@bendyline/gezel-inference 1.0.7 → 1.0.9

package/dist/engines/native-manifest.d.ts

@@ -40,12 +40,12 @@
  * but logs that it could not anchor to a baked-in digest.
  */
 /** Native release version this build pins. Placeholder until first public release. */
-export declare const NATIVE_ENGINE_RELEASE = "0.1.57";
+export declare const NATIVE_ENGINE_RELEASE = "0.1.62";
 /** sha256 of the pinned release's `SHA256SUMS` asset. All-zeros = unpinned. */
-export declare const SHA256SUMS_DIGEST = "977b0ad4cdb8983ed44d7a3223267038b1f35b2806d8a1807d7d15c5c7e94f47";
-/** Exact SHA256 values for every engine archive published by native-v0.1.57. */
+export declare const SHA256SUMS_DIGEST = "9c2e726c1cef4152b5ffb8cee8e9970cf24127786d3fafcce3582dcf0df9badb";
+/** Exact SHA256 values for every engine archive published by native-v0.1.62. */
 export declare const NATIVE_ENGINE_ARCHIVE_SHA256: Readonly<Record<string, string>>;
-/** Download size in bytes of every engine archive published by native-v0.1.57. */
+/** Download size in bytes of every engine archive published by native-v0.1.62. */
 export declare const NATIVE_ENGINE_ARCHIVE_BYTES: Readonly<Record<string, number>>;
 /**
  * True only when the pinned release's standalone macOS archives were

