gitnexus 1.6.9 → 1.6.10-aptos.0

package/dist/core/embeddings/http-client.js

@@ -14,13 +14,102 @@ import { CircuitOpenError, ResilientFetchExhaustedError, resilientFetch } from '
 const HTTP_TIMEOUT_MS = 30_000;
 const HTTP_MAX_RETRIES = 2;
 const HTTP_RETRY_BACKOFF_MS = 1_000;
+const HTTP_RETRY_CAP_MS = 5_000;
 const HTTP_BATCH_SIZE = 64;
 const DEFAULT_DIMS = 384;
 const HTTP_BREAKER_KEY = 'embeddings-http';
+let lastHttpRequestStartedAt;
+let httpPaceQueue = Promise.resolve();
+const parsePositiveIntegerEnv = (name, fallback, max) => {
+    const raw = process.env[name];
+    if (raw === undefined || raw === '')
+        return fallback;
+    if (!/^\d+$/u.test(raw)) {
+        throw new Error(`${name} must be a positive integer, got "${raw}"`);
+    }
+    const parsed = Number(raw);
+    if (!Number.isSafeInteger(parsed) || parsed <= 0 || parsed > max) {
+        throw new Error(`${name} must be a positive integer <= ${max}, got "${raw}"`);
+    }
+    return parsed;
+};
+const parseNonNegativeIntegerEnv = (name, fallback, max) => {
+    const raw = process.env[name];
+    if (raw === undefined || raw === '')
+        return fallback;
+    if (!/^\d+$/u.test(raw)) {
+        throw new Error(`${name} must be a non-negative integer, got "${raw}"`);
+    }
+    const parsed = Number(raw);
+    if (!Number.isSafeInteger(parsed) || parsed < 0 || parsed > max) {
+        throw new Error(`${name} must be a non-negative integer <= ${max}, got "${raw}"`);
+    }
+    return parsed;
+};
+const cancelledError = () => new DOMException('Embedding request cancelled', 'AbortError');
+const throwIfAborted = (signal) => {
+    if (signal?.aborted)
+        throw cancelledError();
+};
+const abortableSleep = (ms, signal) => {
+    throwIfAborted(signal);
+    if (ms <= 0)
+        return Promise.resolve();
+    return new Promise((resolve, reject) => {
+        const timer = setTimeout(() => {
+            signal?.removeEventListener('abort', onAbort);
+            resolve();
+        }, ms);
+        const onAbort = () => {
+            clearTimeout(timer);
+            signal?.removeEventListener('abort', onAbort);
+            reject(cancelledError());
+        };
+        signal?.addEventListener('abort', onAbort, { once: true });
+    });
+};
+const paceHttpRequest = async (minIntervalMs, signal) => {
+    throwIfAborted(signal);
+    if (minIntervalMs <= 0)
+        return;
+    const waitTurn = httpPaceQueue.then(async () => {
+        throwIfAborted(signal);
+        const waitMs = lastHttpRequestStartedAt === undefined
+            ? 0
+            : Math.max(0, lastHttpRequestStartedAt + minIntervalMs - Date.now());
+        await abortableSleep(waitMs, signal);
+        throwIfAborted(signal);
+        lastHttpRequestStartedAt = Date.now();
+    });
+    httpPaceQueue = waitTurn.catch(() => undefined);
+    await waitTurn;
+};
+/**
+ * Stable lead of a {@link readConfig} malformed dims-env error. `readConfig`
+ * throws a plain `Error` (not an {@link HttpEmbeddingError}) for a malformed
+ * `GITNEXUS_EMBEDDING_DIMS` or `GITNEXUS_EMBEDDING_REQUEST_DIMS` because it's a
+ * *config* mistake, not an endpoint failure — so the CLI recognizes it by this
+ * lead ({@link isHttpEmbeddingDimsError}) and prints a clean config message
+ * instead of a raw stack dump. Each var names itself so the message points the
+ * operator at the variable they actually set, not a sibling. See #2385.
