@inflexa-ai/harness 0.1.1 → 0.3.0

package/dist/providers/prompt-cache.js

@@ -0,0 +1,77 @@
+/**
+ * Prompt-cache policy → provider wire options.
+ *
+ * This is the **only** place in the harness that names a vendor for caching.
+ * Everything upstream — `PromptCachePolicy`, `ChatUsage`, `RunAgentOptions` —
+ * speaks the neutral harness concept; this module translates it once, at the
+ * provider seam, so the loop never learns which vendor it is talking to.
+ *
+ * ## Why the emitted options are safe on every provider
+ *
+ * The AI SDK's `providerOptions` is a namespaced bag: each provider reads only
+ * `providerOptions[<its own name>]` and ignores every other key
+ * (`parseProviderOptions` in `@ai-sdk/provider-utils` returns `undefined` when
+ * its namespace is absent). So the `anthropic` namespace is inert — not an
+ * error — on an OpenAI-compatible model, which is exactly what we want: that
+ * family does automatic server-side prefix caching and needs no directive.
+ *
+ * ## What the Anthropic namespace does
+ *
+ * A request-level `cacheControl` makes the provider emit a single top-level
+ * `cache_control` marker; the server then places the breakpoint on the last
+ * cacheable block, so the whole prefix (tools → system → history) is cached
+ * without the harness hand-placing per-block markers.
+ *
+ * ## Cache defeaters — what silently kills the hit rate
+ *
+ * The cache keys on an *exact prefix*. Anything that perturbs the head of the
+ * request invalidates everything after it. One known defeater remains in this
+ * codebase, flagged at its own site as a separate change:
+ *
+ *  1. `runAgent`'s forced wrap-up swaps the tool set to `{}` — tools sit at the
+ *     very front of the prefix, so that one call reads nothing back and rewrites
+ *     the cache from scratch (`loop/run-agent.ts`).
+ *
+ * `loadRecent`'s history eviction used to be a second defeater — it advanced the
+ * window one turn per turn, shifting the message prefix every request. It now
+ * evicts in whole blocks so the prefix holds still between block boundaries
+ * (`memory/thread-history.ts`).
+ *
+ * A sandbox agent's system prompt is NOT one of them: it is a pure function of
+ * its agent type, byte-identical across every step of every run, and the per-step
+ * paths ride in the step's briefing instead (`agents/sandbox/shared.ts`,
+ * `prompts/briefing.ts`). Keep it that way — one interpolated id or path there
+ * makes every step's ~20k-char prefix unique.
+ *
+ * ## Where caching is a no-op regardless
+ *
+ * The Claude Max OAuth path does not honour cache directives, and the OSS CLI
+ * defaults to routing through a local CLIProxyAPI on exactly that path. Caching
+ * only engages against a direct API key or a gateway. The cache-token metrics
+ * (`loop/metrics.ts`) are what tell the two apart at runtime — a flat-zero
+ * `cache_read_tokens` counter is the symptom.
+ */
+/**
+ * The default policy: cache with the 5-minute TTL.
+ *
+ * A cache write costs a premium over a plain input token and only pays for
+ * itself once something reads it back, so caching is worth it exactly when a
+ * shared prefix is re-sent within the TTL. An agent loop always re-sends: every
+ * iteration replays the same tools + system + the whole transcript so far, and
+ * even a two-iteration run breaks even. The 5m TTL covers a live loop; `1h` is
+ * for hosts that also want the *next* turn on a thread to land warm, and costs
+ * more per write. One-shot LLM calls should stay `"off"` — they pay the write
+ * premium for a cache nothing ever reads.
+ */
+export const DEFAULT_PROMPT_CACHE = { ttl: "5m" };
+/**
+ * Translate a neutral cache policy into provider wire options.
+ *
+ * Returns `undefined` for `"off"` so the caller can leave `providerOptions`
+ * entirely unset rather than sending an empty bag.
+ */
+export function promptCacheProviderOptions(policy) {
+    if (policy === "off")
+        return undefined;
+    return { anthropic: { cacheControl: { type: "ephemeral", ttl: policy.ttl } } };
+}

package/dist/providers/types.d.ts

@@ -24,12 +24,45 @@ export interface ChatRequest {
     };
     readonly providerOptions?: ProviderOptions;
 }
+/**
+ * Vendor-neutral prompt-cache policy — a harness concept, not a vendor one.
