@lunora/testing 1.0.0-alpha.111 → 1.0.0-alpha.113

package/dist/index.d.mts

@@ -140,7 +140,26 @@ declare const agentHarness: (agent: AgentDefinition, options: AgentHarnessOption
 declare const finalTurn: (text: string, extra?: Partial<AgentGenerateResult>) => AgentGenerateResult;
 /** A single-tool-call turn: the loop runs the named tool with `input`. */
 declare const toolCallTurn: (id: string, name: string, input: unknown, text?: string) => AgentGenerateResult;
-/** The primitive an eval attaches to a span: a numeric score or a string label. */
+/**
+ * Shared, bundler-inlined builder for the OpenTelemetry `gen_ai.evaluation.*`
+ * attributes an AI **evaluation** verdict (a scorer's `{name, score, label?}`)
+ * contributes to a **generation span**.
+ *
+ * This is the emit-time counterpart of the cloud OTLP decoder, which reads
+ * `gen_ai.evaluation.<name>.score` (number) and optional
+ * `gen_ai.evaluation.<name>.label` (string) attribute pairs back off a generation
+ * span (`EVALUATION_PREFIX = "gen_ai.evaluation."`). The framework emits exactly
+ * that pair here so a score rides the same trace as the generation it grades.
+ *
+ * It lives in `shared/` because more than one layer needs the identical wire
+ * format with no runtime dependency edge between them: `@lunora/do` builds it into
+ * a live `ctx.trace` span's post-hoc attributes (`SpanHandle.recordEvaluation`),
+ * and `@lunora/server` mirrors the handle shape structurally. `@lunora/testing`
+ * ships a parallel test-time helper (`recordEvaluation` / `evaluationAttributes`)
+ * that targets the same `gen_ai.evaluation.*` contract for span-less eval
+ * events/metrics. Keep this file genuinely zero-dependency so inlining stays sound.
+ */
+/** The primitive an eval contributes to a span: a numeric score or a string label. */
 type EvaluationAttributeValue = number | string;
 /**
  * Structural slice of the post-hoc span handle `ctx.trace` hands its body (see the
@@ -194,6 +213,11 @@ interface RecordEvaluationInput {
  * Build the `gen_ai.evaluation.NAME.*` attribute bag for one eval verdict — the
  * `.score` (number) always, the `.label` (string) when a label is given. Exported
  * so a caller can emit the score as a standalone event/metric without a span.
+ *
+ * Delegates to the shared `shared/evaluation-attributes.ts` builder so
+ * `@lunora/testing`'s scorers and the runtime's own `recordEvaluation` emit the
+ * identical wire format — only the thrown error type differs, wrapped here as a
+ * `LunoraError` to match this package's public error contract.
  */
 declare const evaluationAttributes: (input: Pick<RecordEvaluationInput, "label" | "name" | "score">) => Record<string, EvaluationAttributeValue>;
 /**
@@ -210,11 +234,14 @@ interface FakeScheduledJob extends ScheduledJob {
   args: Record<string, unknown>;
 }
 /**
- * A single scheduled-job failure captured during an `advance()` / `runPending()`
- * sweep. Production's scheduler isolates per-job failures (one bad job does not
- * abort the rest), so the fake scheduler does the same — but, being a test
- * harness, it never swallows the error: every failure is recorded here so tests
- * can still assert on it.
+ * A single scheduled-job failure that exhausted its retry budget, captured
+ * during an `advance()` / `runPending()` sweep. Mirrors `SchedulerDO`'s
+ * dead-letter park (`recordRetry()`, `packages/scheduler/src/scheduler-do.ts:875-909`):
+ * a job that fails while it still has retries left is silently re-enqueued
+ * with backoff and is NOT recorded here. Only once a job's `attempts` exceeds
+ * `@lunora/scheduler`'s `MAX_RETRY_ATTEMPTS` does it land here — being a test
+ * harness, the fake scheduler never swallows a terminal failure, so tests can
+ * still assert on it.
  */
 interface ScheduledJobFailure {
   /** The args the job was dispatched with. */
@@ -235,40 +262,55 @@ interface FakeSchedulerControls {
    * Advance the virtual clock by `ms` milliseconds, executing all jobs whose
    * `scheduledFor` timestamp is now at or before the new virtual "now". Jobs
    * are dispatched in `scheduledFor` order (oldest first). Newly queued jobs
-   * (scheduled by an executed job during the advance) are NOT re-evaluated in
-   * the same advance call — callers should advance again if needed.
+   * (scheduled by an executed job, or re-enqueued as a retry, during the
+   * advance) are NOT re-evaluated in the same advance call — callers should
+   * advance again if needed.
    *
    * Per-job failures are isolated (matching production): a job that throws does
-   * NOT prevent the remaining due jobs from running. After every due job has
-   * run, the failures are surfaced — they are recorded on
-   * {@link FakeSchedulerControls.failures} and, by default, re-thrown so a test
-   * still sees the error. A single failure is re-thrown verbatim; multiple
-   * failures are aggregated into an `AggregateError`. Pass
+   * NOT prevent the remaining due jobs from running. A failure with retries
+   * left is silently re-enqueued with exponential backoff on the virtual clock
+   * — mirroring `SchedulerDO`'s default retry policy (`MAX_RETRY_ATTEMPTS`
+   * retries, `RETRY_BASE_DELAY_MS` base delay, doubling; both imported from
+   * `@lunora/scheduler`) — rather than surfaced. Advancing far enough to
+   * observe a terminal failure therefore costs the WHOLE backoff schedule
+   * (30s + 60s + 120s + 240s + 480s = 930s of virtual clock at today's
+   * defaults), not a single tick.
+   * Only once a job's retry budget is exhausted is the failure surfaced: it is
+   * recorded on {@link FakeSchedulerControls.failures} and, by default,
+   * re-thrown so a test still sees the error. A single such failure is
+   * re-thrown verbatim; multiple are aggregated into an `AggregateError`. Pass
    * `{ throwOnError: false }` to suppress the re-throw and inspect
    * {@link FakeSchedulerControls.failures} (and the returned count) instead.
    *
-   * Returns the number of jobs that were executed (including failed ones).
+   * Returns the number of jobs dispatched this sweep, including ones that
+   * failed and were silently retried, and ones that failed terminally.
    */
   advance: (ms: number, options?: SweepOptions) => Promise<number>;
   /**
-   * All scheduled-job failures captured so far, in execution order, across
-   * every `advance()` / `runPending()` call on this harness. Always available,
-   * even when `throwOnError: false` suppressed the re-throw. The list is a
-   * snapshot — mutating it does not affect the scheduler.
+   * All scheduled-job failures that exhausted their retry budget, in
+   * execution order, across every `advance()` / `runPending()` call on this
+   * harness. A failure with retries remaining is NOT recorded here — see
+   * {@link ScheduledJobFailure}. Always available, even when
+   * `throwOnError: false` suppressed the re-throw. The list is a snapshot —
+   * mutating it does not affect the scheduler.
    */
   failures: () => ScheduledJobFailure[];
   /**
    * List all pending jobs (those not yet executed or cancelled) in the order