+ */
+const dimsEnvErrorLead = (name) => `${name} must be a positive integer`;
+const EMBEDDING_DIMS_ENV_ERROR_LEAD = dimsEnvErrorLead('GITNEXUS_EMBEDDING_DIMS');
+const EMBEDDING_REQUEST_DIMS_ENV_ERROR_LEAD = dimsEnvErrorLead('GITNEXUS_EMBEDDING_REQUEST_DIMS');
+/**
+ * @internal Exported for the CLI analyze error handler. True when `message` is a
+ * {@link readConfig} malformed dims-env config error (a plain `Error`) — for
+ * either `GITNEXUS_EMBEDDING_DIMS` or `GITNEXUS_EMBEDDING_REQUEST_DIMS`.
+ */
+export const isHttpEmbeddingDimsError = (message) => message.includes(EMBEDDING_DIMS_ENV_ERROR_LEAD) ||
+    message.includes(EMBEDDING_REQUEST_DIMS_ENV_ERROR_LEAD);
 /**
  * Build config from the current process.env snapshot.
  * Returns null when GITNEXUS_EMBEDDING_URL + GITNEXUS_EMBEDDING_MODEL are unset.
  * Not cached — env vars are read fresh so late configuration takes effect.
+ * Validates GITNEXUS_EMBEDDING_DIMS and throws on a malformed value; callers
+ * that only need to know whether HTTP mode is *configured* must use
+ * {@link isHttpMode} (a presence probe that never throws), not this.
  */
 const readConfig = () => {
     const baseUrl = process.env.GITNEXUS_EMBEDDING_URL;
@@ -31,25 +120,52 @@ const readConfig = () => {
     let dimensions;
     if (rawDims !== undefined) {
         if (!/^\d+$/.test(rawDims)) {
-            throw new Error(`GITNEXUS_EMBEDDING_DIMS must be a positive integer, got "${rawDims}"`);
+            throw new Error(`${EMBEDDING_DIMS_ENV_ERROR_LEAD}, got "${rawDims}"`);
         }
         const parsed = parseInt(rawDims, 10);
         if (parsed <= 0) {
-            throw new Error(`GITNEXUS_EMBEDDING_DIMS must be a positive integer, got "${rawDims}"`);
+            throw new Error(`${EMBEDDING_DIMS_ENV_ERROR_LEAD}, got "${rawDims}"`);
         }
         dimensions = parsed;
     }
+    const rawRequestDims = process.env.GITNEXUS_EMBEDDING_REQUEST_DIMS?.trim();
+    let requestDimensions = dimensions;
+    if (rawRequestDims) {
+        if (/^(omit|none|off|false|0)$/i.test(rawRequestDims)) {
+            requestDimensions = undefined;
+        }
+        else {
+            if (!/^\d+$/.test(rawRequestDims)) {
+                throw new Error(`${EMBEDDING_REQUEST_DIMS_ENV_ERROR_LEAD}, got "${rawRequestDims}"`);
+            }
+            const parsed = parseInt(rawRequestDims, 10);
+            if (parsed <= 0) {
+                throw new Error(`${EMBEDDING_REQUEST_DIMS_ENV_ERROR_LEAD}, got "${rawRequestDims}"`);
+            }
+            requestDimensions = parsed;
+        }
+    }
     return {
         baseUrl: baseUrl.replace(/\/+$/, ''),
         model,
         apiKey: process.env.GITNEXUS_EMBEDDING_API_KEY ?? 'unused',
         dimensions,
+        maxAttempts: parsePositiveIntegerEnv('GITNEXUS_EMBEDDING_MAX_ATTEMPTS', HTTP_MAX_RETRIES + 1, 20),
+        retryCapMs: parsePositiveIntegerEnv('GITNEXUS_EMBEDDING_RETRY_CAP_MS', HTTP_RETRY_CAP_MS, 300_000),
+        minIntervalMs: parseNonNegativeIntegerEnv('GITNEXUS_EMBEDDING_MIN_INTERVAL_MS', 0, 300_000),
+        requestDimensions,
     };
 };
 /**
- * Check whether HTTP embedding mode is active (env vars are set).
