@sanity/ailf 2.7.1 → 2.9.0

package/dist/_vendor/ailf-core/types/index.d.ts

@@ -9,9 +9,11 @@
  * Ports & Adapters migration (Phase 0c). The original file is now a
  * re-export barrel that preserves backward compatibility.
  */
-import type { DocumentRef as _DocumentRef, EvalMode } from "../../ailf-shared/index.d.ts";
+import type { DocumentRef as _DocumentRef, EvalMode, RunContext } from "../../ailf-shared/index.d.ts";
+import type { ArtifactType } from "../artifact-registry.js";
+import type { AssociationValues, RunId } from "./branded-ids.js";
 export type { ActualScoreEntry, ComponentResult, TestResult, UrlMetadata, } from "./scoring-input.js";
-export type { DocumentRef } from "../../ailf-shared/index.d.ts";
+export type { DocumentRef, RunContext, RunTrigger } from "../../ailf-shared/index.d.ts";
 export type { StoredBaseline, StoredReport, StoredRun, StoredTaskResult, StoredTrace, SchemaVersioned, } from "./storage-schema.js";
 export { CURRENT_SCHEMA_VERSION, isSchemaVersioned, migrateDocument, } from "./storage-schema.js";
 export type { AssertionRegistration, FixtureResolverRegistration, ModeBase, ModeRegistration, PluginManifest, PluginRegistry, PresetDefinition, ReportSinkRegistration, RubricTemplateRegistration, } from "./plugin-registry.js";
@@ -21,8 +23,8 @@ export { evalModeType } from "./eval-mode-config.js";
 export type { DependencyEdge, ResolvedFixture, TaskGraph, TaskNode, } from "./task-graph.js";
 export type { VariableDeclaration, VariableEnvelope, VariableProvenance, VariableSource, } from "./variable-envelope.js";
 export type { EvalTrace, ToolCallCategory, ToolCallRecord, TraceEvent, TraceSpan, TraceTokenUsage, } from "./trace.js";
-export type { ArtifactId, Brand, Err, FixtureId, IdValidationError, NewReportId, Ok, ProviderId, PromptId, Result, ResultId, RubricId, RunFingerprint, RunId, SuiteId, TaskId, TaskSlug, TraceId, } from "./branded-ids.js";
-export { err, fixtureId, ok, providerId, resultId, runId, suiteId, taskId, traceId, } from "./branded-ids.js";
+export type { ArtifactId, AssociationAxis, AssociationValues, Brand, EntryKey, Err, FixtureId, IdValidationError, NewReportId, Ok, ProviderId, PromptId, Result, ResultId, RubricId, RunFingerprint, RunId, SuiteId, TaskId, TaskSlug, TraceId, } from "./branded-ids.js";
+export { err, fixtureId, generateRunId, ok, providerId, resultId, runId, suiteId, taskId, traceId, } from "./branded-ids.js";
 export type { AgentHarnessTaskDefinition, CustomTaskDefinition, GeneralizedAssertionDefinition, GeneralizedDocRef, GeneralizedTaskDefinition, GeneralizedTemplatedAssertion, GeneralizedValueAssertion, IdDocRef, KnowledgeProbeTaskDefinition, LiteracyTaskDefinition, MCPServerTaskDefinition, PathDocRef, PerspectiveDocRef, RubricRef, SlugDocRef, TaskCommonFields, TaskDifficulty, TaskOptions, TaskProviderConfig, TaskStatus, } from "./generalized-task.js";
 type DocumentRef = _DocumentRef;
 /** Aggregated retrieval metrics for a feature area */
@@ -536,6 +538,19 @@ export interface PromptfooUrlEntry {
  * only matter during a single pipeline execution.
  */
 export interface PipelineState {
+    /**
+     * Artifact refs produced by upstream steps during the run.
+     * Populated incrementally (CalculateScoresStep writes testOutputs; future
+     * steps will write renderedPrompts, traces, etc.). Read by FinalizeRunStep
+     * to build the `RunManifest.artifacts` catalog.
+     */
+    artifactRefs?: Partial<ArtifactManifest>;
+    /**
+     * The run manifest, finalized and written to `runs/{runId}/manifest.json`.
+     * Populated by FinalizeRunStep. Consumed by PublishReportStep to snapshot
+     * the `artifacts` slice into `Report.artifactManifest` (D0032).
+     */
+    runManifest?: RunManifest;
     /** Report ID generated by PublishReportStep, consumed by CallbackStep + orchestrator job update */
     reportId?: string;
     /** Eval fingerprint computed by RunEvalStep, consumed by PublishReportStep */
@@ -887,6 +902,142 @@ export interface ScoreSummary {
     };
     timestamp: string;
 }
+/**
+ * The subset of `ScoreSummary` the `compare()` primitive reads and the
+ * only fields consumers of `ComparisonReport.baseline` / `.experiment`