-   * they were enqueued.
+   * they were enqueued. A job currently waiting out its retry backoff is
+   * still pending (visible here with its `attempts` count incremented and
+   * `scheduledFor` pushed out) until its budget is exhausted.
    */
   list: () => FakeScheduledJob[];
   /**
    * Execute all currently pending jobs regardless of their `scheduledFor`
    * time. Equivalent to advancing to `Infinity`. Returns the number of jobs
-   * executed (including failed ones).
+   * dispatched this sweep (including failed ones, retried or terminal).
    *
-   * Failure isolation and surfacing match {@link FakeSchedulerControls.advance}:
-   * one failing job does not abort the rest, and failures are recorded on
+   * Failure isolation, retry, and surfacing match
+   * {@link FakeSchedulerControls.advance}: one failing job does not abort the
+   * rest, a failure under the retry budget is silently re-enqueued rather
+   * than surfaced, and only a terminal failure is recorded on
    * {@link FakeSchedulerControls.failures} and re-thrown unless
    * `{ throwOnError: false }` is passed.
    */
@@ -578,6 +620,20 @@ interface Scorer {
   name: string;
   score: (sample: ScorerSample) => Promise<ScoreResult | number> | ScoreResult | number;
 }
+/**
+ * What a `produce` runner may return: the output text alone, or the text plus
+ * metadata describing what the run actually did.
+ *
+ * The metadata form exists for scorers that judge more than the final string —
+ * a retrieval scorer needs the ranked ids the run retrieved, which only the run
+ * can know. It is merged OVER the case's own metadata before scoring.
+ */
+interface ProducedOutput {
+  /** Merged over the case's `metadata`, then handed to every scorer. */
+  metadata?: Record<string, unknown>;
+  /** The output text under test. */
+  output: string;
+}
 /** One dataset case: an input and its optional gold answer/metadata. */
 interface EvalCase {
   expected?: string;
@@ -628,5 +684,70 @@ declare const scoreSample: (sample: ScorerSample, scorers: ReadonlyArray<Scorer>
  * give each its own thread/key inside `produce` if the producer is stateful.
  * Returns per-case results plus the mean of their averages.
  */
-declare const evaluate: (cases: ReadonlyArray<EvalCase>, produce: (input: string) => Promise<string> | string, scorers: ReadonlyArray<Scorer>) => Promise<EvalResult>;
-export { type AgentHarness, type AgentHarnessOptions, type AgentRunOverrides, type EvalCase, type EvalItemResult, type EvalResult, type EvaluationAttributeValue, type EvaluationMetrics, type EvaluationSpanHandle, type FakeScheduledJob, type FakeSchedulerControls, type FunctionRegistry, type HarnessDispatch, type HarnessMessage, type HarnessThread, type LunoraTestOptions, type RecordEvaluationInput, type ScheduledJobFailure, type ScoreResult, type Scorer, type ScorerSample, type SweepOptions, type TestHarness, type TestIdentity, type TestSubscription, agentHarness, containsScorer, evaluate, evaluationAttributes, exactMatchScorer, finalTurn, keywordScorer, llmScorer, lunoraTest, recordEvaluation, regexScorer, scoreSample, toolCallTurn };
+declare const evaluate: (cases: ReadonlyArray<EvalCase>, produce: (input: string) => Promise<ProducedOutput | string> | ProducedOutput | string, scorers: ReadonlyArray<Scorer>) => Promise<EvalResult>;
+/** Where a retrieval scorer reads its two id lists from. */
+interface RetrievalScorerOptions {
+  /** Metadata key holding the gold relevant ids. Default `"relevant"`. */
+  relevantKey?: string;
+  /** Metadata key holding the run's ranked retrieved ids. Default `"retrieved"`. */
+  retrievedKey?: string;
+}
+/**
+ * Recall@k scorer — what fraction of the gold passages made it into the top
+ * `k`.
+ *
+ * This is the ceiling on everything downstream: a passage retrieval never
+ * returned is one no reranker can promote and no prompt can cite. Omit `k` to
+ * score the whole retrieved list.
+ */
+declare const recallAtK: (k?: number, options?: RetrievalScorerOptions) => Scorer;
+/**
+ * Precision@k scorer — what fraction of the top `k` retrieved passages are
+ * gold.
+ *
+ * This is the counterweight to recall: padding `topK` raises recall for free
+ * while burying the answer in noise the model has to read past, and pay for.
+ * Omit `k` to score the whole retrieved list.
+ */
+declare const precisionAtK: (k?: number, options?: RetrievalScorerOptions) => Scorer;
+/**
+ * Mean Reciprocal Rank scorer — `1 / rank` of the first gold passage: 1 if it
+ * is first, 0.5 if second, 0 if absent.
+ *
+ * This is the metric that notices ordering, which recall cannot. A gold passage
+ * at rank 20 counts the same as rank 1 for recall@20, but only one of those
+ * survives a context-window trim or a model that skims the top of its prompt.
+ */
+declare const mrrScorer: (options?: RetrievalScorerOptions) => Scorer;
+/**
+ * Normalized Discounted Cumulative Gain scorer — relevance discounted
+ * logarithmically by rank, divided by the best achievable arrangement.
+ *
+ * This is the one to gate on when comparing retrieval strategies. Unlike recall
+ * it is sensitive to order, and unlike MRR it credits every gold passage rather
+ * than only the first — so it is the metric that can actually say whether a
+ * reranker or a hybrid leg helped.
+ *
+ * Relevance is binary: an id is gold or it is not, which is what a gold-id set
+ * expresses. Graded relevance would need per-id weights.
+ */
+declare const ndcgAtK: (k?: number, options?: RetrievalScorerOptions) => Scorer;
+/**
+ * Groundedness scorer — does the answer assert only what the retrieved context
+ * supports?
+ *
+ * This is the generation-side counterpart to the metrics above: perfect
+ * retrieval still fails if the model answers from its own weights. It scores
+ * the output against the context via an injected `judge`, the same shape
+ * `llmScorer` takes, so this stays model-agnostic and mockable.
+ *
+ * Reads the context from `metadata.context` by default — the `context` string
+ * `retrieve()` already returns. Fails closed: no context means nothing could
+ * have grounded the answer.
+ */
+declare const groundednessScorer: (options: {
+  contextKey?: string;
+  judge: (prompt: string) => Promise<string>;
+  name?: string;
+}) => Scorer;
+export { type AgentHarness, type AgentHarnessOptions, type AgentRunOverrides, type EvalCase, type EvalItemResult, type EvalResult, type EvaluationAttributeValue, type EvaluationMetrics, type EvaluationSpanHandle, type FakeScheduledJob, type FakeSchedulerControls, type FunctionRegistry, type HarnessDispatch, type HarnessMessage, type HarnessThread, type LunoraTestOptions, type ProducedOutput, type RecordEvaluationInput, type RetrievalScorerOptions, type ScheduledJobFailure, type ScoreResult, type Scorer, type ScorerSample, type SweepOptions, type TestHarness, type TestIdentity, type TestSubscription, agentHarness, containsScorer, evaluate, evaluationAttributes, exactMatchScorer, finalTurn, groundednessScorer, keywordScorer, llmScorer, lunoraTest, mrrScorer, ndcgAtK, precisionAtK, recallAtK, recordEvaluation, regexScorer, scoreSample, toolCallTurn };