+ *
+ * `{ ttl }` asks the provider to cache the request prefix (system + tools +
+ * message history) for that lifetime; `"off"` sends no cache directive at all.
+ * Vendors that cache automatically (the OpenAI-compatible family does prefix
+ * caching server-side, unprompted) need no directive, so the policy is a no-op
+ * for them rather than an error — see `./prompt-cache.ts`, the single place the
+ * harness translates this into vendor wire options.
+ */
+export type PromptCachePolicy = {
+    readonly ttl: "5m" | "1h";
+} | "off";
+/**
+ * Token accounting for one chat call, in harness-neutral names.
+ *
+ * `inputTokens` is the *total* billed prefix — cached and uncached alike — so a
+ * cache hit rate is `cacheReadInputTokens / inputTokens`, not a ratio against
+ * some separate uncached figure. `cacheCreationInputTokens` is the write that
+ * seeds the cache (billed at a premium; it only pays for itself once a later
+ * call reads it back).
+ *
+ * Every field is optional: a provider that reports no usage at all, or reports
+ * totals without a cache breakdown, is legitimate. Absent means "not reported",
+ * never "zero".
+ */
+export interface ChatUsage {
+    readonly inputTokens?: number;
+    readonly outputTokens?: number;
+    readonly cacheCreationInputTokens?: number;
+    readonly cacheReadInputTokens?: number;
+}
 export interface ChatResponse {
     readonly message: Extract<ModelMessage, {
         role: "assistant";
     }>;
     readonly finishReason: FinishReason;
     readonly rawFinishReason?: string;
+    readonly usage?: ChatUsage;
 }
 export type ChatStreamEvent = {
     readonly type: "text-delta";

package/dist/reference-data/catalog.d.ts

@@ -0,0 +1,131 @@
+import { type Result } from "neverthrow";
+import { z } from "zod";
+/** Schema version for the checked-in reference-data catalog. */
+export declare const REFERENCE_DATA_CATALOG_VERSION: 1;
+/** Dataset-relative destination of one artifact below its install path. */
+export declare const ReferenceArtifactPathSchema: z.ZodString;
+/** Lowercase SHA-256 digest of one artifact's bytes. */
+export declare const ReferenceSha256Schema: z.ZodString;
+/** One immutable final file distributed for a reference dataset. */
+export declare const ReferenceArtifactSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
+    integrity: z.ZodLiteral<"pinned">;
+    path: z.ZodString;
+    url: z.ZodURL;
+    bytes: z.ZodNumber;
+    sha256: z.ZodString;
+}, z.core.$strict>, z.ZodObject<{
+    integrity: z.ZodLiteral<"unpinned">;
+    path: z.ZodString;
+    url: z.ZodURL;
+}, z.core.$strict>], "integrity">;
+/** Which integrity guarantee an artifact can actually offer. */
+export type ReferenceIntegrity = z.infer<typeof ReferenceArtifactSchema>["integrity"];
+/** Provenance and licensing information for one supported reference dataset. */
+export declare const ReferenceDatasetSchema: z.ZodObject<{
+    id: z.ZodString;
+    version: z.ZodString;
+    title: z.ZodString;
+    description: z.ZodString;
+    sourceUrl: z.ZodURL;
+    license: z.ZodObject<{
+        identifier: z.ZodString;
+        url: z.ZodOptional<z.ZodURL>;
+    }, z.core.$strip>;
+    recommendation: z.ZodObject<{
+        group: z.ZodString;
+        recommended: z.ZodBoolean;
+    }, z.core.$strip>;
+    artifacts: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
+        integrity: z.ZodLiteral<"pinned">;
+        path: z.ZodString;
+        url: z.ZodURL;
+        bytes: z.ZodNumber;
+        sha256: z.ZodString;
+    }, z.core.$strict>, z.ZodObject<{
+        integrity: z.ZodLiteral<"unpinned">;
+        path: z.ZodString;
+        url: z.ZodURL;
+    }, z.core.$strict>], "integrity">>;
+}, z.core.$strip>;
+/** Versioned canonical catalog consumed identically by every harness embedder. */
+export declare const ReferenceDataCatalogSchema: z.ZodObject<{
+    version: z.ZodLiteral<1>;
+    datasets: z.ZodArray<z.ZodObject<{
+        id: z.ZodString;
+        version: z.ZodString;
+        title: z.ZodString;
+        description: z.ZodString;
+        sourceUrl: z.ZodURL;
+        license: z.ZodObject<{
+            identifier: z.ZodString;
+            url: z.ZodOptional<z.ZodURL>;
+        }, z.core.$strip>;