+ * access at runtime. Both the full pipeline `ScoreSummary` and the slim
+ * `ReportSummary` (W0051) carry these fields unchanged, so stored Reports
+ * can participate in auto-compare without re-hydrating prose fields.
+ */
+export type ComparableSummary = Pick<ScoreSummary, "overall" | "perModel" | "scores">;
+/**
+ * Slim pointer to a single low-scoring grader judgment. Replaces the full
+ * `StoredJudgment` inlined on pre-W0051 reports. The `id` field IS the
+ * `graderJudgments` manifest entry key — Studio looks up `reasonPreview`
+ * there for list rendering and hydrates the full reason on drill-down.
+ *
+ * `reasonPreview` is ALSO carried inline as a graceful-degradation fallback
+ * for runs whose `artifactManifest` is empty (offline dev, failed GCS
+ * upload, cache-skip before manifest aggregation lands). At ~280 chars per
+ * judgment × the 50-judgment cap ≈ 14 KB on the Report — tiny against the
+ * 500 KB budget. Studio's dispatch prefers the manifest-entry preview when
+ * present so GCS drill-down still renders the hydrated-on-demand copy.
+ *
+ * `graderId` replaces the historical `dimension` field at the slim-shape
+ * boundary to match D0033 axis naming; runtime value is identical.
+ */
+export interface SlimJudgmentRef {
+    /** Manifest entry key = `formatEntryKey({mode, task, model, grader})` for `graderJudgments`. */
+    id: string;
+    taskId: string;
+    modelId: string;
+    /** Rubric dimension name at runtime (what `GraderJudgment.dimension` carries). */
+    graderId: string;
+    /**
+     * Alias of `graderId` — carried for pre-W0051 Studio readers that still
+     * access `.dimension` (JudgmentList, judgment-formatting, et al). Remove
+     * in Slice 6 when those consumers migrate to `graderId`. Identical
+     * runtime value; no additional ambiguity.
+     */
+    dimension: string;
+    /** Normalized 0–100 score. */
+    score: number;
+    /**
+     * Truncated grader reason (≤280 chars). Inline fallback used when the
+     * manifest entry's preview is unavailable; the authoritative full reason
+     * lives in the `graderJudgments` external artifact.
+     */
+    reasonPreview?: string;
+    /**
+     * Alias of `reasonPreview` — legacy Studio renderers read `.reason`;
+     * expose the same truncated text under both names for the compat window.
+     * Remove alongside `dimension` in Slice 6.
+     */
+    reason?: string;
+}
+/**
+ * Slim failure-mode entry on the Report summary. One per classified
+ * `FailureModeType` bucket; `id` is the `failureModes` manifest entry key
+ * so drill-down can fetch the full category payload.
+ */
+export interface SlimFailureModeTopTitle {
+    /** Manifest entry key = `formatEntryKey({mode, category})` for `failureModes`. */
+    id: string;
+    category: string;
+    severity: "low" | "medium" | "high" | "critical";
+    title: string;
+    count: number;
+}
+/** Counts + top-N per-category summary on the Report. */
+export interface SlimFailureModesSummary {
+    /** Count by FailureModeType-ish id. */
+    counts: Record<string, number>;
+    /** Top-N categories by count, descending. */
+    topTitles: SlimFailureModeTopTitle[];
+    /** Total classified judgments across all categories. */
+    totalJudgments: number;
+    /** Percentage of judgments that landed in a non-unclassified bucket. */
+    classificationRate: number;
+}
+/**
+ * Slim gap pointer. The `gapReport` artifact is bulk (axes: `{run}`) so
+ * there is no per-gap manifest entry to point at — `id` is a stable
+ * synthetic composite so the UI can deduplicate and deep-link. Drill-down
+ * reads the full `gapReport` artifact and filters by id.
+ */
+export interface SlimRecommendationGap {
+    /** Synthetic id: `${area}--${failureMode}`, kebab-safe. */
+    id: string;
+    area: string;
+    title: string;
+    /** Priority bucketing for triage UI ordering. */
+    priority: number;
+}
+/** Counts + top-3 summary on the Report. */
+export interface SlimRecommendations {
+    /** Count of gaps by area. */
+    counts: Record<string, number>;
+    /** Top-3 gaps by priority, descending. */
+    top3: SlimRecommendationGap[];