package/dist/index.d.ts

@@ -140,7 +140,26 @@ declare const agentHarness: (agent: AgentDefinition, options: AgentHarnessOption
 declare const finalTurn: (text: string, extra?: Partial<AgentGenerateResult>) => AgentGenerateResult;
 /** A single-tool-call turn: the loop runs the named tool with `input`. */
 declare const toolCallTurn: (id: string, name: string, input: unknown, text?: string) => AgentGenerateResult;
-/** The primitive an eval attaches to a span: a numeric score or a string label. */
+/**
+ * Shared, bundler-inlined builder for the OpenTelemetry `gen_ai.evaluation.*`
+ * attributes an AI **evaluation** verdict (a scorer's `{name, score, label?}`)
+ * contributes to a **generation span**.
+ *
+ * This is the emit-time counterpart of the cloud OTLP decoder, which reads
+ * `gen_ai.evaluation.<name>.score` (number) and optional
+ * `gen_ai.evaluation.<name>.label` (string) attribute pairs back off a generation
+ * span (`EVALUATION_PREFIX = "gen_ai.evaluation."`). The framework emits exactly
+ * that pair here so a score rides the same trace as the generation it grades.
+ *
+ * It lives in `shared/` because more than one layer needs the identical wire
+ * format with no runtime dependency edge between them: `@lunora/do` builds it into
+ * a live `ctx.trace` span's post-hoc attributes (`SpanHandle.recordEvaluation`),
+ * and `@lunora/server` mirrors the handle shape structurally. `@lunora/testing`
+ * ships a parallel test-time helper (`recordEvaluation` / `evaluationAttributes`)
+ * that targets the same `gen_ai.evaluation.*` contract for span-less eval
+ * events/metrics. Keep this file genuinely zero-dependency so inlining stays sound.
+ */
+/** The primitive an eval contributes to a span: a numeric score or a string label. */
 type EvaluationAttributeValue = number | string;
 /**
  * Structural slice of the post-hoc span handle `ctx.trace` hands its body (see the
@@ -194,6 +213,11 @@ interface RecordEvaluationInput {
  * Build the `gen_ai.evaluation.NAME.*` attribute bag for one eval verdict — the
  * `.score` (number) always, the `.label` (string) when a label is given. Exported
  * so a caller can emit the score as a standalone event/metric without a span.
+ *
+ * Delegates to the shared `shared/evaluation-attributes.ts` builder so
+ * `@lunora/testing`'s scorers and the runtime's own `recordEvaluation` emit the
+ * identical wire format — only the thrown error type differs, wrapped here as a
+ * `LunoraError` to match this package's public error contract.
  */
 declare const evaluationAttributes: (input: Pick<RecordEvaluationInput, "label" | "name" | "score">) => Record<string, EvaluationAttributeValue>;
 /**
@@ -210,11 +234,14 @@ interface FakeScheduledJob extends ScheduledJob {
   args: Record<string, unknown>;
 }
 /**
- * A single scheduled-job failure captured during an `advance()` / `runPending()`
- * sweep. Production's scheduler isolates per-job failures (one bad job does not
- * abort the rest), so the fake scheduler does the same — but, being a test
- * harness, it never swallows the error: every failure is recorded here so tests
- * can still assert on it.
+ * A single scheduled-job failure that exhausted its retry budget, captured
+ * during an `advance()` / `runPending()` sweep. Mirrors `SchedulerDO`'s
+ * dead-letter park (`recordRetry()`, `packages/scheduler/src/scheduler-do.ts:875-909`):
+ * a job that fails while it still has retries left is silently re-enqueued
+ * with backoff and is NOT recorded here. Only once a job's `attempts` exceeds
+ * `@lunora/scheduler`'s `MAX_RETRY_ATTEMPTS` does it land here — being a test
+ * harness, the fake scheduler never swallows a terminal failure, so tests can
+ * still assert on it.
  */
 interface ScheduledJobFailure {
   /** The args the job was dispatched with. */
@@ -235,40 +262,55 @@ interface FakeSchedulerControls {
    * Advance the virtual clock by `ms` milliseconds, executing all jobs whose
    * `scheduledFor` timestamp is now at or before the new virtual "now". Jobs
    * are dispatched in `scheduledFor` order (oldest first). Newly queued jobs
-   * (scheduled by an executed job during the advance) are NOT re-evaluated in
-   * the same advance call — callers should advance again if needed.
+   * (scheduled by an executed job, or re-enqueued as a retry, during the
+   * advance) are NOT re-evaluated in the same advance call — callers should
+   * advance again if needed.
    *
    * Per-job failures are isolated (matching production): a job that throws does
-   * NOT prevent the remaining due jobs from running. After every due job has
-   * run, the failures are surfaced — they are recorded on
-   * {@link FakeSchedulerControls.failures} and, by default, re-thrown so a test
-   * still sees the error. A single failure is re-thrown verbatim; multiple
-   * failures are aggregated into an `AggregateError`. Pass
+   * NOT prevent the remaining due jobs from running. A failure with retries
+   * left is silently re-enqueued with exponential backoff on the virtual clock
+   * — mirroring `SchedulerDO`'s default retry policy (`MAX_RETRY_ATTEMPTS`
+   * retries, `RETRY_BASE_DELAY_MS` base delay, doubling; both imported from
+   * `@lunora/scheduler`) — rather than surfaced. Advancing far enough to
+   * observe a terminal failure therefore costs the WHOLE backoff schedule
+   * (30s + 60s + 120s + 240s + 480s = 930s of virtual clock at today's
+   * defaults), not a single tick.
+   * Only once a job's retry budget is exhausted is the failure surfaced: it is
+   * recorded on {@link FakeSchedulerControls.failures} and, by default,
+   * re-thrown so a test still sees the error. A single such failure is
+   * re-thrown verbatim; multiple are aggregated into an `AggregateError`. Pass
    * `{ throwOnError: false }` to suppress the re-throw and inspect
    * {@link FakeSchedulerControls.failures} (and the returned count) instead.
    *
-   * Returns the number of jobs that were executed (including failed ones).
+   * Returns the number of jobs dispatched this sweep, including ones that
+   * failed and were silently retried, and ones that failed terminally.
    */
   advance: (ms: number, options?: SweepOptions) => Promise<number>;
   /**
-   * All scheduled-job failures captured so far, in execution order, across
-   * every `advance()` / `runPending()` call on this harness. Always available,
-   * even when `throwOnError: false` suppressed the re-throw. The list is a
-   * snapshot — mutating it does not affect the scheduler.
+   * All scheduled-job failures that exhausted their retry budget, in
+   * execution order, across every `advance()` / `runPending()` call on this
+   * harness. A failure with retries remaining is NOT recorded here — see
+   * {@link ScheduledJobFailure}. Always available, even when
+   * `throwOnError: false` suppressed the re-throw. The list is a snapshot —
+   * mutating it does not affect the scheduler.
    */
   failures: () => ScheduledJobFailure[];
   /**
    * List all pending jobs (those not yet executed or cancelled) in the order
-   * they were enqueued.
+   * they were enqueued. A job currently waiting out its retry backoff is
+   * still pending (visible here with its `attempts` count incremented and
+   * `scheduledFor` pushed out) until its budget is exhausted.
    */
   list: () => FakeScheduledJob[];
   /**
    * Execute all currently pending jobs regardless of their `scheduledFor`
    * time. Equivalent to advancing to `Infinity`. Returns the number of jobs
-   * executed (including failed ones).
+   * dispatched this sweep (including failed ones, retried or terminal).
    *
-   * Failure isolation and surfacing match {@link FakeSchedulerControls.advance}:
-   * one failing job does not abort the rest, and failures are recorded on
+   * Failure isolation, retry, and surfacing match
+   * {@link FakeSchedulerControls.advance}: one failing job does not abort the
+   * rest, a failure under the retry budget is silently re-enqueued rather
+   * than surfaced, and only a terminal failure is recorded on
    * {@link FakeSchedulerControls.failures} and re-thrown unless
    * `{ throwOnError: false }` is passed.
    */
@@ -578,6 +620,20 @@ interface Scorer {
   name: string;
   score: (sample: ScorerSample) => Promise<ScoreResult | number> | ScoreResult | number;
 }
+/**
+ * What a `produce` runner may return: the output text alone, or the text plus
+ * metadata describing what the run actually did.
+ *
+ * The metadata form exists for scorers that judge more than the final string —
+ * a retrieval scorer needs the ranked ids the run retrieved, which only the run
+ * can know. It is merged OVER the case's own metadata before scoring.
+ */
+interface ProducedOutput {
+  /** Merged over the case's `metadata`, then handed to every scorer. */
+  metadata?: Record<string, unknown>;
+  /** The output text under test. */
+  output: string;
+}
 /** One dataset case: an input and its optional gold answer/metadata. */
 interface EvalCase {
   expected?: string;
@@ -628,5 +684,70 @@ declare const scoreSample: (sample: ScorerSample, scorers: ReadonlyArray<Scorer>
  * give each its own thread/key inside `produce` if the producer is stateful.
  * Returns per-case results plus the mean of their averages.
  */
-declare const evaluate: (cases: ReadonlyArray<EvalCase>, produce: (input: string) => Promise<string> | string, scorers: ReadonlyArray<Scorer>) => Promise<EvalResult>;
-export { type AgentHarness, type AgentHarnessOptions, type AgentRunOverrides, type EvalCase, type EvalItemResult, type EvalResult, type EvaluationAttributeValue, type EvaluationMetrics, type EvaluationSpanHandle, type FakeScheduledJob, type FakeSchedulerControls, type FunctionRegistry, type HarnessDispatch, type HarnessMessage, type HarnessThread, type LunoraTestOptions, type RecordEvaluationInput, type ScheduledJobFailure, type ScoreResult, type Scorer, type ScorerSample, type SweepOptions, type TestHarness, type TestIdentity, type TestSubscription, agentHarness, containsScorer, evaluate, evaluationAttributes, exactMatchScorer, finalTurn, keywordScorer, llmScorer, lunoraTest, recordEvaluation, regexScorer, scoreSample, toolCallTurn };
+declare const evaluate: (cases: ReadonlyArray<EvalCase>, produce: (input: string) => Promise<ProducedOutput | string> | ProducedOutput | string, scorers: ReadonlyArray<Scorer>) => Promise<EvalResult>;
+/** Where a retrieval scorer reads its two id lists from. */
+interface RetrievalScorerOptions {
+  /** Metadata key holding the gold relevant ids. Default `"relevant"`. */
+  relevantKey?: string;
+  /** Metadata key holding the run's ranked retrieved ids. Default `"retrieved"`. */
+  retrievedKey?: string;
+}
+/**
+ * Recall@k scorer — what fraction of the gold passages made it into the top
+ * `k`.
+ *
+ * This is the ceiling on everything downstream: a passage retrieval never
+ * returned is one no reranker can promote and no prompt can cite. Omit `k` to
+ * score the whole retrieved list.
+ */
+declare const recallAtK: (k?: number, options?: RetrievalScorerOptions) => Scorer;
+/**
+ * Precision@k scorer — what fraction of the top `k` retrieved passages are
+ * gold.
+ *
+ * This is the counterweight to recall: padding `topK` raises recall for free
+ * while burying the answer in noise the model has to read past, and pay for.
+ * Omit `k` to score the whole retrieved list.
+ */
+declare const precisionAtK: (k?: number, options?: RetrievalScorerOptions) => Scorer;
+/**
+ * Mean Reciprocal Rank scorer — `1 / rank` of the first gold passage: 1 if it
+ * is first, 0.5 if second, 0 if absent.
+ *
+ * This is the metric that notices ordering, which recall cannot. A gold passage
+ * at rank 20 counts the same as rank 1 for recall@20, but only one of those
+ * survives a context-window trim or a model that skims the top of its prompt.
+ */
+declare const mrrScorer: (options?: RetrievalScorerOptions) => Scorer;
+/**
+ * Normalized Discounted Cumulative Gain scorer — relevance discounted
+ * logarithmically by rank, divided by the best achievable arrangement.
+ *
+ * This is the one to gate on when comparing retrieval strategies. Unlike recall
+ * it is sensitive to order, and unlike MRR it credits every gold passage rather
+ * than only the first — so it is the metric that can actually say whether a
+ * reranker or a hybrid leg helped.
+ *
+ * Relevance is binary: an id is gold or it is not, which is what a gold-id set
+ * expresses. Graded relevance would need per-id weights.
+ */
+declare const ndcgAtK: (k?: number, options?: RetrievalScorerOptions) => Scorer;
+/**
+ * Groundedness scorer — does the answer assert only what the retrieved context
+ * supports?
+ *
+ * This is the generation-side counterpart to the metrics above: perfect
+ * retrieval still fails if the model answers from its own weights. It scores
+ * the output against the context via an injected `judge`, the same shape
+ * `llmScorer` takes, so this stays model-agnostic and mockable.
+ *
+ * Reads the context from `metadata.context` by default — the `context` string
+ * `retrieve()` already returns. Fails closed: no context means nothing could
+ * have grounded the answer.
+ */
+declare const groundednessScorer: (options: {
+  contextKey?: string;
+  judge: (prompt: string) => Promise<string>;
+  name?: string;
+}) => Scorer;
+export { type AgentHarness, type AgentHarnessOptions, type AgentRunOverrides, type EvalCase, type EvalItemResult, type EvalResult, type EvaluationAttributeValue, type EvaluationMetrics, type EvaluationSpanHandle, type FakeScheduledJob, type FakeSchedulerControls, type FunctionRegistry, type HarnessDispatch, type HarnessMessage, type HarnessThread, type LunoraTestOptions, type ProducedOutput, type RecordEvaluationInput, type RetrievalScorerOptions, type ScheduledJobFailure, type ScoreResult, type Scorer, type ScorerSample, type SweepOptions, type TestHarness, type TestIdentity, type TestSubscription, agentHarness, containsScorer, evaluate, evaluationAttributes, exactMatchScorer, finalTurn, groundednessScorer, keywordScorer, llmScorer, lunoraTest, mrrScorer, ndcgAtK, precisionAtK, recallAtK, recordEvaluation, regexScorer, scoreSample, toolCallTurn };