+        recommendation: z.ZodObject<{
+            group: z.ZodString;
+            recommended: z.ZodBoolean;
+        }, z.core.$strip>;
+        artifacts: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
+            integrity: z.ZodLiteral<"pinned">;
+            path: z.ZodString;
+            url: z.ZodURL;
+            bytes: z.ZodNumber;
+            sha256: z.ZodString;
+        }, z.core.$strict>, z.ZodObject<{
+            integrity: z.ZodLiteral<"unpinned">;
+            path: z.ZodString;
+            url: z.ZodURL;
+        }, z.core.$strict>], "integrity">>;
+    }, z.core.$strip>>;
+}, z.core.$strip>;
+type DeepReadonly<T> = T extends (...args: never[]) => unknown ? T : T extends readonly (infer Item)[] ? readonly DeepReadonly<Item>[] : T extends object ? {
+    readonly [Key in keyof T]: DeepReadonly<T[Key]>;
+} : T;
+export type ReferenceArtifact = DeepReadonly<z.infer<typeof ReferenceArtifactSchema>>;
+export type ReferenceDataset = DeepReadonly<z.infer<typeof ReferenceDatasetSchema>>;
+export type ReferenceDataCatalog = DeepReadonly<z.infer<typeof ReferenceDataCatalogSchema>>;
+/**
+ * Stable identity of one artifact within the catalog, independent of its URL.
+ * Embedders key resumable partials off this, so a moved upstream URL does not
+ * orphan a half-finished transfer.
+ */
+export declare function referenceArtifactKey(dataset: {
+    readonly id: string;
+    readonly version: string;
+}, artifact: {
+    readonly path: string;
+}): string;
+/**
+ * Canonical release catalog. Every artifact is fetched directly from the third
+ * party that publishes it; this project hosts, mirrors, and redistributes
+ * nothing. Entries are added only with real upstream URLs, provenance, and
+ * licensing data — plus a size and digest whenever the upstream publishes
+ * immutable bytes we can pin to.
+ *
+ * `version` is the dataset's upstream release identifier. Datasets built on an
+ * upstream that has no immutable release — NCBI regenerates `gene_info` daily,
+ * Reactome overwrites `current` and deletes prior releases — are versioned
+ * `current` and carry `unpinned` artifacts.
+ */
+export declare const REFERENCE_DATA_CATALOG: ReferenceDataCatalog;
+/** One selected dataset and its host-neutral installation location. */
+export type ReferenceInstallPlanDataset = ReferenceDataset & {
+    readonly installPath: string;
+};
+/** Deterministic plan; every artifact carries the upstream URL it is fetched from. */
+export interface ReferenceInstallPlan {
+    readonly catalogVersion: typeof REFERENCE_DATA_CATALOG_VERSION;
+    readonly datasets: readonly ReferenceInstallPlanDataset[];
+}
+/** A requested dataset id was not present in the selected catalog. */
+export declare class UnknownReferenceDatasetError extends Error {
+    readonly code: "unknown_reference_dataset";
+    readonly unknownId: string;
+    readonly availableIds: readonly string[];
+    constructor(unknownId: string, availableIds: readonly string[]);
+}
+/** Resolve dataset ids without I/O, preserving a stable dataset-id ordering. */
+/** @internal Fixture seam for harness contract tests; not exported by the package barrel. */
+export declare function resolveReferenceInstallPlanFromCatalogForTesting(datasetIds: readonly string[], catalog: ReferenceDataCatalog): Result<ReferenceInstallPlan, UnknownReferenceDatasetError>;
+/** Resolve ids strictly against the immutable canonical catalog shipped by this package. */
+export declare function resolveReferenceInstallPlan(datasetIds: readonly string[]): Result<ReferenceInstallPlan, UnknownReferenceDatasetError>;
+export {};

package/dist/reference-data/catalog.js

@@ -0,0 +1,348 @@
+import { posix as posixPath } from "node:path";
+import { err, ok } from "neverthrow";
+import { z } from "zod";
+/** Schema version for the checked-in reference-data catalog. */
+export const REFERENCE_DATA_CATALOG_VERSION = 1;
+const SAFE_ID = /^[a-z0-9]+(?:[._-][a-z0-9]+)*$/;
+const SAFE_VERSION = /^[A-Za-z0-9]+(?:[._-][A-Za-z0-9]+)*$/;
+const SHA256 = /^[a-f0-9]{64}$/;
+function isSafeRelativePath(value) {
+    if (value.length === 0 || value.includes("\\") || value.startsWith("/"))
+        return false;