+    /** Total actionable gaps identified (sum of counts). */
+    totalGaps: number;
+    /** Aggregate estimated lift (matches `GapAnalysisReport.totalPotentialLift`). */
+    totalPotentialLift: number;
+}
+/**
+ * Slim per-feature agent-behavior summary. Full `searchQueries` and
+ * `docSlugsVisited` arrays move to `traces` NDJSON; the Report keeps only
+ * counts + first-N samples for triage preview. The `firstN` cap is the
+ * producer's choice (we default to 5).
+ */
+export interface SlimAgentBehaviorFeature {
+    feature: string;
+    avgDocPagesVisited: number;
+    avgNetworkTimeMs: number;
+    avgSearchesPerformed: number;
+    tasksWithBehaviorData: number;
+    externalDomains: string[];
+    /** Distinct count of unique search queries across tasks in this feature. */
+    searchQueriesCount: number;
+    /** First-N unique search queries (bounded samples). */
+    searchQueriesSample: string[];
+    /** Distinct count of unique doc slugs visited. */
+    docSlugsVisitedCount: number;
+    /** First-N unique doc slugs (bounded samples). */
+    docSlugsVisitedSample: string[];
+}
+/**
+ * Slim `summary` field on a published `Report`. Structurally
+ * `Omit<ScoreSummary, slimmed-fields> & slim-replacements` — every
+ * pipeline-produced field survives except the four prose/array fields
+ * W0051 moves to external artifacts.
+ */
+export type ReportSummary = Omit<ScoreSummary, "agentBehavior" | "failureModes" | "lowScoringJudgments" | "recommendations"> & {
+    agentBehavior?: SlimAgentBehaviorFeature[];
+    failureModes?: SlimFailureModesSummary;
+    lowScoringJudgments?: SlimJudgmentRef[];
+    recommendations?: SlimRecommendations;
+};
 /** Result of a single pipeline step */
 export type StepResult = {
     status: "failed";
@@ -1060,8 +1211,8 @@ export interface ComparisonReport {
     areas: AreaDelta[];
     /** Per-document attribution (when changed docs are known) */
     attribution?: AttributionReport;
-    /** The "before" or "control" summary */
-    baseline: ScoreSummary;
+    /** The "before" or "control" summary (narrowed in W0051 so slim Reports compare) */
+    baseline: ComparableSummary;
     /** Aggregate deltas */
     deltas: {
         /** Overall score delta (experiment.avgScore − baseline.avgScore) */
@@ -1080,8 +1231,8 @@ export interface ComparisonReport {
             modelId: string;
         }[];
     };
-    /** The "after" or "treatment" summary */
-    experiment: ScoreSummary;
+    /** The "after" or "treatment" summary (narrowed in W0051 so slim Reports compare) */
+    experiment: ComparableSummary;
     /** When this comparison was generated */
     generatedAt: string;
     /** Areas that improved beyond the noise threshold */
@@ -1162,23 +1313,88 @@ export interface PublishResult {
         result: SinkResult;
     }[];
 }
-/** Reference to an artifact in external object storage (GCS). See D0030. */
+/**
+ * A single per-entry row in `ArtifactRef.entries`. Carries enough metadata for
+ * Studio list/triage views to render without fetching the external payload.
+ *
+ * D0033/W0049 extensions:
+ *   - `association` — the axis values that identify this entry (`{task, model}`
+ *     for testOutputs, etc.). Present only on manifests written by
+ *     `emit()`; legacy manifests omit it and readers treat absence as `{}`.
+ *   - `truncated` — whether the entry payload was capped by the descriptor's
+ *     `capBytes`. Readers treat absence as `false` (pre-W0049 manifests were
+ *     never truncated because no caps were enforced).
+ *   - `preview` — an inline summary produced by the descriptor's
+ *     `manifestPreview.extract()`. Typed via the descriptor's preview schema;
+ *     omitted when the descriptor has no `manifestPreview`. Wiring lands in
+ *     W0051; the field is reserved here so manifests written now parse there.
+ */
+export interface ArtifactRefEntry {
+    key: string;
+    bytes: number;
+    association?: AssociationValues;
+    truncated?: boolean;
+    preview?: unknown;
+}
+/**
+ * Reference to an artifact in external object storage.
+ *
+ * `store` discriminates the backend: `"gcs"` uses `bucket` as the bucket