package/dist/index.mjs

@@ -1 +1 @@
-import{agentHarness as o,finalTurn as t,toolCallTurn as a}from"./packem_shared/agentHarness-DA9onGVO.mjs";import{evaluationAttributes as c,recordEvaluation as n}from"./packem_shared/evaluationAttributes-BA1uUP-z.mjs";import{lunoraTest as u}from"./packem_shared/lunoraTest-CC-mCvor.mjs";import{containsScorer as m,evaluate as p,exactMatchScorer as s,keywordScorer as f,llmScorer as S,regexScorer as d,scoreSample as v}from"./packem_shared/containsScorer-DDK7WEuL.mjs";import{extractLink as T,listCapturedMail as g,waitForMail as k}from"@lunora/mail/testing";export{o as agentHarness,m as containsScorer,p as evaluate,c as evaluationAttributes,s as exactMatchScorer,T as extractLink,t as finalTurn,f as keywordScorer,g as listCapturedMail,S as llmScorer,u as lunoraTest,n as recordEvaluation,d as regexScorer,v as scoreSample,a as toolCallTurn,k as waitForMail};
+import{agentHarness as o,finalTurn as t,toolCallTurn as a}from"./packem_shared/agentHarness-DA9onGVO.mjs";import{evaluationAttributes as l,recordEvaluation as n}from"./packem_shared/evaluationAttributes-CYLgTP1P.mjs";import{lunoraTest as s}from"./packem_shared/lunoraTest-UV1TxREP.mjs";import{groundednessScorer as p,mrrScorer as u,ndcgAtK as x,precisionAtK as S,recallAtK as f}from"./packem_shared/groundednessScorer-CznJBIjn.mjs";import{containsScorer as g,evaluate as A,exactMatchScorer as v,keywordScorer as K,llmScorer as M,regexScorer as T,scoreSample as k}from"./packem_shared/containsScorer-gv0Sdd86.mjs";import{extractLink as C,listCapturedMail as b,waitForMail as h}from"@lunora/mail/testing";export{o as agentHarness,g as containsScorer,A as evaluate,l as evaluationAttributes,v as exactMatchScorer,C as extractLink,t as finalTurn,p as groundednessScorer,K as keywordScorer,b as listCapturedMail,M as llmScorer,s as lunoraTest,u as mrrScorer,x as ndcgAtK,S as precisionAtK,f as recallAtK,n as recordEvaluation,T as regexScorer,k as scoreSample,a as toolCallTurn,h as waitForMail};