+    const segments = value.split("/");
+    return segments.every((segment) => segment.length > 0 && segment !== "." && segment !== "..") && posixPath.normalize(value) === value;
+}
+/** Dataset-relative destination of one artifact below its install path. */
+export const ReferenceArtifactPathSchema = z.string().refine(isSafeRelativePath, "Expected a safe dataset-relative path");
+/** Lowercase SHA-256 digest of one artifact's bytes. */
+export const ReferenceSha256Schema = z.string().regex(SHA256, "Expected a lowercase SHA-256 digest");
+/**
+ * Artifacts are fetched from the upstream that publishes them — this project
+ * redistributes nothing — so the URL must be a real third-party https endpoint.
+ */
+const ReferenceArtifactUrlSchema = z.url().refine((value) => value.startsWith("https://"), "Reference artifacts must be fetched over https");
+/**
+ * An artifact whose upstream publishes immutable, versioned bytes. The catalog
+ * carries the size and digest, so a download is verified against *this file*
+ * before it is ever activated: a mismatch means the bytes are not what we
+ * reviewed, and the install fails.
+ */
+const PinnedReferenceArtifactSchema = z.strictObject({
+    integrity: z.literal("pinned"),
+    path: ReferenceArtifactPathSchema,
+    url: ReferenceArtifactUrlSchema,
+    bytes: z.number().int().positive(),
+    sha256: ReferenceSha256Schema,
+});
+/**
+ * An artifact whose upstream regenerates the same URL in place — NCBI rebuilds
+ * `gene_info` continuously and Reactome's `current` release is overwritten and
+ * its predecessors deleted. No checked-in digest can survive that, and pinning
+ * one would only guarantee a broken download. Integrity is therefore
+ * trust-on-first-use: the installer records the bytes it actually received in
+ * the receipt, and `verify` proves the files have not changed *since install*.
+ * This is a weaker guarantee than `pinned` and is surfaced as such to the user.
+ */
+const UnpinnedReferenceArtifactSchema = z.strictObject({
+    integrity: z.literal("unpinned"),
+    path: ReferenceArtifactPathSchema,
+    url: ReferenceArtifactUrlSchema,
+});
+/** One immutable final file distributed for a reference dataset. */
+export const ReferenceArtifactSchema = z.discriminatedUnion("integrity", [PinnedReferenceArtifactSchema, UnpinnedReferenceArtifactSchema]);
+/** Provenance and licensing information for one supported reference dataset. */
+export const ReferenceDatasetSchema = z.object({
+    id: z.string().regex(SAFE_ID, "Expected a stable lowercase dataset id"),
+    version: z.string().regex(SAFE_VERSION, "Expected a safe dataset version path segment"),
+    title: z.string().min(1),
+    description: z.string().min(1),
+    sourceUrl: z.url(),
+    license: z.object({
+        identifier: z.string().min(1),
+        url: z.url().optional(),
+    }),
+    recommendation: z.object({
+        group: z.string().min(1),
+        recommended: z.boolean(),
+    }),
+    artifacts: z.array(ReferenceArtifactSchema).min(1),
+});
+/** Versioned canonical catalog consumed identically by every harness embedder. */
+export const ReferenceDataCatalogSchema = z
+    .object({
+    version: z.literal(REFERENCE_DATA_CATALOG_VERSION),
+    datasets: z.array(ReferenceDatasetSchema),
+})
+    .superRefine((catalog, ctx) => {
+    const ids = new Set();
+    for (const [datasetIndex, dataset] of catalog.datasets.entries()) {
+        if (ids.has(dataset.id)) {
+            ctx.addIssue({ code: "custom", message: `Duplicate dataset id: ${dataset.id}`, path: ["datasets", datasetIndex, "id"] });
+        }
+        ids.add(dataset.id);
+        const paths = new Set();
+        for (const [artifactIndex, artifact] of dataset.artifacts.entries()) {
+            if (paths.has(artifact.path)) {
+                ctx.addIssue({
+                    code: "custom",
+                    message: `Duplicate artifact path in ${dataset.id}: ${artifact.path}`,
+                    path: ["datasets", datasetIndex, "artifacts", artifactIndex, "path"],
+                });
+            }
+            paths.add(artifact.path);
+        }
+    }
+});
+/**
+ * Stable identity of one artifact within the catalog, independent of its URL.