+ * name (D0032); `"local"` uses `bucket` as the absolute rootDir path
+ * under which `path` resolves to a file on disk (D0033 / W0050 M4).
+ * Consumers (Studio retrieval, contract tests) branch on `store` only
+ * when constructing the fetch URL — the `path` is store-relative and
+ * identical across backends for the same logical artifact.
+ *
+ * `layout` determines the on-disk shape:
+ *   - `"bulk"` — a single object at `path`. `entries` is absent.
+ *   - `"per-entry"` — `path` is a directory prefix. Each entry is a
+ *     separate object at `{path}/{sanitizedKey}.json`. `entries` inlines
+ *     the catalog so consumers can render drill-down states without a
+ *     second list call.
+ *
+ * D0033/W0049 extensions (optional — legacy manifests parse without them):
+ *   - `truncated` on the bulk row indicates the single-object body was capped.
+ *   - `preview` on the bulk row carries a descriptor-typed summary for list
+ *     views; wiring lands in W0051.
+ */
 export interface ArtifactRef {
-    store: "gcs";
+    store: "gcs" | "local";
+    /**
+     * GCS bucket name for `store: "gcs"`; absolute rootDir path for
+     * `store: "local"`. Kept as a single field so callers iterating
+     * manifest entries don't need to branch on `store` just to read the
+     * storage container.
+     */
     bucket: string;
     path: string;
     bytes?: number;
     entryCount?: number;
+    layout: "bulk" | "per-entry";
+    entries?: ArtifactRefEntry[];
+    truncated?: boolean;
+    preview?: unknown;
 }
+/**
+ * Catalog of artifact refs produced by a single pipeline run.
+ *
+ * Lives on `RunManifest.artifacts` (source of truth in GCS) and is
+ * snapshotted onto `Report.artifactManifest` at publish time.
+ *
+ * Derived from `ArtifactType` so adding a descriptor to the registry
+ * automatically admits it to the manifest catalog — drift between the
+ * two becomes a compile error (W0049 review finding C1).
+ */
+export type ArtifactManifest = Partial<Record<ArtifactType, ArtifactRef>>;
 /** A published evaluation report — the atomic unit of the report store */
 export interface Report {
-    /** External artifact references — set by publish step when uploader is available (D0030) */
-    artifacts?: {
-        testOutputs?: ArtifactRef;
-        renderedPrompts?: ArtifactRef;
-        rawResults?: ArtifactRef;
-        traces?: ArtifactRef;
-    };
+    /**
+     * Snapshot of the run manifest's `artifacts` slice at publish time (D0032).
+     * The source of truth lives in `gs://…/runs/{runId}/manifest.json`; this
+     * field denormalizes enough information for Studio to render drill-down
+     * states without an extra manifest fetch.
+     */
+    artifactManifest?: ArtifactManifest;
     /** Optional auto-comparison against the most recent comparable report */
     comparison?: ComparisonReport;
     /** When the evaluation completed */
@@ -1189,8 +1405,14 @@ export interface Report {
     id: ReportId;
     /** What produced this report */
     provenance: ReportProvenance;
-    /** The full score summary */
-    summary: ScoreSummary;
+    /**
+     * The slim published summary. Inlined prose fields (grader reasons, full
+     * failure-mode text, gap prose, agent-behavior arrays) moved to external
+     * artifacts in W0051; see `ReportSummary` for the retained shape and
+     * `docs/decisions/D0033-unified-run-anchored-artifact-capture.md` §§ M7 for
+     * the full migration table.
+     */
+    summary: ReportSummary;
     /** Optional human-supplied label */
     tag?: string;
     /** Auto-generated descriptive title for discoverability and sharing */
@@ -1239,76 +1461,75 @@ export interface ReportLineage {
      */
     rerunOf?: ReportId;
 }
-/** Full provenance metadata for an evaluation report */
-export interface ReportProvenance {
-    /** Which feature areas were evaluated */
-    areas: string[];
+/**
+ * Full provenance metadata for an evaluation report.
+ *
+ * Inherits the 9 run-description fields (mode, areas, taskIds, models,
+ * graderModel, source, evalFingerprint, trigger, git) from `RunContext`.
+ * Adding a field to `RunContext` makes it available here automatically —