+ * Whether HTTP embedding mode is active — i.e. both `GITNEXUS_EMBEDDING_URL` and
+ * `GITNEXUS_EMBEDDING_MODEL` are set. A pure presence probe: it deliberately does
+ * NOT call {@link readConfig}, so it never throws on a malformed
+ * `GITNEXUS_EMBEDDING_DIMS`. This lets its ~13 call sites (analyze, doctor,
+ * run-analyze, embedder, mcp) probe the mode without a defensive try/catch; the
+ * DIMS value is validated where it is actually used (`readConfig` in
+ * `httpEmbed`/`httpEmbedQuery`), surfacing a recognizable config error. See #2385.
  */
-export const isHttpMode = () => readConfig() !== null;
+export const isHttpMode = () => Boolean(process.env.GITNEXUS_EMBEDDING_URL && process.env.GITNEXUS_EMBEDDING_MODEL);
 /**
  * Return the configured embedding dimensions for HTTP mode, or undefined
  * if HTTP mode is not active or no explicit dimensions are set.
@@ -70,6 +186,68 @@ export const safeUrl = (url) => {
         return '<invalid-url>';
     }
 };
+/**
+ * Strip credentials from an underlying transport error message before it is
+ * surfaced. A credential-bearing endpoint URL (`https://user:secret@host/v1`)
+ * makes undici throw `TypeError: Request cannot be constructed from a URL that
+ * includes credentials: <that full URL>`; interpolating `err.message` verbatim
+ * would re-leak the secret to stderr + logs even though the URL argument is
+ * already masked with {@link safeUrl}. First swap the exact configured `url` for
+ * its masked form, then strip any residual `scheme://userinfo@` the transport may
+ * have echoed in a normalized (non-exact) form. See #2385.
+ */
+const sanitizeReason = (reason, url, apiKey) => {
+    const withoutUrlCredentials = reason
+        .split(url)
+        .join(safeUrl(url))
+        .replace(/([a-z][a-z0-9+.-]*:\/\/)[^/@\s]*@/gi, '$1');
+    return apiKey && apiKey !== 'unused'
+        ? withoutUrlCredentials.split(apiKey).join('[redacted]')
+        : withoutUrlCredentials;
+};
+/**
+ * Error thrown by this module's HTTP embedding path (`httpEmbedBatch` /
+ * `httpEmbed` / `httpEmbedQuery`) for any endpoint failure — a
+ * connection/timeout/DNS error, an open circuit, a non-OK status, an
+ * unparseable or wrong-shape response body, an empty response, or a dimension
+ * mismatch.
+ *
+ * Carrying a distinct type (rather than a plain `Error`) lets the CLI tell a
+ * *custom endpoint* failure apart from a HuggingFace *model download* failure
+ * without matching message text: the two share the same underlying network
+ * substrings (`fetch failed`, `ECONNREFUSED`, …), which is exactly why
+ * `isNetworkFetchError` in `hf-env.ts` cannot tell them apart. Keying on the
+ * type instead of the message is also locale-proof and survives message
+ * rewording. The human-readable `.message` (built with `safeUrl` and the
+ * underlying reason) is what the CLI surfaces to the user. See #2385.
+ */
+export class HttpEmbeddingError extends Error {
+    constructor(message, options) {
+        super(message, options?.cause !== undefined ? { cause: options.cause } : undefined);
+        this.name = 'HttpEmbeddingError';
+    }
+}
+/**
+ * @internal Exported for the CLI analyze error handler and unit tests.
+ *
+ * Type-guard for {@link HttpEmbeddingError}. The `name` fallback keeps the
+ * check working across module-realm boundaries where `instanceof` can fail
+ * (two loaded copies of the class) — mirroring the codebase's existing
+ * `err.name === 'TimeoutError'` idiom. Matches on the stable class
+ * discriminator, never on the human-readable (potentially localized) message.
