@sanity/ailf 7.0.1 → 7.1.2

package/dist/pipeline/run-context.d.ts

@@ -13,7 +13,7 @@
  * @see docs/decisions/D0032-run-anchored-artifact-store.md (§ Move 5 — Drift Prevention)
  */
 import { type Logger, type RunContext } from "../_vendor/ailf-core/index.d.ts";
-import { type RunClassification, type RunExecutor, type RunExecutorSurface, type RunHost, type RunLineage, type RunOwner, type RunTool } from "../_vendor/ailf-shared/index.d.ts";
+import { type LiteracyVariant, type RunClassification, type RunExecutor, type RunExecutorSurface, type RunHost, type RunLineage, type RunOwner, type RunTool } from "../_vendor/ailf-shared/index.d.ts";
 import type { ResolvedSourceConfig } from "../sources.js";
 import type { EvalMode } from "./types.js";
 /**
@@ -74,6 +74,21 @@ export interface RunContextInput {
     source: ResolvedSourceConfig;
     /** Specific task IDs evaluated (if scoped) */
     taskIds?: string[];
+    /**
+     * Literacy mode variant (`baseline | agentic | observed | full`). Only
+     * meaningful when `mode === "literacy"`; ignored for other modes. Lands
+     * on `RunContext.variant` and `ReportProvenance.variant` so consumers
+     * can disambiguate which literacy variant the run executed.
+     */
+    variant?: LiteracyVariant;
+    /**
+     * Model IDs the caller requested via `PipelineRequest.models`. When
+     * present, `RunContext.models` is filtered to this subset so the report's
+     * `provenance.models` reflects what was actually evaluated. Unknown IDs
+     * are silently filtered out — the upstream rejection path (W0281
+     * `filterModelsByRequest`) has already failed the run or warned.
+     */
+    requestedModelIds?: string[];
 }
 /**
  * Derive `RunContext` from pipeline inputs. The only construction path.

package/dist/pipeline/run-context.js

@@ -68,8 +68,18 @@ export function buildRunContext(input) {
     // config/models.ts model matrix — listing those models would be
     // misleading. Only include them for literacy mode where they're the
     // actual eval targets.
+    //
+    // When `PipelineRequest.models` pinned a subset, filter here too so
+    // `provenance.models` matches what actually ran (W0281). Without this
+    // the report would advertise the full cohort even though only the
+    // requested subset reached the LLMs.
+    const requestedSet = input.requestedModelIds?.length
+        ? new Set(input.requestedModelIds)
+        : undefined;
     const evaluatedModels = input.mode === "literacy"
-        ? models.models.map((m) => ({ id: m.id, label: m.label }))
+        ? models.models
+            .filter((m) => !requestedSet || requestedSet.has(m.id))
+            .map((m) => ({ id: m.id, label: m.label }))
         : [];
     return {
         areas: input.areas,
@@ -95,6 +105,7 @@ export function buildRunContext(input) {
         taskIds: input.taskIds,
         tool,
         trigger,
+        variant: input.mode === "literacy" ? input.variant : undefined,
     };
 }
 // ---------------------------------------------------------------------------

package/dist/pipeline/validate.d.ts

@@ -14,11 +14,15 @@ import type { ValidationIssue, ValidationResult } from "./types.js";
  */
 export declare function validateConfiguration(rootDir: string): ValidationResult;
 /**
- * Check that canonical context files exist. These are the per-task
- * gold-retrieval contexts actually referenced by task definitions.
+ * Check that the canonical-contexts directory exists.
  *
- * Contexts are generated by fetch-docs and may not exist yet —
- * returns warnings, not errors.
+ * Contexts are populated by fetch-docs, which scopes to the tasks
+ * actually being evaluated (not every task in the registry). Warning
+ * on individual missing files here would fire for every task the user
+ * didn't select — pure noise that previously crowded out real errors
+ * in the GHA safety-net's tail-of-log capture (W0282). The per-task
+ * precondition is enforced by `run-eval-step.ts:checkCanonicalContextsExist`
+ * against the filtered task set, where missing files are real errors.
  */
 export declare function validateContexts(rootDir: string): ValidationIssue[];
 /**