+ * the structural extension is the drift-prevention mechanism described in
+ * D0032 § "Drift Prevention".
+ */
+export interface ReportProvenance extends RunContext {
     /** Release auto-scope metadata (when perspective evaluation was scoped to affected tasks) */
     autoScope?: ReportAutoScope;
     /** Content hash of the documentation context at eval time */
     contextHash?: string;
-    /**
-     * Evaluation fingerprint — SHA-256 of all inputs that affect eval output.
-     * Used for cross-environment cache lookup (CI → Content Lake).
-     * @see docs/design-docs/content-lake-eval-caching.md
-     */
-    evalFingerprint?: string;
-    /** Git metadata (when run from CI) */
-    git?: {
-        branch: string;
-        prNumber?: number;
-        repo: string;
-        sha: string;
-    };
-    /** Grader model used for scoring */
-    graderModel: string;
     /** Typed relationships with other reports (re-run, comparison) */
     lineage?: ReportLineage;
-    /** Evaluation mode */
-    mode: EvalMode;
-    /** Models under evaluation */
-    models: {
-        id: string;
-        label: string;
-    }[];
     /** @deprecated Use `promptfooUrls` — kept for backward compatibility */
     promptfooUrl?: string;
     /** Per-mode Promptfoo share URLs (one per sub-eval that produced a shareable link) */
     promptfooUrls?: PromptfooUrlEntry[];
-    /** Documentation source configuration */
-    source: {
-        baseUrl: string;
-        dataset?: string;
-        name: string;
-        perspective?: string;
-        projectId?: string;
-    };
+    /**
+     * Identity of the pipeline run that produced this report. Links the
+     * Content Lake document back to the GCS run manifest and its artifacts.
+     * @see docs/decisions/D0032-run-anchored-artifact-store.md
+     */
+    runId: RunId;
     /** Sanity document IDs that were targeted (if using --sanity-document) */
     targetDocuments?: string[];
-    /** Which specific task IDs were evaluated (if scoped) */
-    taskIds?: string[];
-    /** What initiated this evaluation */
-    trigger: ReportTrigger;
-}
-/** What triggered this evaluation */
-export type ReportTrigger = {
-    type: "ci";
-    runId: string;
-    workflow: string;
-} | {
-    type: "cross-repo";
-    callerRef?: string;
-    callerRepo: string;
-} | {
-    type: "manual";
-} | {
-    type: "scheduled";
-    schedule: string;
-} | {
-    type: "webhook";
-    documentId?: string;
-    source: string;
-};
+}
+/**
+ * A run's manifest in GCS (`runs/{runId}/manifest.json`). Source of truth
+ * for artifact locations, run identity, and outcome. Reports snapshot the
+ * `artifacts` slice into `Report.artifactManifest` at publish time.
+ *
+ * Written once by `FinalizeRunStep`; `reportIds` may be appended by
+ * `PublishReportStep` via strongly-consistent GCS overwrite.
+ */
+export interface RunManifest {
+    /** Schema version — bumped when the manifest shape changes */
+    version: 1;
+    /** Identity for this pipeline run */
+    runId: RunId;
+    /** When the manifest was written (pipeline finalization time) */
+    createdAt: ISOTimestamp;
+    /** Total pipeline duration */
+    durationMs: number;
+    /** Outcome of the run */
+    status: "completed" | "failed" | "partial";
+    /** Failure classification when status is not "completed" */
+    failureReason?: PipelineFailureReason;
+    /** What ran — shared shape with ReportProvenance */
+    context: RunContext;
+    /** Run-level aggregates (self-describing without a report) */
+    outcomes?: {
+        testSummary?: TestSummary;
+        usage?: PipelineUsage;
+        cache?: {
+            hits: number;
+            misses: number;
+            skipped: number;
+        };
+    };
+    /** Reports published from this run (0..N). Authoritative link is Report.provenance.runId. */
+    reportIds?: ReportId[];
+    /** Promptfoo share URLs collected during the run */
+    promptfooUrls?: PromptfooUrlEntry[];
+    /** Artifact catalog — per-type refs with inline per-entry indexes */
+    artifacts: ArtifactManifest;
+}
 /** Health check result for a sink */
 export type SinkHealthStatus = {
     healthy: false;