+ */
+export const isHttpEmbeddingError = (err) => err instanceof HttpEmbeddingError || (err instanceof Error && err.name === 'HttpEmbeddingError');
+/**
+ * Runtime guard for a single response item. The `Array.isArray(data.data)` shape
+ * check only validates the outer array — a 200 body like `{"data":[null]}` passes
+ * it, then crashes at `new Float32Array(item.embedding)` (`httpEmbed`) or
+ * `items[0].embedding` (`httpEmbedQuery`) with a raw `TypeError` that escapes the
+ * typed boundary, landing on the CLI's generic stack-dump path — the exact class
+ * #2385 closes. Validate each item so every wrong-shape body stays classifiable.
+ */
+const isEmbeddingItem = (item) => typeof item === 'object' &&
+    item !== null &&
+    Array.isArray(item.embedding);
 /**
  * Send a single batch of texts to the embedding endpoint with retry.
  *
@@ -82,11 +260,11 @@ export const safeUrl = (url) => {
  *   the `dimensions` field in the request body. Endpoints that implement
  *   Matryoshka truncation (OpenAI text-embedding-3-*, Cohere embed-v3,
  *   Voyage) return a truncated vector at that size; endpoints that do not
- *   recognise the field may ignore it or return 400. Leave
- *   `GITNEXUS_EMBEDDING_DIMS` unset for strict backends that reject
- *   unknown fields.
+ *   recognise the field may ignore it or return 400. Set
+ *   `GITNEXUS_EMBEDDING_REQUEST_DIMS=omit` for strict backends while keeping
+ *   `GITNEXUS_EMBEDDING_DIMS` set to the returned vector size.
  */
-const httpEmbedBatch = async (url, batch, model, apiKey, batchIndex = 0, dimensions) => {
+const httpEmbedBatch = async (url, batch, model, apiKey, batchIndex = 0, dimensions, requestOptions = {}, maxAttempts = HTTP_MAX_RETRIES + 1, retryCapMs = HTTP_RETRY_CAP_MS, minIntervalMs = 0) => {
     const requestBody = {
         input: batch,
         model,
@@ -96,38 +274,72 @@ const httpEmbedBatch = async (url, batch, model, apiKey, batchIndex = 0, dimensi
     }
     let resp;
     try {
+        throwIfAborted(requestOptions.signal);
         resp = await resilientFetch(url, {
             method: 'POST',
-            signal: AbortSignal.timeout(HTTP_TIMEOUT_MS),
             headers: {
                 'Content-Type': 'application/json',
                 Authorization: `Bearer ${apiKey}`,
             },
             body: JSON.stringify(requestBody),
         }, {
+            fetchImpl: async (input, init) => {
+                await paceHttpRequest(minIntervalMs, requestOptions.signal);
+                throwIfAborted(requestOptions.signal);
+                const timeoutSignal = AbortSignal.timeout(HTTP_TIMEOUT_MS);
+                const signal = requestOptions.signal
+                    ? AbortSignal.any([requestOptions.signal, timeoutSignal])
+                    : timeoutSignal;
+                return globalThis.fetch(input, { ...init, signal });
+            },
             breakerKey: HTTP_BREAKER_KEY,
-            retry: { maxAttempts: HTTP_MAX_RETRIES + 1, baseDelayMs: HTTP_RETRY_BACKOFF_MS },
+            retry: {
+                maxAttempts,
+                baseDelayMs: HTTP_RETRY_BACKOFF_MS,
+                capDelayMs: retryCapMs,
+                retryAfterCapMs: retryCapMs,
+                sleep: (ms) => abortableSleep(ms, requestOptions.signal),
+            },
         });
     }
     catch (err) {
+        if (requestOptions.signal?.aborted ||
+            (err instanceof DOMException && err.name === 'AbortError')) {
+            throw new HttpEmbeddingError(`Embedding request cancelled (${safeUrl(url)}, batch ${batchIndex})`, { cause: err });
+        }
         if (err instanceof CircuitOpenError) {
-            throw new Error(`Embedding endpoint circuit open (${safeUrl(url)}, batch ${batchIndex}): retry in ${Math.ceil(err.retryAfterMs / 1000)}s`);
+            throw new HttpEmbeddingError(`Embedding endpoint circuit open (${safeUrl(url)}, batch ${batchIndex}): retry in ${Math.ceil(err.retryAfterMs / 1000)}s`, { cause: err });
         }
         if (err instanceof DOMException && err.name === 'TimeoutError') {