+ * Embedders key resumable partials off this, so a moved upstream URL does not
+ * orphan a half-finished transfer.
+ */
+export function referenceArtifactKey(dataset, artifact) {
+    return `${dataset.id}/${dataset.version}/${artifact.path}`;
+}
+function deepFreeze(value) {
+    if (value && typeof value === "object" && !Object.isFrozen(value)) {
+        for (const nested of Object.values(value))
+            deepFreeze(nested);
+        Object.freeze(value);
+    }
+    return value;
+}
+/**
+ * Canonical release catalog. Every artifact is fetched directly from the third
+ * party that publishes it; this project hosts, mirrors, and redistributes
+ * nothing. Entries are added only with real upstream URLs, provenance, and
+ * licensing data — plus a size and digest whenever the upstream publishes
+ * immutable bytes we can pin to.
+ *
+ * `version` is the dataset's upstream release identifier. Datasets built on an
+ * upstream that has no immutable release — NCBI regenerates `gene_info` daily,
+ * Reactome overwrites `current` and deletes prior releases — are versioned
+ * `current` and carry `unpinned` artifacts.
+ */
+export const REFERENCE_DATA_CATALOG = deepFreeze(ReferenceDataCatalogSchema.parse({
+    version: REFERENCE_DATA_CATALOG_VERSION,
+    datasets: [
+        {
+            id: "ncbi-gene-human",
+            version: "current",
+            title: "NCBI human gene identifiers",
+            description: "Entrez Gene records for Homo sapiens (NCBI taxonomy 9606): identifiers, approved symbols, synonyms, and cross-references (Ensembl, HGNC) in the dbXrefs column. Tab-separated, gzipped. NCBI rebuilds this file in place, so it is verified against what you downloaded rather than a checked-in digest.",
+            sourceUrl: "https://ftp.ncbi.nlm.nih.gov/gene/DATA/GENE_INFO/Mammalia/",
+            license: {
+                identifier: "NCBI-Molecular-Data-Usage-Policy",
+                url: "https://www.ncbi.nlm.nih.gov/home/about/policies/",
+            },
+            recommendation: { group: "gene-identifiers", recommended: true },
+            artifacts: [
+                {
+                    integrity: "unpinned",
+                    path: "Homo_sapiens.gene_info.gz",
+                    url: "https://ftp.ncbi.nlm.nih.gov/gene/DATA/GENE_INFO/Mammalia/Homo_sapiens.gene_info.gz",
+                },
+            ],
+        },
+        {
+            id: "ncbi-gene-mouse",
+            version: "current",
+            title: "NCBI mouse gene identifiers",
+            description: "Entrez Gene records for Mus musculus (NCBI taxonomy 10090): identifiers, approved symbols, synonyms, and cross-references (Ensembl, MGI) in the dbXrefs column. Tab-separated, gzipped. NCBI rebuilds this file in place, so it is verified against what you downloaded rather than a checked-in digest.",
+            sourceUrl: "https://ftp.ncbi.nlm.nih.gov/gene/DATA/GENE_INFO/Mammalia/",
+            license: {
+                identifier: "NCBI-Molecular-Data-Usage-Policy",
+                url: "https://www.ncbi.nlm.nih.gov/home/about/policies/",
+            },
+            recommendation: { group: "gene-identifiers", recommended: true },
+            artifacts: [
+                {
+                    integrity: "unpinned",
+                    path: "Mus_musculus.gene_info.gz",
+                    url: "https://ftp.ncbi.nlm.nih.gov/gene/DATA/GENE_INFO/Mammalia/Mus_musculus.gene_info.gz",
+                },
+            ],
+        },
+        {
+            id: "ncbi-gene-rat",
+            version: "current",
+            title: "NCBI rat gene identifiers",
+            description: "Entrez Gene records for Rattus norvegicus (NCBI taxonomy 10116): identifiers, approved symbols, synonyms, and cross-references (Ensembl, RGD) in the dbXrefs column. Tab-separated, gzipped. NCBI rebuilds this file in place, so it is verified against what you downloaded rather than a checked-in digest.",