package/dist/_vendor/ailf-core/types/index.js

@@ -16,7 +16,7 @@ export { InMemoryPluginRegistry } from "./plugin-registry.js";
 // version is used internally by LiteracyModeConfig. If consumers need
 // the mode-specific version, they import from "./eval-mode-config.js".
 export { evalModeType } from "./eval-mode-config.js";
-export { err, fixtureId, ok, providerId, resultId, runId, suiteId, taskId, traceId, } from "./branded-ids.js";
+export { err, fixtureId, generateRunId, ok, providerId, resultId, runId, suiteId, taskId, traceId, } from "./branded-ids.js";
 // ---------------------------------------------------------------------------
 // Comparison (Approach 2: structured comparison output)
 // ---------------------------------------------------------------------------

package/dist/_vendor/ailf-shared/index.d.ts

@@ -13,3 +13,5 @@ export * from "./document-ref.js";
 export * from "./score-grades.js";
 export * from "./noise-threshold.js";
 export * from "./eval-modes.js";
+export * from "./run-trigger.js";
+export * from "./run-context.js";

package/dist/_vendor/ailf-shared/index.js

@@ -13,3 +13,5 @@ export * from "./document-ref.js";
 export * from "./score-grades.js";
 export * from "./noise-threshold.js";
 export * from "./eval-modes.js";
+export * from "./run-trigger.js";
+export * from "./run-context.js";