package/dist/pipeline/validate.js

@@ -34,11 +34,15 @@ export function validateConfiguration(rootDir) {
     return { issues, valid };
 }
 /**
- * Check that canonical context files exist. These are the per-task
- * gold-retrieval contexts actually referenced by task definitions.
+ * Check that the canonical-contexts directory exists.
  *
- * Contexts are generated by fetch-docs and may not exist yet —
- * returns warnings, not errors.
+ * Contexts are populated by fetch-docs, which scopes to the tasks
+ * actually being evaluated (not every task in the registry). Warning
+ * on individual missing files here would fire for every task the user
+ * didn't select — pure noise that previously crowded out real errors
+ * in the GHA safety-net's tail-of-log capture (W0282). The per-task
+ * precondition is enforced by `run-eval-step.ts:checkCanonicalContextsExist`
+ * against the filtered task set, where missing files are real errors.
  */
 export function validateContexts(rootDir) {
     const source = "validateContexts";
@@ -46,20 +50,6 @@ export function validateContexts(rootDir) {
     const canonicalDir = path.join(rootDir, "contexts", "canonical");
     if (!fs.existsSync(canonicalDir)) {
         issues.push(warning(source, "contexts/canonical/ directory not found — run 'pnpm fetch-docs' to generate", canonicalDir));
-        return issues;
-    }
-    const mappings = resolveMappings(rootDir);
-    for (const [, areaConfig] of Object.entries(mappings.feature_areas)) {
-        if (!areaConfig?.tasks)
-            continue;
-        for (const task of areaConfig.tasks) {
-            if (!task.id)
-                continue;
-            const contextFile = path.join(canonicalDir, `${task.id}.md`);
-            if (!fs.existsSync(contextFile)) {
-                issues.push(warning(source, `Missing canonical context for task '${task.id}' — run 'pnpm fetch-docs' to generate`, contextFile));
-            }
-        }
     }
     return issues;
 }

package/dist/report-store.d.ts

@@ -15,7 +15,7 @@
  * @see docs/design-docs/report-store/domain-model.md
  */
 import type { SanityClient } from "@sanity/client";
-import type { ArtifactRef, ArtifactType, SynthesisCostTelemetry } from "./_vendor/ailf-core/index.d.ts";
+import type { ArtifactRef, ArtifactType, LoadBaselineResult, SynthesisCostTelemetry } from "./_vendor/ailf-core/index.d.ts";
 import type { ComparisonReport, ISOTimestamp, LineageQuery, Report, ReportId, ReportProvenance, ScoreSummary } from "./pipeline/types.js";
 /**
  * Result of an auto-comparison, bundling the ComparisonReport with the
@@ -113,6 +113,19 @@ export declare class ReportStore {
      *   W0191 runtime schema gate. Sanity API failures still return null.
      */
     read(id: ReportId): Promise<null | Report>;
+    /**
+     * Load a previously-published report's score summary as a baseline
+     * for comparison. Returns a discriminated result so the caller can
+     * distinguish a genuine 404 (skip compare with a clear reason) from
+     * a transport failure (fail the step — the user pinned a baseline
+     * and deserves to know it didn't actually compare).
+     *
+     * The report's `summary` field is a `ReportSummary` — a superset of
+     * `ComparableSummary` — so the projection below carries everything
+     * the `compare()` primitive needs (`overall`, `perModel`, `scores`)
+     * without re-hydrating the slim prose/array fields.
+     */
+    loadBaselineFromReport(reportId: string): Promise<LoadBaselineResult>;
     /**
      * Write a report to the Sanity Content Lake.
      *

package/dist/report-store.js

@@ -270,6 +270,38 @@ export class ReportStore {
             return null;
         }
     }
+    /**
+     * Load a previously-published report's score summary as a baseline
+     * for comparison. Returns a discriminated result so the caller can
+     * distinguish a genuine 404 (skip compare with a clear reason) from
+     * a transport failure (fail the step — the user pinned a baseline
+     * and deserves to know it didn't actually compare).
+     *
+     * The report's `summary` field is a `ReportSummary` — a superset of
+     * `ComparableSummary` — so the projection below carries everything
+     * the `compare()` primitive needs (`overall`, `perModel`, `scores`)
+     * without re-hydrating the slim prose/array fields.
+     */
+    async loadBaselineFromReport(reportId) {
+        try {
+            const doc = await this.client.fetch(`*[_type == $type && reportId == $id][0]{ summary }`, { id: reportId, type: REPORT_TYPE });
+            const summary = doc?.summary;
+            if (!summary)
+                return { kind: "not_found" };
+            return {
+                kind: "ok",
+                baseline: {
+                    overall: summary.overall,
+                    perModel: summary.perModel,
+                    scores: summary.scores,
+                },
+            };
+        }
+        catch (error) {
+            const message = error instanceof Error ? error.message : String(error);
+            return { kind: "error", message };
+        }
+    }
     /**
      * Write a report to the Sanity Content Lake.
      *

package/dist/sanity/client.js

@@ -108,8 +108,8 @@ export function getSanityClient(overrides, source) {
  * fall back to `SANITY_DATASET` so existing CI workflows that pin a
  * test/staging dataset (e.g. Tier 2 with `SANITY_DATASET=ailf-test`)
  * continue to work without a new env var. The hard-coded fallback is
- * the editorial dataset name during the D0043 cutover window — the flip
- * to `ailf-prod-private` happens after the migration script runs.
+ * `AILF_DATASET_DEFAULT` (`ailf-prod-private`, D0043) — only reached for
+ * ad-hoc runs with no env at all.
  *
  * Token resolution prefers the AILF-scoped token, falling back to
  * the shared `SANITY_API_TOKEN`.