+            sourceUrl: "https://ftp.ncbi.nlm.nih.gov/gene/DATA/GENE_INFO/Mammalia/",
+            license: {
+                identifier: "NCBI-Molecular-Data-Usage-Policy",
+                url: "https://www.ncbi.nlm.nih.gov/home/about/policies/",
+            },
+            recommendation: { group: "gene-identifiers", recommended: true },
+            artifacts: [
+                {
+                    integrity: "unpinned",
+                    path: "Rattus_norvegicus.gene_info.gz",
+                    url: "https://ftp.ncbi.nlm.nih.gov/gene/DATA/GENE_INFO/Mammalia/Rattus_norvegicus.gene_info.gz",
+                },
+            ],
+        },
+        {
+            id: "reactome-pathways",
+            version: "current",
+            title: "Reactome pathways",
+            description: "Reactome pathway gene sets (GMT) and the pathway index — the enrichment workhorse, ~2 MB. Reactome overwrites its `current` release each quarter and removes the prior one, so this is verified against what you downloaded rather than a checked-in digest. For identifier-to-pathway mapping tables, add `reactome-mappings`.",
+            sourceUrl: "https://reactome.org/download-data",
+            license: { identifier: "CC0-1.0", url: "https://reactome.org/license" },
+            recommendation: { group: "pathways", recommended: true },
+            artifacts: [
+                { integrity: "unpinned", path: "ReactomePathways.gmt", url: "https://reactome.org/download/current/ReactomePathways.gmt" },
+                { integrity: "unpinned", path: "ReactomePathways.txt", url: "https://reactome.org/download/current/ReactomePathways.txt" },
+            ],
+        },
+        {
+            // Split out of `reactome-pathways` and NOT recommended by default: upstream serves these
+            // as uncompressed TSV and they total roughly 700 MB, two orders of magnitude more than the
+            // GMT most workflows actually want. Anyone who needs Entrez/Ensembl/UniProt -> pathway
+            // joins opts in deliberately, rather than every `setup` paying for them by accident.
+            id: "reactome-mappings",
+            version: "current",
+            title: "Reactome identifier mappings (large)",
+            description: "Entrez/Ensembl/UniProt-to-Reactome mapping tables at all pathway levels. Large: roughly 700 MB of uncompressed TSV, because Reactome publishes no compressed form. Only needed to join external identifiers onto Reactome pathways; the gene sets themselves live in `reactome-pathways`.",
+            sourceUrl: "https://reactome.org/download-data",
+            license: { identifier: "CC0-1.0", url: "https://reactome.org/license" },
+            recommendation: { group: "pathways", recommended: false },
+            artifacts: [
+                {
+                    integrity: "unpinned",
+                    path: "NCBI2Reactome_All_Levels.txt",
+                    url: "https://reactome.org/download/current/NCBI2Reactome_All_Levels.txt",
+                },
+                {
+                    integrity: "unpinned",
+                    path: "Ensembl2Reactome_All_Levels.txt",
+                    url: "https://reactome.org/download/current/Ensembl2Reactome_All_Levels.txt",
+                },
+                {
+                    integrity: "unpinned",
+                    path: "UniProt2Reactome_All_Levels.txt",
+                    url: "https://reactome.org/download/current/UniProt2Reactome_All_Levels.txt",
+                },
+            ],
+        },
+        {
+            id: "wikipathways-human",
+            version: "2026.07.10",
+            title: "WikiPathways human pathways",
+            description: "Community-curated Homo sapiens pathway gene sets (GMT) from the immutable 2026-07-10 WikiPathways snapshot.",
+            sourceUrl: "https://data.wikipathways.org/20260710/gmt/",
+            license: { identifier: "CC0-1.0", url: "https://www.wikipathways.org/about/terms.html" },