package/dist/_vendor/ailf-shared/run-context.d.ts

@@ -0,0 +1,55 @@
+/**
+ * RunContext — the set of fields that describe an AILF pipeline run.
+ *
+ * This is the single source of truth for run-description data. Both
+ * `RunManifest.context` (in GCS, written by FinalizeRunStep) and
+ * `ReportProvenance` (in Content Lake, built by PublishReportStep)
+ * carry this shape. `ReportProvenance extends RunContext` to
+ * structurally enforce parity — adding a field here becomes a
+ * compile-time failure until every consumer threads it through.
+ *
+ * Fields are alphabetized to match the surrounding codebase convention
+ * (see `ReportProvenance` in `@sanity/ailf-core`).
+ *
+ * @see docs/decisions/D0032-run-anchored-artifact-store.md
+ * @see docs/design-docs/run-artifact-store.md (§ Drift Prevention)
+ */
+import type { EvalMode } from "./eval-modes.js";
+import type { RunTrigger } from "./run-trigger.js";
+export interface RunContext {
+    /** Which feature areas were evaluated */
+    areas: string[];
+    /**
+     * Evaluation fingerprint — SHA-256 of all inputs that affect eval output.
+     * Used for cross-environment cache lookup (CI → Content Lake).
+     */
+    evalFingerprint?: string;
+    /** Git metadata (when run from CI) */
+    git?: {
+        branch: string;
+        prNumber?: number;
+        repo: string;
+        sha: string;
+    };
+    /** Grader model used for scoring */
+    graderModel: string;
+    /** Evaluation mode */
+    mode: EvalMode;
+    /** Models under evaluation */
+    models: {
+        id: string;
+        label: string;
+    }[];
+    /** Documentation source configuration */
+    source: {
+        baseUrl: string;
+        dataset?: string;
+        name: string;
+        perspective?: string;
+        projectId?: string;
+    };
+    /** Specific task IDs evaluated when scoped to a subset */
+    taskIds?: string[];
+    /** What initiated this run */
+    trigger: RunTrigger;
+}

package/dist/_vendor/ailf-shared/run-context.js

@@ -0,0 +1,17 @@
+/**
+ * RunContext — the set of fields that describe an AILF pipeline run.
+ *
+ * This is the single source of truth for run-description data. Both
+ * `RunManifest.context` (in GCS, written by FinalizeRunStep) and
+ * `ReportProvenance` (in Content Lake, built by PublishReportStep)
+ * carry this shape. `ReportProvenance extends RunContext` to
+ * structurally enforce parity — adding a field here becomes a
+ * compile-time failure until every consumer threads it through.
+ *
+ * Fields are alphabetized to match the surrounding codebase convention
+ * (see `ReportProvenance` in `@sanity/ailf-core`).
+ *
+ * @see docs/decisions/D0032-run-anchored-artifact-store.md
+ * @see docs/design-docs/run-artifact-store.md (§ Drift Prevention)
+ */
+export {};

package/dist/_vendor/ailf-shared/run-trigger.d.ts

@@ -0,0 +1,30 @@
+/**
+ * RunTrigger — what initiated a pipeline run.
+ *
+ * Lives in shared so it can be referenced from RunContext (the single
+ * source of truth for run-description fields) and consumed by both
+ * RunManifest (GCS) and ReportProvenance (Content Lake) without creating
+ * a core → core cycle.
+ *
+ * The `ci.runId` field is the external workflow run identifier (e.g.
+ * GitHub Actions run ID). It is unrelated to the pipeline-level `RunId`
+ * brand in `@sanity/ailf-core`, which identifies an AILF pipeline run.
+ */
+export type RunTrigger = {
+    type: "ci";
+    runId: string;
+    workflow: string;
+} | {
+    type: "cross-repo";
+    callerRef?: string;
+    callerRepo: string;
+} | {
+    type: "manual";
+} | {
+    type: "scheduled";
+    schedule: string;
+} | {
+    type: "webhook";
+    documentId?: string;
+    source: string;
+};