-            throw new Error(`Embedding request timed out after ${HTTP_TIMEOUT_MS}ms (${safeUrl(url)}, batch ${batchIndex})`);
+            throw new HttpEmbeddingError(`Embedding request timed out after ${HTTP_TIMEOUT_MS}ms (${safeUrl(url)}, batch ${batchIndex})`, { cause: err });
         }
         if (err instanceof ResilientFetchExhaustedError) {
-            throw new Error(`Embedding endpoint returned ${err.response.status} (${safeUrl(url)}, batch ${batchIndex})`);
+            throw new HttpEmbeddingError(`Embedding endpoint returned ${err.response.status} (${safeUrl(url)}, batch ${batchIndex})`, { cause: err });
         }
-        const reason = err instanceof Error ? err.message : String(err);
-        throw new Error(`Embedding request failed (${safeUrl(url)}, batch ${batchIndex}): ${reason}`);
+        const reason = sanitizeReason(err instanceof Error ? err.message : String(err), url, apiKey);
+        const safeCause = new Error(reason);
+        safeCause.name = err instanceof Error ? err.name : 'EmbeddingTransportError';
+        throw new HttpEmbeddingError(`Embedding request failed (${safeUrl(url)}, batch ${batchIndex}): ${reason}`, { cause: safeCause });
     }
     if (!resp.ok) {
         // resilientFetch already retried 5xx/429; any non-OK response here is
         // a terminal client error (4xx other than 429).
-        throw new Error(`Embedding endpoint returned ${resp.status} (${safeUrl(url)}, batch ${batchIndex})`);
+        throw new HttpEmbeddingError(`Embedding endpoint returned ${resp.status} (${safeUrl(url)}, batch ${batchIndex})`);
+    }
+    // A reachable-but-wrong endpoint (e.g. a captive portal or a non-embeddings
+    // service) can answer 200 with an HTML/truncated body. Parse inside the
+    // typed-error boundary so that lands as an endpoint failure the CLI can
+    // classify, not a raw SyntaxError/TypeError on the generic stack-dump path.
+    let data;
+    try {
+        data = (await resp.json());
+    }
+    catch (err) {
+        throw new HttpEmbeddingError(`Embedding endpoint returned an unparseable response (${safeUrl(url)}, batch ${batchIndex})`, { cause: err });
+    }
+    if (!Array.isArray(data?.data) || !data.data.every(isEmbeddingItem)) {
+        throw new HttpEmbeddingError(`Embedding endpoint returned an unexpected response shape (${safeUrl(url)}, batch ${batchIndex})`);
     }
-    const data = (await resp.json());
     return data.data;
 };
 /**
@@ -137,7 +349,7 @@ const httpEmbedBatch = async (url, batch, model, apiKey, batchIndex = 0, dimensi
  * @param texts - Array of texts to embed
  * @returns Array of Float32Array embedding vectors
  */
-export const httpEmbed = async (texts) => {
+export const httpEmbed = async (texts, requestOptions = {}) => {
     if (texts.length === 0)
         return [];
     const config = readConfig();
@@ -148,9 +360,9 @@ export const httpEmbed = async (texts) => {
     for (let i = 0; i < texts.length; i += HTTP_BATCH_SIZE) {
         const batch = texts.slice(i, i + HTTP_BATCH_SIZE);
         const batchIndex = Math.floor(i / HTTP_BATCH_SIZE);
-        const items = await httpEmbedBatch(url, batch, config.model, config.apiKey, batchIndex, config.dimensions);
+        const items = await httpEmbedBatch(url, batch, config.model, config.apiKey, batchIndex, config.requestDimensions, requestOptions, config.maxAttempts, config.retryCapMs, config.minIntervalMs);
         if (items.length !== batch.length) {
-            throw new Error(`Embedding endpoint returned ${items.length} vectors for ${batch.length} texts ` +
+            throw new HttpEmbeddingError(`Embedding endpoint returned ${items.length} vectors for ${batch.length} texts ` +
                 `(${safeUrl(url)}, batch ${batchIndex})`);
         }
         for (const item of items) {
@@ -162,7 +374,7 @@ export const httpEmbed = async (texts) => {
                 const hint = config.dimensions
                     ? 'Update GITNEXUS_EMBEDDING_DIMS to match your model output.'
