAA,QACZ,SAAS,IAAI;AAAA,QACb,UAAU,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,EAAE;AAAA,QAChD,WAAW,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE;AAAA,QAClD,YAAY,OAAO;AAAA,QACnB,kBAAkB;AAAA,MACpB,CAAC;AAAA,IACH;AACA,WAAO;AAAA,EACT;AACF;;;ACzEA,eAAsB,aACpB,OACA,WACkC;AAClC,QAAM,UAAU,MAAM,MAAM,SAAS,EAAE,UAAU,CAAC;AAClD,QAAM,UAAU,cAAc,SAAS,SAAS;AAChD,QAAM,QAAQ,cAAc,SAAS,WAAW;AAChD,QAAM,UAAU,QAAQ,OAAO,CAAC,MAAM,EAAE,UAAU,aAAa;AAE/D,QAAM,YAAY,UAAU,MAAM,iBAAiB,OAAO,QAAQ,KAAK,IAAI;AAC3E,QAAM,aAAa,OAAO,SAAS,SAAS;AAC5C,QAAM,gBAAgB,QACnB,IAAI,CAAC,MAAM,EAAE,SAAS,KAAK,EAC3B,OAAO,CAAC,MAAmB,OAAO,MAAM,QAAQ;AACnD,QAAM,eACJ,cAAc,SAAS,IACnB,cAAc,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,cAAc,SACzD;AACN,QAAM,gBAAgB,QAAQ,OAAO,CAAC,MAAM,EAAE,SAAS,SAAS,IAAI,EAAE;AACtE,QAAM,kBAAkB,QAAQ,SAAS,IAAI,gBAAgB,QAAQ,SAAS;AAE9E,QAAM,OAAoB,QAAQ,WAAW,IAAI,kBAAkB;AACnE,QAAM,WACJ,SAAS,kBACL,cAAc,QAAQ,eAAe,OACrC,cAAc,QAAQ,eAAe,QAAQ,iBAAiB;AAEpE,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,cAAc,SAAS;AAAA,IACvB;AAAA,IACA,YAAY,OAAO;AAAA,IACnB;AAAA,IACA,kBAAkB,QAAQ,IAAI,CAAC,MAAM,EAAE,KAAK;AAAA,IAC5C;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAGA,eAAsB,iBAAiB,OAAuD;AAC5F,QAAM,OAAO,MAAM,MAAM,SAAS;AAClC,QAAM,aAAa,CAAC,GAAG,IAAI,IAAI,KAAK,IAAI,CAAC,MAAM,EAAE,SAAS,EAAE,OAAO,CAAC,MAAmB,CAAC,CAAC,CAAC,CAAC,CAAC;AAC5F,SAAO,QAAQ,IAAI,WAAW,IAAI,CAAC,MAAM,aAAa,OAAO,CAAC,CAAC,CAAC;AAClE;AAEA,SAAS,cAAc,MAAa,OAAsC;AACxE,QAAM,WAAW,KAAK,OAAO,CAAC,MAAM,EAAE,UAAU,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC/F,SAAO,SAAS,CAAC;AACnB;AAEA,eAAe,iBAAiB,OAAmB,cAA8C;AAC/F,QAAM,KAAK,MAAM,WAAW,OAAO,YAAY;AAC/C,QAAM,OAAO,GAAG;AAAA,IACd,CAAC,MAAM,EAAE,YAAY,kBAAkB,EAAE,cAAc;AAAA,EACzD;AACA,MAAI,CAAC,KAAM,QAAO;AAElB,MAAI,KAAK,SAAS,KAAK,KAAK,SAAS,EAAG,QAAO,KAAK;AACpD,MAAI,KAAK,SAAS,KAAK,KAAK,SAAS,GAAI,QAAO,KAAK,QAAQ;AAC7D,SAAO;AACT;","names":[]}