package/dist/_vendor/ailf-shared/run-trigger.js

@@ -0,0 +1,13 @@
+/**
+ * RunTrigger — what initiated a pipeline run.
+ *
+ * Lives in shared so it can be referenced from RunContext (the single
+ * source of truth for run-description fields) and consumed by both
+ * RunManifest (GCS) and ReportProvenance (Content Lake) without creating
+ * a core → core cycle.
+ *
+ * The `ci.runId` field is the external workflow run identifier (e.g.
+ * GitHub Actions run ID). It is unrelated to the pipeline-level `RunId`
+ * brand in `@sanity/ailf-core`, which identifies an AILF pipeline run.
+ */
+export {};

package/dist/artifact-capture/accumulating-artifact-writer.d.ts

@@ -0,0 +1,50 @@
+/**
+ * accumulating-artifact-writer.ts
+ *
+ * Decorator that wraps any `ArtifactWriter` and accumulates every
+ * successful `emit()` / `appendNdjson()` return into a run-scoped
+ * manifest slice. FinalizeRunStep reads the accumulator at the end of
+ * a pipeline and writes a `runs/{runId}/manifest.json` populated with
+ * one entry per produced artifact type.
+ *
+ * Before W0051 revisit: only `calculate-scores-step` registered its
+ * ref (for `testOutputs`) into `state.artifactRefs`; every other
+ * producer discarded the returned ref. The result was empty
+ * `Report.artifactManifest` fields and no per-entry preview lookup
+ * for Studio hooks. Wrapping the writer at the composition-root level
+ * closes that gap without per-producer bookkeeping.
+ *
+ * Merging rules (per type):
+ *   - `bulk`: last-writer-wins. A pipeline that emits the same bulk
+ *     artifact twice overwrites — matches GCS semantics.
+ *   - `per-entry`: entries accumulate into a keyed map. A later emit
+ *     at the same `entries[].key` replaces the earlier one.
+ *
+ * The decorator holds no disk state; the `_resetAccumulated()` hook is
+ * for unit tests that rerun emit sequences within a single writer.
+ *
+ * @see docs/decisions/D0033-unified-run-anchored-artifact-capture.md (§ M5)
+ */
+import type { ArtifactEntry, ArtifactManifest, ArtifactRef, ArtifactType, ArtifactWriter, AssociationValues, RunId, RunManifest } from "../_vendor/ailf-core/index.d.ts";
+export declare class AccumulatingArtifactWriter implements ArtifactWriter {
+    /**
+     * Exposed so composition-root tests can assert on the underlying backend
+     * (LocalFilesystemArtifactWriter, FanoutArtifactWriter, etc.) without
+     * plumbing a separate accessor. Treat as read-only.
+     */
+    readonly inner: ArtifactWriter;
+    private readonly accumulated;
+    constructor(inner: ArtifactWriter);
+    /** Snapshot of every ref produced this far, keyed by artifact type. */
+    getAccumulatedArtifactRefs(): ArtifactManifest;
+    /** Test-only. Clears accumulated refs without touching the inner writer. */
+    _resetAccumulated(): void;
+    emit<T extends ArtifactType>(type: T, association: AssociationValues, payload: unknown): Promise<ArtifactRef | null>;
+    appendNdjson<T extends ArtifactType>(type: T, association: AssociationValues, rows: readonly unknown[]): Promise<ArtifactRef | null>;
+    writeManifest(runId: RunId, manifest: RunManifest): Promise<ArtifactRef | null>;
+    /** @deprecated — forwarded to the inner writer without accumulation. */
+    writeBulk(type: ArtifactType, runId: RunId, data: unknown): Promise<ArtifactRef | null>;
+    /** @deprecated — forwarded to the inner writer without accumulation. */
+    writePerEntry(type: ArtifactType, runId: RunId, entries: readonly ArtifactEntry[]): Promise<ArtifactRef | null>;
+    private mergeRef;
+}