@bendyline/gezel-inference 1.0.5 → 1.0.6

package/dist/catalog/provenance.json

@@ -1,6 +1,6 @@
 {
   "package": "@bendyline/gilde",
-  "version": "0.3.0",
+  "version": "0.3.1",
   "kinds": [
     "chat-models",
     "image-models",

package/dist/engines/resolver.d.ts

@@ -32,7 +32,7 @@ export interface ResolveEngineOptions {
     signaturePolicy?: SignaturePolicy;
     fetchImpl?: typeof fetch;
     signal?: AbortSignal;
-    /** Pre-resolved token: `undefined` = auto-resolve; `null` = explicitly none. */
+    /** Pre-resolved token: `undefined` = from the environment; `null` = explicitly none. */
     token?: string | null;
     /** Override the source-pinned SHA256SUMS digest (tests). */
     expectedSha256sumsDigest?: string;

package/dist/engines/resolver.js

@@ -3,7 +3,7 @@ import { createReadStream, existsSync, readdirSync } from 'node:fs';
 import { chmod, mkdir, readFile, rename, rm, writeFile } from 'node:fs/promises';
 import { dirname, join } from 'node:path';
 import { HttpStatusError, createLogger, nowIso, resolveDistributionProfile, retryTransient, } from '@bendyline/gezel';
-import { DUCKDB_APPLE_TEAM_ID, DUCKDB_ARCHIVE_SHA256, DUCKDB_ASSET, DUCKDB_BINARY_SHA256, DUCKDB_VERSION, DUCKDB_WINDOWS_PUBLISHER, MXC_EXECUTOR_NAME, MXC_EXECUTOR_SHA256, MXC_LICENSE_SHA256, MXC_TARBALL_INTEGRITY, MXC_TARBALL_LICENSE_ENTRY, MXC_TARBALL_SHA256, MXC_TARBALL_URL, MXC_VERIFIED_SENTINEL, MXC_VERSION, MXC_WINDOWS_PUBLISHER, duckdbAssetUrl, duckdbBinaryName, duckdbInstallDir, duckdbPlatformKey, mxcArch, mxcInstallDir, mxcTarballEntry, resolvePlatformKey, windowsHeadlessSpawnOptions, } from '@bendyline/gezel/native';
+import { DUCKDB_APPLE_TEAM_ID, DUCKDB_ARCHIVE_SHA256, DUCKDB_ASSET, DUCKDB_BINARY_SHA256, DUCKDB_VERSION, DUCKDB_WINDOWS_PUBLISHER, MXC_EXECUTOR_NAME, MXC_EXECUTOR_SHA256, MXC_LICENSE_SHA256, MXC_TARBALL_INTEGRITY, MXC_TARBALL_LICENSE_ENTRY, MXC_TARBALL_SHA256, MXC_TARBALL_URL, MXC_VERIFIED_SENTINEL, MXC_VERSION, MXC_WINDOWS_PUBLISHER, duckdbAssetUrl, duckdbBinaryName, duckdbInstallDir, duckdbPlatformKey, mxcArch, mxcInstallDir, mxcTarballEntry, resolvePlatformKey, } from '@bendyline/gezel/native';
 import * as tar from 'tar';
 import { downloadWithRetry, friendlyErrors, } from '../utils/download-with-retry.js';
 import { extractZip } from '../utils/extract-zip.js';
@@ -541,14 +541,16 @@ export async function* resolveEngine(opts) {
     // `bendyline/gezel` is public, so release metadata and assets are readable
     // anonymously. A token stays OPTIONAL and is used when present purely to
     // lift GitHub's 60-request/hour unauthenticated API rate limit. Do not
-    // reinstate a hard gate here: users who install from npm have no `gh`
-    // login, and requiring one would make on-device engines unreachable for
-    // every consumer outside this repo.
-    const token = opts.token !== undefined ? opts.token : await resolveGitHubToken();
+    // reinstate a hard gate here: users who install from npm have no token,
+    // and requiring one would make on-device engines unreachable for every
+    // consumer outside this repo.
+    const auth = {
+        token: opts.token !== undefined ? opts.token : resolveGitHubToken(),
+    };
     const baseFetch = opts.fetchImpl ?? globalThis.fetch;
     const tag = `native-v${version}`;
     // ── 1. Resolve release + assets ──
-    const release = await ghJson(`${apiBase}/repos/${REPO}/releases/tags/${tag}`, baseFetch, token);
+    const release = await ghJson(`${apiBase}/repos/${REPO}/releases/tags/${tag}`, baseFetch, auth);
     const archiveAsset = release.assets?.find((a) => a.name === archiveName);
     const sumsAsset = release.assets?.find((a) => a.name === 'SHA256SUMS');
     if (!archiveAsset) {
@@ -558,7 +560,7 @@ export async function* resolveEngine(opts) {
         yield* fail(`release ${tag} is missing its SHA256SUMS manifest — refusing to install unverified`);
     }
     // ── 2. SHA256SUMS: verify own digest against the source pin, then parse ──
-    const sumsBytes = await ghBytes(sumsAsset.url, baseFetch, token);
+    const sumsBytes = await ghBytes(sumsAsset.url, baseFetch, auth);
     const sumsDigest = createHash('sha256').update(sumsBytes).digest('hex');
     if (!unpinned && sumsDigest !== expectedDigest) {
         yield* fail(`SHA256SUMS digest mismatch for ${tag}: this build pins ${expectedDigest.slice(0, 16)}… but the release served ${sumsDigest.slice(0, 16)}…. Refusing — the release may have been tampered with or re-cut.`);
@@ -584,7 +586,7 @@ export async function* resolveEngine(opts) {
         url: archiveAsset.url,
         destPath: archivePath,
         approxSizeBytes: archiveAsset.size ?? 0,
-        fetchImpl: makeAuthedFetch(baseFetch, token),
+        fetchImpl: gitHubFetch(baseFetch, auth, 'application/octet-stream'),