package/dist/packem_shared/containsScorer-gv0Sdd86.mjs

@@ -0,0 +1,2 @@
+import{LunoraError as g}from"@lunora/errors";const l=/^\s*(-?\d+(?:\.\d+)?)/u,i=e=>Number.isFinite(e)?Math.min(1,Math.max(0,e)):0,h=e=>typeof e=="number"?{score:i(e)}:{score:i(e.score),...e.reason===void 0?{}:{reason:e.reason}},u=e=>e.length===0?0:e.reduce((t,o)=>t+o,0)/e.length,y=(e,t={})=>({name:`contains:${e}`,score:({output:o})=>{const a=t.caseSensitive?o:o.toLowerCase(),r=t.caseSensitive?e:e.toLowerCase();return a.includes(r)?1:0}}),v=(e,t="regex")=>({name:t,score:({output:o})=>e.test(o)?1:0}),$=()=>({name:"exact-match",score:({expected:e,output:t})=>t.trim()===e?.trim()?1:0}),L=e=>{if(e.length===0)throw new g("BAD_REQUEST","@lunora/testing: keywordScorer requires at least one keyword");return{name:"keyword-coverage",score:({output:t})=>{const o=t.toLowerCase(),a=e.filter(r=>o.includes(r.toLowerCase())).length;return{reason:`${String(a)}/${String(e.length)} keywords present`,score:a/e.length}}}},S=(e,t)=>[`Rate the ASSISTANT OUTPUT against this criterion: ${e}`,"Respond with a single number from 0 (fails) to 1 (fully meets), then a dash and a one-line reason.",...t.input===void 0?[]:["",`Input: ${t.input}`],...t.expected===void 0?[]:["",`Reference answer: ${t.expected}`],"",`Assistant output: ${t.output}`].join(`
+`),w=e=>{const t=l.exec(e);return{reason:e.trim(),score:t?i(Number(t[1])):0}},b=e=>({name:e.name??"llm-judge",score:async t=>w(await e.judge(S(e.criteria,t)))}),f=async(e,t)=>{const o=await Promise.all(t.map(async n=>({name:n.name,result:h(await n.score(e))}))),a={},r=new Map;for(const{name:n,result:s}of o){const c=r.get(n)??0;r.set(n,c+1),a[c===0?n:`${n}#${String(c+1)}`]=s}return{average:u(o.map(({result:n})=>n.score)),scores:a}},A=async(e,t,o)=>{const a=await Promise.all(e.map(async r=>{const n=await t(r.input),s=typeof n=="string"?n:n.output,c=typeof n=="string"||n.metadata===void 0?r.metadata:{...r.metadata,...n.metadata},m={input:r.input,output:s,...r.expected===void 0?{}:{expected:r.expected},...c===void 0?{}:{metadata:c}},{average:d,scores:p}=await f(m,o);return{average:d,input:r.input,output:s,scores:p}}));return{average:u(a.map(r=>r.average)),items:a}};export{y as containsScorer,A as evaluate,$ as exactMatchScorer,L as keywordScorer,b as llmScorer,w as parseJudgeScore,v as regexScorer,f as scoreSample};

package/dist/packem_shared/evaluationAttributes-CYLgTP1P.mjs

@@ -0,0 +1 @@
+import{LunoraError as a}from"@lunora/errors";const n=/[\w.-]/u,t=e=>{let r="";for(const o of e)r+=n.test(o)?o:"_";return r},s=e=>{if(typeof e.name!="string"||e.name.length===0)throw new TypeError("recordEvaluation requires a non-empty `name`");if(typeof e.score!="number"||!Number.isFinite(e.score))throw new TypeError("recordEvaluation `score` must be a finite number");const r=t(e.name),o={[`gen_ai.evaluation.${r}.score`]:e.score};return e.label!==void 0&&(o[`gen_ai.evaluation.${r}.label`]=e.label),o},c=e=>{try{return s(e)}catch(r){throw r instanceof TypeError?new a("BAD_REQUEST",`@lunora/testing: ${r.message}`):r}},l=e=>{const r=c(e);return e.span?.setAttributes(r),e.metrics?.gauge(`gen_ai.evaluation.${t(e.name)}.score`,e.score,e.label===void 0?void 0:{label:e.label}),r};export{c as evaluationAttributes,l as recordEvaluation};

package/dist/packem_shared/groundednessScorer-CznJBIjn.mjs

@@ -0,0 +1,3 @@
+import{LunoraError as g}from"@lunora/errors";import{parseJudgeScore as h}from"./containsScorer-gv0Sdd86.mjs";const w="retrieved",v="relevant",l=(e,t)=>{const n=e.metadata?.[t];return Array.isArray(n)?n.filter(r=>typeof r=="string"&&r.length>0):[]},s=(e,t)=>{const n=l(e,t?.relevantKey??v);if(n.length!==0)return{relevant:new Set(n),retrieved:l(e,t?.retrievedKey??w)}},a=e=>({reason:`no gold ids under metadata.${e?.relevantKey??v} — cannot score retrieval`,score:0}),c=(e,t)=>{if(e!==void 0&&(!Number.isInteger(e)||e<1))throw new g("BAD_REQUEST",`@lunora/testing: ${t} \`k\` must be a positive integer`)},u=(e,t)=>t===void 0?e.retrieved:e.retrieved.slice(0,t),f=(e,t)=>{const n=new Set;for(const r of e)t.has(r)&&n.add(r);return n.size},S=(e,t)=>{const n=new Set;let r=0;for(const[o,i]of e.entries())t.has(i)&&!n.has(i)&&(n.add(i),r+=1/Math.log2(o+2));return r},m=e=>{let t=0;for(let n=0;n<e;n+=1)t+=1/Math.log2(n+2);return t},p=(e,t)=>(c(e,"recallAtK"),{name:e===void 0?"recall":`recall@${String(e)}`,score:n=>{const r=s(n,t);if(r===void 0)return a(t);const o=f(u(r,e),r.relevant);return{reason:`${String(o)}/${String(r.relevant.size)} gold ids retrieved`,score:o/r.relevant.size}}}),x=(e,t)=>(c(e,"precisionAtK"),{name:e===void 0?"precision":`precision@${String(e)}`,score:n=>{const r=s(n,t);if(r===void 0)return a(t);const o=u(r,e);if(o.length===0)return{reason:"nothing retrieved",score:0};const i=f(o,r.relevant);return{reason:`${String(i)}/${String(o.length)} retrieved ids are gold`,score:i/o.length}}}),y=e=>({name:"mrr",score:t=>{const n=s(t,e);if(n===void 0)return a(e);const r=n.retrieved.findIndex(o=>n.relevant.has(o));return r===-1?{reason:"no gold id retrieved",score:0}:{reason:`first gold id at rank ${String(r+1)}`,score:1/(r+1)}}}),$=(e,t)=>(c(e,"ndcgAtK"),{name:e===void 0?"ndcg":`ndcg@${String(e)}`,score:n=>{const r=s(n,t);if(r===void 0)return a(t);const o=u(r,e);if(o.length===0)return{reason:"nothing retrieved",score:0};const i=S(o,r.relevant),d=m(e===void 0?r.relevant.size:Math.min(r.relevant.size,e));return{reason:`dcg ${i.toFixed(3)} / ideal ${d.toFixed(3)}`,score:d===0?0:i/d}}}),K=e=>{if(typeof e.judge!="function")throw new g("BAD_REQUEST","@lunora/testing: groundednessScorer requires an injected `judge` function");const t=e.contextKey??"context";return{name:e.name??"groundedness",score:async n=>{const r=n.metadata?.[t];if(typeof r!="string"||r.trim().length===0)return{reason:`no retrieved context under metadata.${t}`,score:0};const o=await e.judge(["Rate how well the ASSISTANT ANSWER is supported by the RETRIEVED CONTEXT below.","Score 1 if every claim in the answer is supported by the context, 0 if the answer asserts","anything the context does not support. Judge support only — do NOT reward correctness","the context does not contain.","Respond with a single number from 0 to 1, then a dash and a one-line reason.","",`Retrieved context:
+${r}`,"",`Assistant answer: ${n.output}`].join(`
+`));return h(o)}}};export{K as groundednessScorer,y as mrrScorer,$ as ndcgAtK,x as precisionAtK,p as recallAtK};