                     : `Set GITNEXUS_EMBEDDING_DIMS=${vec.length} to match your model output.`;
-                throw new Error(`Embedding dimension mismatch: endpoint returned ${vec.length}d vector, ` +
+                throw new HttpEmbeddingError(`Embedding dimension mismatch: endpoint returned ${vec.length}d vector, ` +
                     `but expected ${expected}d. ${hint}`);
             }
             allVectors.push(vec);
@@ -177,14 +389,14 @@ export const httpEmbed = async (texts) => {
  * @param text - Query text to embed
  * @returns Embedding vector as number array
  */
-export const httpEmbedQuery = async (text) => {
+export const httpEmbedQuery = async (text, requestOptions = {}) => {
     const config = readConfig();
     if (!config)
         throw new Error('HTTP embedding not configured');
     const url = `${config.baseUrl}/embeddings`;
-    const items = await httpEmbedBatch(url, [text], config.model, config.apiKey, 0, config.dimensions);
+    const items = await httpEmbedBatch(url, [text], config.model, config.apiKey, 0, config.requestDimensions, requestOptions, config.maxAttempts, config.retryCapMs, config.minIntervalMs);
     if (!items.length) {
-        throw new Error(`Embedding endpoint returned empty response (${safeUrl(url)})`);
+        throw new HttpEmbeddingError(`Embedding endpoint returned empty response (${safeUrl(url)})`);
     }
     const embedding = items[0].embedding;
     // Same dimension checks as httpEmbed — catch mismatches before they
@@ -194,7 +406,7 @@ export const httpEmbedQuery = async (text) => {
         const hint = config.dimensions
             ? 'Update GITNEXUS_EMBEDDING_DIMS to match your model output.'
             : `Set GITNEXUS_EMBEDDING_DIMS=${embedding.length} to match your model output.`;
-        throw new Error(`Embedding dimension mismatch: endpoint returned ${embedding.length}d vector, ` +
+        throw new HttpEmbeddingError(`Embedding dimension mismatch: endpoint returned ${embedding.length}d vector, ` +
             `but expected ${expected}d. ${hint}`);
     }
     return embedding;

package/dist/core/embeddings/node-module-compat.d.ts

@@ -0,0 +1,25 @@
+/**
+ * The single access point for `module.registerHooks` (#2372).
+ *
+ * `module.registerHooks` — the synchronous ESM/CJS resolution-hook API the
+ * embedding-stack resolvers rely on — was added in Node 22.15.0 (and 23.5.0 on
+ * the 23.x line). The gitnexus engines floor is `^22.18.0 || >=24.11.0`, so
+ * every supported runtime exposes it — but `engines` is advisory (not
+ * engine-strict), so a below-floor Node (22.0–22.14, or the unsupported
+ * 23.0–23.4 line) can still run, where the export is absent.
+ *
+ * In this `"type": "module"` package, a *static named* import of a missing
+ * builtin export (`import { registerHooks } from 'node:module'`) is a
+ * `SyntaxError` at ESM link time — thrown before any `typeof registerHooks`
+ * guard in the module body can run, so every module carrying that import fails
+ * to load on those Node versions. This module owns the only namespace import of
+ * `node:module` and hands callers a value-or-`undefined` they guard at runtime,
+ * so the graceful-degradation path is finally reachable.
+ *
+ * `@types/node` types `registerHooks` as always-present, so `nodeModule.registerHooks`
+ * would type as defined while being `undefined` at runtime on older Node. The
+ * `Partial` narrow surfaces the real optionality without an `any` cast.
+ */
+import * as nodeModule from 'node:module';
+/** `module.registerHooks` if this Node exposes it (>=22.15 / >=23.5), else `undefined`. */
+export declare const getRegisterHooks: () => typeof nodeModule.registerHooks | undefined;

package/dist/core/embeddings/node-module-compat.js

@@ -0,0 +1,25 @@
+/**
+ * The single access point for `module.registerHooks` (#2372).