         ...(opts.signal ? { signal: opts.signal } : {}),
     });
     let outcome;
@@ -765,35 +767,37 @@ function stampEnv(engine, binPath, variant, versionDir) {
         process.env.GEZEL_SQLITE_VEC_PATH = sqliteVec;
     }
 }
-async function resolveGitHubToken() {
-    if (process.env.GEZEL_GITHUB_TOKEN)
-        return process.env.GEZEL_GITHUB_TOKEN;
-    if (process.env.GITHUB_TOKEN)
-        return process.env.GITHUB_TOKEN;
-    try {
-        const { execFileSync } = await import('node:child_process');
-        // Named host: a bare `gh auth token` answers for gh's default host, which
-        // can be a GitHub Enterprise server whose token must not go to github.com.
-        const out = execFileSync('gh', ['auth', 'token', '--hostname', 'github.com'], {
-            encoding: 'utf8',
-            stdio: ['ignore', 'pipe', 'ignore'],
-            ...windowsHeadlessSpawnOptions(),
-        });
-        return out.trim() || null;
-    }
-    catch {
-        return null;
-    }
+/**
+ * An optional token, from the environment only. This module runs inside apps
+ * that embed Gezel, so it never borrows the user's GitHub CLI login: `gh` is
+ * not its credential to use, and a stale one used to fail public downloads.
+ */
+function resolveGitHubToken() {
+    return process.env.GEZEL_GITHUB_TOKEN || process.env.GITHUB_TOKEN || null;
 }
-function makeAuthedFetch(base, token) {
-    return ((input, init) => base(input, {
-        ...init,
-        headers: {
-            ...init?.headers,
-            Accept: 'application/octet-stream',
-            ...(token ? { Authorization: `Bearer ${token}` } : {}),
-        },
-    }));
+/**
+ * Fetch from GitHub with the optional token. The release is public, so a token
+ * GitHub rejects (expired, revoked) is dropped for the rest of the resolve and
+ * the request repeats anonymously instead of failing the install.
+ */
+function gitHubFetch(base, auth, accept) {
+    return (async (input, init) => {
+        const send = () => base(input, {
+            ...init,
+            headers: {
+                ...init?.headers,
+                Accept: accept,
+                ...(auth.token ? { Authorization: `Bearer ${auth.token}` } : {}),
+            },
+        });
+        const res = await send();
+        if (res.status !== 401 || !auth.token)
+            return res;
+        await res.body?.cancel().catch(() => { });
+        log.warn('[engine-resolver] GitHub rejected the token from GEZEL_GITHUB_TOKEN / GITHUB_TOKEN (401); continuing anonymously');
+        auth.token = null;
+        return send();
+    });
 }
 /**
  * The archive itself goes through `downloadWithRetry`, but the two small
@@ -808,31 +812,23 @@ function ghRetryLog(what) {
         log.warn(`[engine-resolver] ${what} failed (${info.reason}); retrying in ${info.delayMs}ms (attempt ${info.attempt}/${info.maxAttempts})`);
     };
 }
-async function ghJson(url, fetchImpl, token) {
+async function ghJson(url, fetchImpl, auth) {
+    const get = gitHubFetch(fetchImpl, auth, 'application/vnd.github+json');
     return retryTransient(async () => {
-        const res = await fetchImpl(url, {
-            headers: {
-                Accept: 'application/vnd.github+json',
-                ...(token ? { Authorization: `Bearer ${token}` } : {}),
-            },
-        });
+        const res = await get(url);
         if (!res.ok) {
             await res.body?.cancel().catch(() => { });
             // A transient status is thrown as HttpStatusError so retryTransient
             // takes another pass; 4xx becomes the terminal EngineUnavailableError.
-            throw new HttpStatusError(res.status, `GitHub API returned ${res.status} for ${url}${res.status === 404 ? ' (release not found — or no access while private)' : ''}`);
+            throw new HttpStatusError(res.status, `GitHub API returned ${res.status} for ${url}${res.status === 404 ? ' (release not found)' : ''}`);
         }
         return (await res.json());
     }, { attempts: GH_METADATA_ATTEMPTS, onRetry: ghRetryLog('GitHub release lookup') }).catch(rethrowAsEngineUnavailable);
 }
-async function ghBytes(assetApiUrl, fetchImpl, token) {
+async function ghBytes(assetApiUrl, fetchImpl, auth) {
+    const get = gitHubFetch(fetchImpl, auth, 'application/octet-stream');
     return retryTransient(async () => {
-        const res = await fetchImpl(assetApiUrl, {
-            headers: {
-                Accept: 'application/octet-stream',
-                ...(token ? { Authorization: `Bearer ${token}` } : {}),
-            },
-        });
+        const res = await get(assetApiUrl);
         if (!res.ok) {
             await res.body?.cancel().catch(() => { });
             throw new HttpStatusError(res.status, `asset download returned ${res.status} for ${assetApiUrl}`);