package/dist/packem_shared/lunoraTest-UV1TxREP.mjs

@@ -0,0 +1 @@
+import{LunoraError as _}from"@lunora/errors";import{runShardMigrations as O,createShardCtxDb as z,RLS_UNWRAP_SYMBOL as Z}from"@lunora/shard-engine";import{evaluationAttributes as tt}from"./evaluationAttributes-CYLgTP1P.mjs";import{MAX_RETRY_ATTEMPTS as et,RETRY_BASE_DELAY_MS as nt}from"@lunora/scheduler";import{DatabaseSync as ot}from"node:sqlite";const rt=(r,c,s,u,g)=>{let p=g,y=1;const d=new Map,f=[],h=(e,l,a={})=>{const t=`fake-job-${String(y)}`;return y+=1,d.set(t,{args:a,enqueuedAt:p,functionPath:l,id:t,scheduledFor:e}),t},b=e=>typeof e=="string"?e:e.name??e.binding??"",E={cancel:e=>{const l=d.has(e);return d.delete(e),Promise.resolve({cancelled:l})},get:e=>Promise.resolve(d.get(e)??null),list:()=>Promise.resolve([...d.values()]),runAfter:(e,l,a)=>{const t=h(p+e,b(l),a);return Promise.resolve(t)},runAt:(e,l,a)=>{const t=h(e,b(l),a);return Promise.resolve(t)}},k=async e=>{d.delete(e.id);const a=u().get(e.functionPath);if(a===void 0){console.warn(`[fake-scheduler] unknown functionPath "${e.functionPath}" — job ${e.id} dropped`);return}if(a.kind==="mutation"||a.kind==="action"){const t=r(),n=a.kind==="action"?s():c();await t(a.kind,a,n,e.args)}else console.warn(`[fake-scheduler] functionPath "${e.functionPath}" is a ${a.kind} — only mutations and actions can be scheduled; job ${e.id} dropped`)},M=async e=>{const l=[...d.values()].filter(n=>n.scheduledFor<=e).toSorted((n,v)=>n.scheduledFor-v.scheduledFor),a=[];let t=0;for(const n of l)if(d.has(n.id)){t+=1;try{await k(n)}catch(v){const T=(n.attempts??0)+1;if(T>et){const N={args:n.args,error:v,functionPath:n.functionPath,id:n.id};a.push(N),f.push(N)}else{const N=nt*2**(T-1);d.set(n.id,{...n,attempts:T,scheduledFor:p+N})}}}return{executed:t,failed:a}},C=async(e,l)=>{const{executed:a,failed:t}=await M(e);if(t.length>0&&(l?.throwOnError??!0)){const[n]=t;throw t.length===1&&n!==void 0?n.error:new AggregateError(t.map(v=>v.error),`${String(t.length)} scheduled jobs failed: ${t.map(v=>v.functionPath).join(", ")}`)}return a};return{controls:{advance:(e,l)=>(p+=e,C(p,l)),failures:()=>[...f],list:()=>[...d.values()],runPending:e=>C(Number.POSITIVE_INFINITY,e)},scheduler:E}},st=()=>{const r=new ot(":memory:"),c=u=>({one(){if(u.length!==1)throw new _("INTERNAL",`expected exactly one row, received ${String(u.length)}`);const[g]=u;return g},[Symbol.iterator](){return u[Symbol.iterator]()},toArray(){return u}});return{close:()=>{r.close()},sql:{exec:(u,...g)=>{const y=r.prepare(u).all(...g);return c(y)}}}},q=r=>{if(typeof r!="object"||r===null)return;const{kind:c}=r;if(c==="query"||c==="mutation"||c==="action")return c},it=r=>typeof r=="object"&&r!==null&&r.visibility==="internal"?"internal":"public",Q=r=>{throw new _("INTERNAL",`ctx.${r} is not available in the in-memory @lunora/testing harness (v1)`)},w=r=>new Proxy((...c)=>Q(r),{apply:()=>Q(r),get:()=>Q(r)}),K={spanId:"0000000000000001",traceId:"00000000000000000000000000000001"},ct={addEvent:()=>{},addLink:()=>{},recordEvaluation:()=>{},recordException:()=>{},setAttribute:()=>{},setAttributes:()=>{},spanContext:()=>K},at=()=>{const r={attributes:{},events:[],links:[]},c={addEvent:(s,u)=>{r.events.push({...u===void 0?{}:{attributes:{...u}},name:s})},addLink:s=>{r.links.push({spanId:s.spanId,traceId:s.traceId})},recordEvaluation:s=>{Object.assign(r.attributes,tt(s))},recordException:s=>{const u={"exception.message":s instanceof Error?s.message:String(s),"exception.type":s instanceof Error?s.constructor.name:"Error"};s instanceof Error&&s.stack!==void 0&&(u["exception.stacktrace"]=s.stack),c.addEvent("exception",u)},setAttribute:(s,u)=>{r.attributes[s]=u},setAttributes:s=>{Object.assign(r.attributes,s)},spanContext:()=>K};return{handle:c,recorded:r}},D=async(r,c)=>await c(D,ct),Y={count:()=>{},gauge:()=>{},record:()=>{}},j={debug:()=>{},error:()=>{},event:()=>{},fatal:()=>{},info:()=>{},log:()=>{},trace:()=>{},warn:()=>{},with:()=>j},ut=(r,c,s)=>(g,p)=>{let y=!1;const d=[];let f,h,b=0,E=0;const k=()=>q(g)?r("query",g,c,p,!1):Promise.resolve(g(c)),M=(t,n)=>{if(t<E)return;E=t;const v={done:!1,value:n};if(d.length===0)f=v,h=void 0;else{f=void 0,h=void 0;for(const T of d.splice(0))T.resolve(v)}},C=(t,n)=>{if(!(t<E))if(E=t,d.length===0)h={error:n},f=void 0;else{f=void 0,h=void 0;for(const v of d.splice(0))v.reject(n)}},R=t=>n=>{M(t,n)},e=t=>n=>{C(t,n)},l=()=>{if(y)return;b+=1;const t=b;k().then(R(t)).catch(e(t))};s.add(l);const a={[Symbol.asyncIterator](){return a},next:()=>{if(y)return Promise.resolve({done:!0,value:void 0});if(E===b){if(h!==void 0){const{error:t}=h;return h=void 0,Promise.reject(t)}if(f!==void 0){const t=f;return f=void 0,Promise.resolve(t)}}return E<b?new Promise((t,n)=>{d.push({reject:n,resolve:t})}):k().then(t=>{if(h!==void 0){const{error:n}=h;throw h=void 0,n}if(f!==void 0){const n=f;return f=void 0,n}return{done:!1,value:t}})},return:()=>{y=!0,s.delete(l);for(const t of d.splice(0))t.resolve({done:!0,value:void 0});return Promise.resolve({done:!0,value:void 0})}};return k().then(R(0)).catch(e(0)),a},pt=(r,c)=>{const{close:s,sql:u}=st(),g=r;O(u,g);const p=z({enforceRls:c?.enforceRls??!0,schema:g,sql:u}),y=p[Z]??p,d=m=>{u.exec.call(u,m)};let f=Promise.resolve();const h=m=>{const P=async()=>{d("BEGIN");try{const S=await m();return d("COMMIT"),S}catch(S){try{d("ROLLBACK")}catch{}throw S}},x=f.then(P);return f=x.then(()=>{},()=>{}),x};let b=!1;const E=()=>{b||(b=!0,s())},k=new Map(Object.entries(c?.functions??{}).map(([m,P])=>[m,P])),M=new Set,C=()=>{for(const m of M)m()},R=m=>(C(),m);let e,l,a;const t=c?.now??Date.now(),n=at(),v=(m,P)=>{if(m===void 0)throw new _("INTERNAL",`[fake-scheduler] ${P} not yet available — scheduler.advance called before harness construction completed`);return m},{controls:T,scheduler:N}=rt(()=>v(e,"dispatch"),()=>v(l,"mutationContext"),()=>v(a,"actionContext"),()=>k,t),B=m=>{const P={getIdentity:()=>Promise.resolve(m??null),userId:m?.userId??null},x={auth:P,db:p,env:c?.env,log:j,metrics:Y,now:t,span:n.handle,trace:D,runQuery:((o,i)=>I("query",o,x,i)),secrets:w("secrets"),storage:w("storage"),vectors:w("vectors")},S={auth:P,db:p,env:c?.env,log:j,metrics:Y,now:t,span:n.handle,trace:D,runMutation:((o,i)=>I("mutation",o,S,i)),runQuery:((o,i)=>I("query",o,x,i)),scheduler:N,secrets:w("secrets"),storage:w("storage"),vectors:w("vectors"),workflows:w("workflows")};l??=S;const U={...S,db:y},F={auth:P,db:p,env:c?.env,fetch:c?.fetch??w("fetch"),log:j,metrics:Y,now:t,span:n.handle,trace:D,runAction:((o,i)=>I("action",o,F,i)),runMutation:((o,i)=>I("mutation",o,S,i)),runQuery:((o,i)=>I("query",o,x,i)),scheduler:N,secrets:w("secrets"),storage:w("storage"),vectors:w("vectors"),workflows:w("workflows")};a??=F;const $=(o,i,A,L,J)=>{const H=q(i);if(H!==o)throw new _("INTERNAL",`expected a registered ${o}, received a ${H??"non-function"} reference`);if(!J&&it(i)==="internal")throw new _("INTERNAL",`This ${o} is an internal function — it is unreachable from the external RPC boundary in production. Call it through ctx.run${o.charAt(0).toUpperCase()}${o.slice(1)} from another function instead.`);return Promise.resolve(i.handler(A,L??{}))},I=(o,i,A,L)=>$(o,i,A,L,!0);e??=(o,i,A,L)=>o==="mutation"?h(()=>$("mutation",i,A,L,!0)).then(R):I("action",i,A,L);const V=((o,i)=>q(o)?$("query",o,x,i,!1):Promise.resolve(o(x))),W=((o,i)=>{const A=q(o)?()=>$("mutation",o,S,i,!1):()=>o(S);return h(A).then(R)}),X=((o,i)=>q(o)?$("action",o,F,i,!1):Promise.resolve(o(F))),G=ut($,x,M);return{action:X,close:E,mutation:W,query:V,run:o=>h(()=>o(U)).then(R),scheduler:T,subscribe:G,wideEvent:()=>n.recorded,withIdentity:o=>B(o)}};return B(null)};export{pt as lunoraTest};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lunora/testing",
-  "version": "1.0.0-alpha.111",
+  "version": "1.0.0-alpha.113",
   "description": "Testing toolkit for Lunora: an in-memory harness for queries, mutations, and actions",
   "keywords": [
     "cloudflare",
@@ -50,11 +50,12 @@
     "access": "public"
   },
   "dependencies": {
-    "@lunora/agent": "1.0.0-alpha.57",
+    "@lunora/agent": "1.0.0-alpha.59",
     "@lunora/errors": "1.0.0-alpha.22",
-    "@lunora/mail": "1.0.0-alpha.50",
-    "@lunora/server": "1.0.0-alpha.76",
-    "@lunora/shard-engine": "1.0.0-alpha.29"
+    "@lunora/mail": "1.0.0-alpha.51",
+    "@lunora/scheduler": "1.0.0-alpha.34",
+    "@lunora/server": "1.0.0-alpha.78",
+    "@lunora/shard-engine": "1.0.0-alpha.31"
   },
   "peerDependencies": {
     "@playwright/test": "^1.61.1"