+ *
+ * `module.registerHooks` — the synchronous ESM/CJS resolution-hook API the
+ * embedding-stack resolvers rely on — was added in Node 22.15.0 (and 23.5.0 on
+ * the 23.x line). The gitnexus engines floor is `^22.18.0 || >=24.11.0`, so
+ * every supported runtime exposes it — but `engines` is advisory (not
+ * engine-strict), so a below-floor Node (22.0–22.14, or the unsupported
+ * 23.0–23.4 line) can still run, where the export is absent.
+ *
+ * In this `"type": "module"` package, a *static named* import of a missing
+ * builtin export (`import { registerHooks } from 'node:module'`) is a
+ * `SyntaxError` at ESM link time — thrown before any `typeof registerHooks`
+ * guard in the module body can run, so every module carrying that import fails
+ * to load on those Node versions. This module owns the only namespace import of
+ * `node:module` and hands callers a value-or-`undefined` they guard at runtime,
+ * so the graceful-degradation path is finally reachable.
+ *
+ * `@types/node` types `registerHooks` as always-present, so `nodeModule.registerHooks`
+ * would type as defined while being `undefined` at runtime on older Node. The
+ * `Partial` narrow surfaces the real optionality without an `any` cast.
+ */
+import * as nodeModule from 'node:module';
+/** `module.registerHooks` if this Node exposes it (>=22.15 / >=23.5), else `undefined`. */
+export const getRegisterHooks = () => nodeModule.registerHooks;

package/dist/core/embeddings/onnxruntime-common-resolver.js

@@ -52,16 +52,17 @@
  * per-resolution cost is a single string comparison.
  *
  * `module.registerHooks` is marked `@experimental` and requires Node >= 22.15
- * (the gitnexus engines floor is >= 22.0.0). On older runtimes it is absent and
- * this is a graceful no-op: embeddings then resolve onnxruntime-common exactly
+ * (below the gitnexus engines floor of `^22.18.0 || >=24.11.0`). On below-floor
+ * runtimes it is absent and this is a graceful no-op: embeddings then resolve onnxruntime-common exactly
  * as before — fine on hoisted layouts. Any failure during installation is
  * swallowed.
  */
-import { registerHooks, createRequire } from 'node:module';
+import { createRequire } from 'node:module';
 import { pathToFileURL } from 'node:url';
 import { join } from 'node:path';
 import { getEffectiveOnnxRuntimeNodeDir } from './onnxruntime-node-resolver.js';
 import { logger } from '../logger.js';
+import { getRegisterHooks } from './node-module-compat.js';
 let attempted = false;
 /**
  * Compute the file: URL the hook redirects `onnxruntime-common` to.
@@ -98,8 +99,10 @@ export const ensureOnnxRuntimeCommonResolvable = () => {
     // initEmbedder() call, and the hook is process-global — once is enough.
     attempted = true;
     try {
-        // Node < 22.15 (the gitnexus engines floor is >= 22.0.0): no synchronous
-        // hooks API. Degrade gracefully — the import still works on hoisted layouts.
+        // Node < 22.15 / < 23.5 (below the gitnexus engines floor of
+        // ^22.18.0 || >=24.11.0): no synchronous hooks API. Degrade gracefully —
+        // the import still works on hoisted layouts.
+        const registerHooks = getRegisterHooks();
         if (typeof registerHooks !== 'function')
             return;
         const redirectUrl = resolveOnnxRuntimeCommonUrl();

package/dist/core/embeddings/onnxruntime-node-resolver.js

@@ -36,8 +36,8 @@
  * So CUDA-12 hosts, Windows (DirectML), macOS, and CPU-only hosts are
  * untouched. Idempotent; any failure is swallowed and leaves the default
  * resolution exactly as before. `module.registerHooks` requires Node >= 22.15
- * (the gitnexus engines floor is >= 22.0.0); on older runtimes the redirect is
- * a no-op, but the default copy's CUDA major is still probed so an
+ * (below the gitnexus engines floor of `^22.18.0 || >=24.11.0`); on below-floor
+ * runtimes the redirect is a no-op, but the default copy's CUDA major is still probed so an
  * already-matching host (e.g. CUDA 12 + transformers' CUDA-12 build) keeps
  * auto-selecting the GPU.