package/dist/sanity/queries.d.ts

@@ -69,7 +69,7 @@ export declare const ALL_ARTICLES_QUERY = "\n  *[_type == \"article\"\n    && !(
  *
  * Usage: client.fetch(ARTICLES_METADATA_BY_SLUGS_QUERY, { slugs: ["slug-a", "slug-b"] })
  */
-export declare const ARTICLES_METADATA_BY_SLUGS_QUERY = "\n  *[_type == \"article\"\n    && slug.current in $slugs\n    && !(_id in path(\"drafts.**\"))\n  ] {\n    \"slug\": slug.current,\n    _id,\n    _rev,\n    title\n  }\n";
+export declare const ARTICLES_METADATA_BY_SLUGS_QUERY = "\n  *[_type == \"article\"\n    && slug.current in $slugs\n    && !(_id in path(\"drafts.**\"))\n  ] {\n    \"slug\": slug.current,\n    \"sectionSlug\": primarySection->slug.current,\n    _id,\n    _rev,\n    title\n  }\n";
 /**
  * Fetch a single article by its slug — identical to ARTICLE_BY_SLUG_QUERY
  * but designed to be called with a perspective-enabled client.

package/dist/sanity/queries.js

@@ -203,6 +203,7 @@ export const ARTICLES_METADATA_BY_SLUGS_QUERY = `
     && !(_id in path("drafts.**"))
   ] {
     "slug": slug.current,
+    "sectionSlug": primarySection->slug.current,
     _id,
     _rev,
     title

package/dist/sources.js

@@ -37,6 +37,44 @@ const DEFAULT_SOURCE = {
     studioOrigin: "https://admin.sanity.io",
     urls: [],
 };
+/**
+ * Apply `SourceOverrides` + env-var fallbacks to `DEFAULT_SOURCE`.
+ *
+ * The DEFAULT_SOURCE early-return branches are taken when `config/sources`
+ * is missing or empty — the production state, since the named source
+ * definitions actually live in the `sanity-literacy` preset's `sourceDefs`
+ * (which `loadSource` doesn't consult). Returning `DEFAULT_SOURCE`
+ * verbatim drops every override the caller passed in, including
+ * `perspective` — observed live as production-source release evals
+ * fetching the published doc revision (W0295).
+ *
+ * The merge order mirrors the priority-1 (env-baseUrl) branch. The two
+ * paths diverge in three ways, all intentional: this branch (a) pins
+ * `baseUrl` / `llmsTxt` / `name` / `priorityDomain` to `DEFAULT_SOURCE`,
+ * (b) returns `documentIds: []` (the prior `DEFAULT_SOURCE` shape) where
+ * priority-1 would return `undefined` — both fall through the same
+ * `length > 0` consumer check, so behaviorally equivalent.
+ */
+function applyOverridesToDefault(overrides) {
+    const allowedOrigins = overrides?.allowedOrigins ?? parseAllowedOriginsEnv();
+    const headers = overrides?.headers ?? parseHeadersEnv();
+    return {
+        ...DEFAULT_SOURCE,
+        ...(allowedOrigins ? { allowedOrigins } : {}),
+        // oxlint-disable-next-line @typescript-eslint/prefer-nullish-coalescing -- empty string env var should fall back
+        dataset: overrides?.dataset ?? (process.env.SANITY_DATASET || "next"),
+        documentIds: overrides?.documentIds ?? parseDocumentIdsEnv() ?? [],
+        ...(headers ? { headers } : {}),
+        // oxlint-disable-next-line @typescript-eslint/prefer-nullish-coalescing -- empty string env var should fall back
+        perspective: overrides?.perspective ?? (process.env.SANITY_PERSPECTIVE || undefined),
+        // oxlint-disable-next-line @typescript-eslint/prefer-nullish-coalescing -- empty string env var should fall back
+        projectId: overrides?.projectId ?? (process.env.SANITY_PROJECT_ID || "3do82whm"),
+        // oxlint-disable-next-line @typescript-eslint/prefer-nullish-coalescing -- empty string env var should fall back
+        studioOrigin: overrides?.studioOrigin ??
+            (process.env.SANITY_STUDIO_ORIGIN || "https://admin.sanity.io"),
+        urls: overrides?.directUrls ?? parseDirectUrlsEnv(),
+    };
+}
 // ---------------------------------------------------------------------------
 // Validation
 // ---------------------------------------------------------------------------
@@ -117,12 +155,12 @@ export function loadSource(name, overrides, logger) {
             defaultBaseUrl: DEFAULT_SOURCE.baseUrl,
         });
         console.log("  No config/sources found, using built-in default (sanity.io production)");
-        return DEFAULT_SOURCE;
+        return applyOverridesToDefault(overrides);
     }
     if (!rawFile?.sources || Object.keys(rawFile.sources).length === 0) {
         log.debug("config/sources is empty, falling back to built-in default");
         console.log("  config/sources is empty, using built-in default");
-        return DEFAULT_SOURCE;
+        return applyOverridesToDefault(overrides);
     }
     // Resolve which source to use
     const sourceName = 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sanity/ailf",
-  "version": "7.0.1",
+  "version": "7.1.2",
   "private": false,
   "publishConfig": {
     "access": "public"