package/dist/packem_shared/containsScorer-DDK7WEuL.mjs

@@ -1,2 +0,0 @@
-import{LunoraError as d}from"@lunora/errors";const g=/^\s*(-?\d+(?:\.\d+)?)/u,i=e=>Number.isFinite(e)?Math.min(1,Math.max(0,e)):0,p=e=>typeof e=="number"?{score:i(e)}:{score:i(e.score),...e.reason===void 0?{}:{reason:e.reason}},u=e=>e.length===0?0:e.reduce((t,n)=>t+n,0)/e.length,x=(e,t={})=>({name:`contains:${e}`,score:({output:n})=>{const a=t.caseSensitive?n:n.toLowerCase(),r=t.caseSensitive?e:e.toLowerCase();return a.includes(r)?1:0}}),f=(e,t="regex")=>({name:t,score:({output:n})=>e.test(n)?1:0}),v=()=>({name:"exact-match",score:({expected:e,output:t})=>t.trim()===e?.trim()?1:0}),y=e=>{if(e.length===0)throw new d("BAD_REQUEST","@lunora/testing: keywordScorer requires at least one keyword");return{name:"keyword-coverage",score:({output:t})=>{const n=t.toLowerCase(),a=e.filter(r=>n.includes(r.toLowerCase())).length;return{reason:`${String(a)}/${String(e.length)} keywords present`,score:a/e.length}}}},l=(e,t)=>[`Rate the ASSISTANT OUTPUT against this criterion: ${e}`,"Respond with a single number from 0 (fails) to 1 (fully meets), then a dash and a one-line reason.",...t.input===void 0?[]:["",`Input: ${t.input}`],...t.expected===void 0?[]:["",`Reference answer: ${t.expected}`],"",`Assistant output: ${t.output}`].join(`
-`),h=e=>{const t=g.exec(e);return{reason:e.trim(),score:t?i(Number(t[1])):0}},$=e=>({name:e.name??"llm-judge",score:async t=>h(await e.judge(l(e.criteria,t)))}),S=async(e,t)=>{const n=await Promise.all(t.map(async o=>({name:o.name,result:p(await o.score(e))}))),a={},r=new Map;for(const{name:o,result:s}of n){const c=r.get(o)??0;r.set(o,c+1),a[c===0?o:`${o}#${String(c+1)}`]=s}return{average:u(n.map(({result:o})=>o.score)),scores:a}},L=async(e,t,n)=>{const a=await Promise.all(e.map(async r=>{const o=await t(r.input),s={input:r.input,output:o,...r.expected===void 0?{}:{expected:r.expected},...r.metadata===void 0?{}:{metadata:r.metadata}},{average:c,scores:m}=await S(s,n);return{average:c,input:r.input,output:o,scores:m}}));return{average:u(a.map(r=>r.average)),items:a}};export{x as containsScorer,L as evaluate,v as exactMatchScorer,y as keywordScorer,$ as llmScorer,f as regexScorer,S as scoreSample};

package/dist/packem_shared/evaluationAttributes-BA1uUP-z.mjs

@@ -1 +0,0 @@
-import{LunoraError as a}from"@lunora/errors";const n=/[\w.-]/u,t=e=>{let o="";for(const r of e)o+=n.test(r)?r:"_";return o},s=e=>{if(typeof e.name!="string"||e.name.length===0)throw new a("BAD_REQUEST","@lunora/testing: recordEvaluation requires a non-empty `name`");if(typeof e.score!="number"||!Number.isFinite(e.score))throw new a("BAD_REQUEST","@lunora/testing: recordEvaluation `score` must be a finite number");const o=t(e.name),r={[`gen_ai.evaluation.${o}.score`]:e.score};return e.label!==void 0&&(r[`gen_ai.evaluation.${o}.label`]=e.label),r},c=e=>{const o=s(e);return e.span?.setAttributes(o),e.metrics?.gauge(`gen_ai.evaluation.${t(e.name)}.score`,e.score,e.label===void 0?void 0:{label:e.label}),o};export{s as evaluationAttributes,c as recordEvaluation};