+            recommendation: { group: "pathways", recommended: true },
+            artifacts: [
+                {
+                    integrity: "pinned",
+                    path: "wikipathways_Homo_sapiens.gmt",
+                    url: "https://data.wikipathways.org/20260710/gmt/wikipathways-20260710-gmt-Homo_sapiens.gmt",
+                    bytes: 341_474,
+                    sha256: "79615a079246bb0b07cc3505265b1f75ea6cffec88001ce27f644dd86a39c97d",
+                },
+            ],
+        },
+        {
+            id: "collectri-human",
+            version: "2.0",
+            title: "CollecTRI human regulatory network",
+            description: "A literature-curated human transcription-factor target interaction network (CSV) from the immutable Zenodo record 8192729.",
+            sourceUrl: "https://zenodo.org/records/8192729",
+            license: { identifier: "CC-BY-4.0", url: "https://zenodo.org/records/8192729" },
+            recommendation: { group: "regulatory-networks", recommended: true },
+            artifacts: [
+                {
+                    integrity: "pinned",
+                    path: "CollecTRI_regulons.csv",
+                    url: "https://zenodo.org/records/8192729/files/CollecTRI_regulons.csv",
+                    bytes: 4_345_649,
+                    sha256: "4473c9189dd53dacc80297709ad1452dda1086a1cc2185f9a56146c261668701",
+                },
+            ],
+        },
+        {
+            id: "gtex-v8",
+            version: "8",
+            title: "GTEx v8 normal tissue expression",
+            description: "GTEx v8 median gene TPM by tissue (GCT) with the sample attributes table, for normal-reference workflows. Immutable v8 release files.",
+            sourceUrl: "https://gtexportal.org/home/downloads/adult-gtex/bulk_tissue_expression",
+            license: { identifier: "GTEx-Portal-Data-License", url: "https://gtexportal.org/home/license" },
+            recommendation: { group: "normal-expression", recommended: true },
+            artifacts: [
+                {
+                    integrity: "pinned",
+                    path: "GTEx_Analysis_2017-06-05_v8_RNASeQCv1.1.9_gene_median_tpm.gct.gz",
+                    url: "https://storage.googleapis.com/adult-gtex/bulk-gex/v8/rna-seq/GTEx_Analysis_2017-06-05_v8_RNASeQCv1.1.9_gene_median_tpm.gct.gz",
+                    bytes: 6_952_331,
+                    sha256: "ee7201ff2f280b0de5657d4b08e9a240362d9757efed7f7bd5dba35a5f8617b8",
+                },
+                {
+                    integrity: "pinned",
+                    path: "GTEx_Analysis_v8_Annotations_SampleAttributesDS.txt",
+                    url: "https://storage.googleapis.com/adult-gtex/annotations/v8/metadata-files/GTEx_Analysis_v8_Annotations_SampleAttributesDS.txt",
+                    bytes: 11_512_258,
+                    sha256: "74f6ab4c34ed2648d708a0ae6e6dff324f6c86ea723ae7d1c37d76f5221148f0",
+                },
+            ],
+        },
+        {
+            id: "celltypist-immune",
+            version: "2",
+            title: "CellTypist immune cell models",
+            description: "CellTypist Immune All low- and high-resolution models for immune cell-type annotation. Load by absolute path — CellTypist's by-name lookup expects its own cache layout, which the reference store deliberately does not impersonate.",
+            sourceUrl: "https://www.celltypist.org/models",
+            license: { identifier: "NOASSERTION" },
+            recommendation: { group: "cell-typing", recommended: true },
+            artifacts: [
+                {
+                    integrity: "pinned",
+                    path: "Immune_All_Low.pkl",
+                    url: "https://celltypist.cog.sanger.ac.uk/models/Pan_Immune_CellTypist/v2/Immune_All_Low.pkl",
+                    bytes: 2_824_990,
+                    sha256: "290874d35dac039d4c9218c343fde4aac1077709b72a331ce7266f6828c36502",