  * `npm link` / symlinked local-dev checkouts are a known caveat: `resolveOurOrtNodeDir`/
@@ -51,12 +51,14 @@
  * loaded (the probe uses CJS `require.resolve`, which an ESM hook does not
  * affect) — keeping probe and runtime consistent.
  */
-import { registerHooks, createRequire } from 'node:module';
+import { createRequire } from 'node:module';
 import { pathToFileURL } from 'node:url';
 import { existsSync } from 'node:fs';
 import { join, dirname } from 'node:path';
 import { execFileSync } from 'node:child_process';
 import { logger } from '../logger.js';
+import { getEmbeddingRuntimeDir } from './runtime-install.js';
+import { getRegisterHooks } from './node-module-compat.js';
 const require = createRequire(import.meta.url);
 /**
  * Read a shared object's NEEDED entries, tolerating ldd's non-zero exit when a
@@ -152,7 +154,18 @@ const resolveDefaultOrtNodeDir = () => {
         return dirname(createRequire(transformersMain).resolve('onnxruntime-node/package.json'));
     }
     catch {
-        return null;
+        // On-demand runtime prefix (#2370): when the optional stack was pruned at
+        // install time and fetched on demand, the copy that actually loads (via
+        // ensureEmbeddingStackResolvable's fallback hook) lives in the prefix — so
+        // it IS the effective default and must be the one the CUDA probe inspects.
+        try {
+            const prefixRequire = createRequire(join(getEmbeddingRuntimeDir(), 'noop.js'));
+            const transformersMain = prefixRequire.resolve('@huggingface/transformers');
+            return dirname(createRequire(transformersMain).resolve('onnxruntime-node/package.json'));
+        }
+        catch {
+            return null;
+        }
     }
 };
 /** gitnexus' own direct top-level onnxruntime-node dir. */
@@ -161,7 +174,19 @@ const resolveOurOrtNodeDir = () => {
         return dirname(require.resolve('onnxruntime-node/package.json'));
     }
     catch {
-        return null;
+        // On-demand runtime prefix (#2370): when gitnexus' own onnxruntime-node was
+        // pruned at install time and fetched on demand, the prefix copy IS our
+        // effective top-level build — so the CUDA-major redirect must be able to
+        // target it (mirrors resolveDefaultOrtNodeDir's fallback above). Without
+        // this, `embeddings install --cuda` on a pruned install downloads the GPU
+        // binaries but the probe still can't see them and embeddings run on CPU.
+        try {
+            const prefixRequire = createRequire(join(getEmbeddingRuntimeDir(), 'noop.js'));
+            return dirname(prefixRequire.resolve('onnxruntime-node/package.json'));
+        }
+        catch {
+            return null;
+        }
     }
 };
 let cached = null;
@@ -176,7 +201,7 @@ const decide = () => {
     // major is still probed: a CUDA-12 host on Node 22.0–22.14 whose default
     // build already matches must keep auto-selecting the GPU exactly as it did
     // before this redirect existed.
-    const canRedirect = typeof registerHooks === 'function';
+    const canRedirect = typeof getRegisterHooks() === 'function';
     const systemMajor = detectSystemCudaMajor();
     // `defaultDir` resolving is NOT a precondition for checking `ourDir` below —
     // if transformers' own resolution fails outright (defaultMajor stays null),
@@ -264,8 +289,9 @@ export const ensureOnnxRuntimeNodeMatchesSystem = () => {
         return;
     attempted = true;
     try {
+        const registerHooks = getRegisterHooks();
         if (typeof registerHooks !== 'function')
-            return; // Node < 22.15: graceful no-op
+            return; // Node < 22.15 / < 23.5: graceful no-op
         const d = decide();
         if (!d.redirect || !d.effectiveDir)
             return;