package/dist/packem_shared/lunoraTest-CC-mCvor.mjs

@@ -1 +0,0 @@
-import{LunoraError as M}from"@lunora/errors";import{runShardMigrations as X,createShardCtxDb as z,RLS_UNWRAP_SYMBOL as Z}from"@lunora/shard-engine";import{evaluationAttributes as tt}from"./evaluationAttributes-BA1uUP-z.mjs";import{DatabaseSync as et}from"node:sqlite";const nt=(o,c,s,u,g)=>{let p=g,y=1;const d=new Map,f=[],h=(e,l,a={})=>{const t=`fake-job-${String(y)}`;return y+=1,d.set(t,{args:a,enqueuedAt:p,functionPath:l,id:t,scheduledFor:e}),t},b=e=>typeof e=="string"?e:e.name??e.binding??"",P={cancel:e=>{const l=d.has(e);return d.delete(e),Promise.resolve({cancelled:l})},get:e=>Promise.resolve(d.get(e)??null),list:()=>Promise.resolve([...d.values()]),runAfter:(e,l,a)=>{const t=h(p+e,b(l),a);return Promise.resolve(t)},runAt:(e,l,a)=>{const t=h(e,b(l),a);return Promise.resolve(t)}},A=async e=>{d.delete(e.id);const a=u().get(e.functionPath);if(a===void 0){console.warn(`[fake-scheduler] unknown functionPath "${e.functionPath}" — job ${e.id} dropped`);return}if(a.kind==="mutation"||a.kind==="action"){const t=o(),r=a.kind==="action"?s():c();await t(a.kind,a,r,e.args)}else console.warn(`[fake-scheduler] functionPath "${e.functionPath}" is a ${a.kind} — only mutations and actions can be scheduled; job ${e.id} dropped`)},I=async e=>{const l=[...d.values()].filter(r=>r.scheduledFor<=e).toSorted((r,v)=>r.scheduledFor-v.scheduledFor),a=[];let t=0;for(const r of l)if(d.has(r.id)){t+=1;try{await A(r)}catch(v){const $={args:r.args,error:v,functionPath:r.functionPath,id:r.id};a.push($),f.push($)}}return{executed:t,failed:a}},C=async(e,l)=>{const{executed:a,failed:t}=await I(e);if(t.length>0&&(l?.throwOnError??!0)){const[r]=t;throw t.length===1&&r!==void 0?r.error:new AggregateError(t.map(v=>v.error),`${String(t.length)} scheduled jobs failed: ${t.map(v=>v.functionPath).join(", ")}`)}return a};return{controls:{advance:(e,l)=>(p+=e,C(p,l)),failures:()=>[...f],list:()=>[...d.values()],runPending:e=>C(Number.POSITIVE_INFINITY,e)},scheduler:P}},rt=()=>{const o=new et(":memory:"),c=u=>({one(){if(u.length!==1)throw new M("INTERNAL",`expected exactly one row, received ${String(u.length)}`);const[g]=u;return g},[Symbol.iterator](){return u[Symbol.iterator]()},toArray(){return u}});return{close:()=>{o.close()},sql:{exec:(u,...g)=>{const y=o.prepare(u).all(...g);return c(y)}}}},L=o=>{if(typeof o!="object"||o===null)return;const{kind:c}=o;if(c==="query"||c==="mutation"||c==="action")return c},ot=o=>typeof o=="object"&&o!==null&&o.visibility==="internal"?"internal":"public",_=o=>{throw new M("INTERNAL",`ctx.${o} is not available in the in-memory @lunora/testing harness (v1)`)},w=o=>new Proxy((...c)=>_(o),{apply:()=>_(o),get:()=>_(o)}),U={spanId:"0000000000000001",traceId:"00000000000000000000000000000001"},st={addEvent:()=>{},addLink:()=>{},recordEvaluation:()=>{},recordException:()=>{},setAttribute:()=>{},setAttributes:()=>{},spanContext:()=>U},it=()=>{const o={attributes:{},events:[],links:[]},c={addEvent:(s,u)=>{o.events.push({...u===void 0?{}:{attributes:{...u}},name:s})},addLink:s=>{o.links.push({spanId:s.spanId,traceId:s.traceId})},recordEvaluation:s=>{Object.assign(o.attributes,tt(s))},recordException:s=>{const u={"exception.message":s instanceof Error?s.message:String(s),"exception.type":s instanceof Error?s.constructor.name:"Error"};s instanceof Error&&s.stack!==void 0&&(u["exception.stacktrace"]=s.stack),c.addEvent("exception",u)},setAttribute:(s,u)=>{o.attributes[s]=u},setAttributes:s=>{Object.assign(o.attributes,s)},spanContext:()=>U};return{handle:c,recorded:o}},j=async(o,c)=>await c(j,st),Q={count:()=>{},gauge:()=>{},record:()=>{}},D={debug:()=>{},error:()=>{},event:()=>{},fatal:()=>{},info:()=>{},log:()=>{},trace:()=>{},warn:()=>{},with:()=>D},ct=(o,c,s)=>(g,p)=>{let y=!1;const d=[];let f,h,b=0,P=0;const A=()=>L(g)?o("query",g,c,p,!1):Promise.resolve(g(c)),I=(t,r)=>{if(t<P)return;P=t;const v={done:!1,value:r};if(d.length===0)f=v,h=void 0;else{f=void 0,h=void 0;for(const $ of d.splice(0))$.resolve(v)}},C=(t,r)=>{if(!(t<P))if(P=t,d.length===0)h={error:r},f=void 0;else{f=void 0,h=void 0;for(const v of d.splice(0))v.reject(r)}},R=t=>r=>{I(t,r)},e=t=>r=>{C(t,r)},l=()=>{if(y)return;b+=1;const t=b;A().then(R(t)).catch(e(t))};s.add(l);const a={[Symbol.asyncIterator](){return a},next:()=>{if(y)return Promise.resolve({done:!0,value:void 0});if(P===b){if(h!==void 0){const{error:t}=h;return h=void 0,Promise.reject(t)}if(f!==void 0){const t=f;return f=void 0,Promise.resolve(t)}}return P<b?new Promise((t,r)=>{d.push({reject:r,resolve:t})}):A().then(t=>{if(h!==void 0){const{error:r}=h;throw h=void 0,r}if(f!==void 0){const r=f;return f=void 0,r}return{done:!1,value:t}})},return:()=>{y=!0,s.delete(l);for(const t of d.splice(0))t.resolve({done:!0,value:void 0});return Promise.resolve({done:!0,value:void 0})}};return A().then(R(0)).catch(e(0)),a},ft=(o,c)=>{const{close:s,sql:u}=rt(),g=o;X(u,g);const p=z({enforceRls:c?.enforceRls??!0,schema:g,sql:u}),y=p[Z]??p,d=m=>{u.exec.call(u,m)};let f=Promise.resolve();const h=m=>{const x=async()=>{d("BEGIN");try{const S=await m();return d("COMMIT"),S}catch(S){try{d("ROLLBACK")}catch{}throw S}},E=f.then(x);return f=E.then(()=>{},()=>{}),E};let b=!1;const P=()=>{b||(b=!0,s())},A=new Map(Object.entries(c?.functions??{}).map(([m,x])=>[m,x])),I=new Set,C=()=>{for(const m of I)m()},R=m=>(C(),m);let e,l,a;const t=c?.now??Date.now(),r=it(),v=(m,x)=>{if(m===void 0)throw new M("INTERNAL",`[fake-scheduler] ${x} not yet available — scheduler.advance called before harness construction completed`);return m},{controls:$,scheduler:B}=nt(()=>v(e,"dispatch"),()=>v(l,"mutationContext"),()=>v(a,"actionContext"),()=>A,t),H=m=>{const x={getIdentity:()=>Promise.resolve(m??null),userId:m?.userId??null},E={auth:x,db:p,env:c?.env,log:D,metrics:Q,now:t,span:r.handle,trace:j,runQuery:((n,i)=>N("query",n,E,i)),secrets:w("secrets"),storage:w("storage"),vectors:w("vectors")},S={auth:x,db:p,env:c?.env,log:D,metrics:Q,now:t,span:r.handle,trace:j,runMutation:((n,i)=>N("mutation",n,S,i)),runQuery:((n,i)=>N("query",n,E,i)),scheduler:B,secrets:w("secrets"),storage:w("storage"),vectors:w("vectors"),workflows:w("workflows")};l??=S;const V={...S,db:y},F={auth:x,db:p,env:c?.env,fetch:c?.fetch??w("fetch"),log:D,metrics:Q,now:t,span:r.handle,trace:j,runAction:((n,i)=>N("action",n,F,i)),runMutation:((n,i)=>N("mutation",n,S,i)),runQuery:((n,i)=>N("query",n,E,i)),scheduler:B,secrets:w("secrets"),storage:w("storage"),vectors:w("vectors"),workflows:w("workflows")};a??=F;const T=(n,i,k,q,O)=>{const K=L(i);if(K!==n)throw new M("INTERNAL",`expected a registered ${n}, received a ${K??"non-function"} reference`);if(!O&&ot(i)==="internal")throw new M("INTERNAL",`This ${n} is an internal function — it is unreachable from the external RPC boundary in production. Call it through ctx.run${n.charAt(0).toUpperCase()}${n.slice(1)} from another function instead.`);return Promise.resolve(i.handler(k,q??{}))},N=(n,i,k,q)=>T(n,i,k,q,!0);e??=(n,i,k,q)=>n==="mutation"?h(()=>T("mutation",i,k,q,!0)).then(R):N("action",i,k,q);const W=((n,i)=>L(n)?T("query",n,E,i,!1):Promise.resolve(n(E))),Y=((n,i)=>{const k=L(n)?()=>T("mutation",n,S,i,!1):()=>n(S);return h(k).then(R)}),G=((n,i)=>L(n)?T("action",n,F,i,!1):Promise.resolve(n(F))),J=ct(T,E,I);return{action:G,close:P,mutation:Y,query:W,run:n=>h(()=>n(V)).then(R),scheduler:$,subscribe:J,wideEvent:()=>r.recorded,withIdentity:n=>H(n)}};return H(null)};export{ft as lunoraTest};