+                },
+                {
+                    integrity: "pinned",
+                    path: "Immune_All_High.pkl",
+                    url: "https://celltypist.cog.sanger.ac.uk/models/Pan_Immune_CellTypist/v2/Immune_All_High.pkl",
+                    bytes: 1_070_426,
+                    sha256: "a715fec36c2c421f7c4e31cf4cb4bea883eedab7fd7b20a4b76f091c22660448",
+                },
+            ],
+        },
+    ],
+}));
+/** A requested dataset id was not present in the selected catalog. */
+export class UnknownReferenceDatasetError extends Error {
+    code = "unknown_reference_dataset";
+    unknownId;
+    availableIds;
+    constructor(unknownId, availableIds) {
+        super(`Unknown reference dataset "${unknownId}". Available ids: ${availableIds.join(", ") || "(none)"}`);
+        this.name = "UnknownReferenceDatasetError";
+        this.unknownId = unknownId;
+        this.availableIds = availableIds;
+    }
+}
+/** Resolve dataset ids without I/O, preserving a stable dataset-id ordering. */
+/** @internal Fixture seam for harness contract tests; not exported by the package barrel. */
+export function resolveReferenceInstallPlanFromCatalogForTesting(datasetIds, catalog) {
+    const byId = new Map(catalog.datasets.map((dataset) => [dataset.id, dataset]));
+    const availableIds = [...byId.keys()].sort();
+    const selectedIds = [...new Set(datasetIds)].sort();
+    for (const id of selectedIds) {
+        if (!byId.has(id))
+            return err(new UnknownReferenceDatasetError(id, availableIds));
+    }
+    return ok({
+        catalogVersion: catalog.version,
+        datasets: selectedIds.map((id) => {
+            const dataset = byId.get(id);
+            return { ...dataset, installPath: `${dataset.id}/${dataset.version}` };
+        }),
+    });
+}
+/** Resolve ids strictly against the immutable canonical catalog shipped by this package. */
+export function resolveReferenceInstallPlan(datasetIds) {
+    return resolveReferenceInstallPlanFromCatalogForTesting(datasetIds, REFERENCE_DATA_CATALOG);
+}

package/dist/reference-data/receipt.d.ts

@@ -0,0 +1,41 @@
+import { z } from "zod";
+/** Schema version for installation receipts shared by all embedders. */
+export declare const REFERENCE_INSTALL_RECEIPT_VERSION: 1;
+/**
+ * What was actually written to disk for one artifact. The size and digest are
+ * *observed at install*, never copied from the catalog — for a `pinned`
+ * artifact they equal the catalog's (the install would have failed otherwise),
+ * and for an `unpinned` one they are the only record of what the mutable
+ * upstream served, which is what `verify` later checks the files against.
+ */
+declare const ReferenceReceiptArtifactSchema: z.ZodObject<{
+    path: z.ZodString;
+    bytes: z.ZodNumber;
+    sha256: z.ZodString;
+    integrity: z.ZodEnum<{
+        pinned: "pinned";
+        unpinned: "unpinned";
+    }>;
+}, z.core.$strip>;
+/** Optional metadata describing one activated reference dataset version. */
+export declare const ReferenceInstallReceiptSchema: z.ZodObject<{
+    version: z.ZodLiteral<1>;
+    datasetId: z.ZodString;
+    datasetVersion: z.ZodString;
+    activatedAt: z.ZodISODateTime;
+    artifacts: z.ZodArray<z.ZodObject<{
+        path: z.ZodString;
+        bytes: z.ZodNumber;
+        sha256: z.ZodString;
+        integrity: z.ZodEnum<{
+            pinned: "pinned";
+            unpinned: "unpinned";
+        }>;
+    }, z.core.$strip>>;
+}, z.core.$strip>;
+export type ReferenceInstallReceipt = z.infer<typeof ReferenceInstallReceiptSchema>;
+/** One artifact as recorded by a completed installation. */
+export type ReferenceReceiptArtifact = z.infer<typeof ReferenceReceiptArtifactSchema>;
+/** Parse optional/untrusted receipt metadata; invalid receipts are ignored. */
+export declare function parseReferenceInstallReceipt(input: unknown): ReferenceInstallReceipt | undefined;
+export {};