package/dist/providers/llama-cpp/provider.js

@@ -32,6 +32,13 @@ import { classifyStartupLine } from './stdout-parser.js';
 export * from '@bendyline/gezel/local-loop';
 export { isFileRepairTurn as isScenarioFileRepairTurn } from '@bendyline/gezel/local-loop';
 const log = createLogger('llama-cpp');
+/**
+ * How long shutdown waits for the slot caches to reach disk. The app gives
+ * the whole daemon 15 s to quit, and a save the engine never answers must
+ * not spend all of it: an unsaved or half-written slot only means a fresh
+ * prefill next time, because a restore that fails falls back to one.
+ */
+const SHUTDOWN_FLUSH_BUDGET_MS = 8_000;
 export class LlamaCppProvider {
     name = 'llama-cpp';
     queue;
@@ -627,12 +634,21 @@ export class LlamaCppProvider {
         // llama-server exits the slot save endpoint is gone. flushAll is
         // best-effort; failures during shutdown shouldn't block the stop.
         if (this.cacheAdapter) {
+            let timer;
+            const budget = new Promise((resolve) => {
+                timer = setTimeout(() => resolve('timeout'), SHUTDOWN_FLUSH_BUDGET_MS);
+                timer.unref?.();
+            });
             try {
-                await this.cacheAdapter.flushAll();
+                // Best-effort persistence: a failure only means the next boot pays
+                // the prefill cost on resume.
+                const outcome = await Promise.race([this.cacheAdapter.flushAll().catch(() => 0), budget]);
+                if (outcome === 'timeout') {
+                    log.warn(`slot cache save still running after ${SHUTDOWN_FLUSH_BUDGET_MS}ms at shutdown; stopping the engine`);
+                }
             }
-            catch {
-                // Best-effort persistence — failure here just means the next
-                // boot pays the prefill cost on resume, same as before.
+            finally {
+                clearTimeout(timer);
             }
         }
         await this.supervisor?.stop();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bendyline/gezel-inference",
-  "version": "1.0.5",
+  "version": "1.0.6",
   "description": "Gezel's on-device inference runtime: native engines, local model management, embeddings, knowledge search, and OpenAI-compatible routes.",
   "license": "MIT",
   "keywords": [
@@ -73,9 +73,9 @@
     "!dist/**/__pycache__"
   ],
   "dependencies": {
-    "@bendyline/gezel": "1.2.10",
-    "@bendyline/gezel-catalog": "1.2.10",
-    "@bendyline/gezel-client": "1.2.10",
+    "@bendyline/gezel": "1.2.11",
+    "@bendyline/gezel-catalog": "1.2.11",
+    "@bendyline/gezel-client": "1.2.11",
     "@bendyline/gezel-knowledge": "1.2.9",
     "hono": "4.13.9",
     "tar": "7.5.21",