package/dist/engines/native-manifest.js

@@ -40,47 +40,51 @@
  * but logs that it could not anchor to a baked-in digest.
  */
 /** Native release version this build pins. Placeholder until first public release. */
-export const NATIVE_ENGINE_RELEASE = '0.1.57';
+export const NATIVE_ENGINE_RELEASE = '0.1.62';
 /** sha256 of the pinned release's `SHA256SUMS` asset. All-zeros = unpinned. */
-export const SHA256SUMS_DIGEST = '977b0ad4cdb8983ed44d7a3223267038b1f35b2806d8a1807d7d15c5c7e94f47';
+export const SHA256SUMS_DIGEST = '9c2e726c1cef4152b5ffb8cee8e9970cf24127786d3fafcce3582dcf0df9badb';
 // BEGIN PINNED NATIVE ARCHIVE HASHES
-/** Exact SHA256 values for every engine archive published by native-v0.1.57. */
+/** Exact SHA256 values for every engine archive published by native-v0.1.62. */
 export const NATIVE_ENGINE_ARCHIVE_SHA256 = Object.freeze({
-    'gezel-native-0.1.57-darwin-arm64-metal.tar.gz': '6b97be648b032980cf73663db4424140f4cc04577972dc283b956af677ce884a',
-    'gezel-native-0.1.57-darwin-arm64.tar.gz': 'e8ca070c4d3961e95946ed32fe7376ddbc97e407dd3f5b78e26cab20c8fb286a',
-    'gezel-native-0.1.57-linux-arm64-cpu.tar.gz': 'cc87822123602a0024c63a0d65789232a665140cb2fe37b106d5bd680d12f795',
-    'gezel-native-0.1.57-linux-arm64-cuda.tar.gz': '315a441e5698800945f7030e9da961e20f3cf379a4b75c98e1d6ba88c8fee1c5',
-    'gezel-native-0.1.57-linux-arm64.tar.gz': '0684b77d23f008f8a2aa0d9d8931437bacd8569fb6c10d3ddab61a5fb25b6a06',
-    'gezel-native-0.1.57-linux-x64-cpu.tar.gz': 'b2100842f51c393b92667535798e49932173d52626f5ae72c40005b3408f1201',
-    'gezel-native-0.1.57-linux-x64-cuda.tar.gz': '2fca7720614687e8491be17366060c604551257199d395f795b27a87216bcde9',
-    'gezel-native-0.1.57-linux-x64-vulkan.tar.gz': 'c1014a1fe1d241ad1ee9abfa5badafcce208cbff8d38b4b4deb4779f31a95bef',
-    'gezel-native-0.1.57-linux-x64.tar.gz': 'a7f3dbe1941c3d03ed121c32be1bdd199aea47080b03d89989590a0f5d13c5ff',
-    'gezel-native-0.1.57-win32-arm64-cpu.zip': '12d66681cec1872f23b015c7d0bd45a25ed262342da6b0ecea8bea303adf2105',
-    'gezel-native-0.1.57-win32-arm64-cuda.zip': 'b437a9f0894673179f91a1af75d5d7d058dafa1b4f48a57eab123596f1c21ca4',
-    'gezel-native-0.1.57-win32-arm64.zip': '21cddb8b733f9a2669d64b1ada18002f60bb2779f2d5e2aacfb88214db00da7e',
-    'gezel-native-0.1.57-win32-x64-cpu.zip': 'a81f7c6754f1954a0ed5de1efc47d6194dc60d912138a9d2b8e4f992168b12c6',
-    'gezel-native-0.1.57-win32-x64-cuda.zip': '9e3f21713fe77c23536dd4124244ac3834f90c82d6e470cb6836a73bb4f7c592',
-    'gezel-native-0.1.57-win32-x64-vulkan.zip': '0d473300317b4f6cb072bdd7875ce1536a85fe50e781c435e1aef1c4cbdb08e5',
-    'gezel-native-0.1.57-win32-x64.zip': '89a073abcfc97e50b9a0c39f56862b51ac21139b7ca4a3e7a75105dca56292a2',
+    'gezel-native-0.1.62-darwin-arm64-metal.tar.gz': '0ac76122fdc7bd33e6774fb5f62dea6ff22220c483c78026667e33ff8014936d',
+    'gezel-native-0.1.62-darwin-arm64.tar.gz': '23607c3e5c3e6ac050f0594bcf4ca997cff91c57a4b65af985511e1d87843764',
+    'gezel-native-0.1.62-linux-arm64-cpu.tar.gz': 'e9a69cdb07f5ad0103da8d99715e4a1bd257219c55457716ff93e34ee4a544ab',
+    'gezel-native-0.1.62-linux-arm64-cuda.tar.gz': '6389e97d85a2d4b53b4a1c6735669e5cd6022f393fdccabe4c18c9957f9073fc',
+    'gezel-native-0.1.62-linux-arm64.tar.gz': 'cd263de007b14be6de5b828985d3066490a339f097ef199086cd51692495b9f0',
+    'gezel-native-0.1.62-linux-x64-cpu.tar.gz': '2f1b382a02082767187340f8ecd037dfa72953004977e5716ca95127429e4f0b',
+    'gezel-native-0.1.62-linux-x64-cuda.tar.gz': '8f28018fdfd5a38e48dc886a92eab392f92193d3a58c322c85dcd16c6f80fd13',
+    'gezel-native-0.1.62-linux-x64-rocm.tar.gz': '37b88b81bdcbe06fa3c5ebe3b07f0d0112b555449cc2d29775d5ad8373150f18',
+    'gezel-native-0.1.62-linux-x64-vulkan.tar.gz': 'c39fff97dceeacb377a1948ebe9f9553a98485cf21cda9492fbfbe18dd279193',
+    'gezel-native-0.1.62-linux-x64.tar.gz': '2d0c927f98379c0cd1dd05d68f2cc4b87178ede077a65e948bf36e124d560993',
+    'gezel-native-0.1.62-win32-arm64-cpu.zip': 'f867d50591df2da2c2d78cbf844a5b87679da9f5abc9a29ed6b06905c5423bed',
+    'gezel-native-0.1.62-win32-arm64-cuda.zip': 'd42c79845f64ee1fb5e97ea45dde44c6b4705ad6bc4cd6ce9360caffc27cf67f',
+    'gezel-native-0.1.62-win32-arm64.zip': 'd4e660181cd387051d4f8728d6ae008004969786f68da15c76e44704db247af0',
+    'gezel-native-0.1.62-win32-x64-cpu.zip': '85a350c01454e77d49e2a7e463089025d71b6f7cd25daa30b5b0fe8498cf3f2a',
+    'gezel-native-0.1.62-win32-x64-cuda.zip': 'edddeeb820f4612453c036bccd4882c507811464520eb772dfa5bf7b4aa2658f',
+    'gezel-native-0.1.62-win32-x64-rocm.zip': '8640280b49fbc4bc6152b734fcb2dd02b0b9c3375797685470da52406a46aa34',
+    'gezel-native-0.1.62-win32-x64-vulkan.zip': 'a57071775f8876800d6e5689d38a3f7362c4b585e8137596ee19e45ee38b6991',
+    'gezel-native-0.1.62-win32-x64.zip': '357e6c79a4d02968a2ba67299e78efed8dc2ab2d4b40a1e8f866073506b6b2b6',
 });
-/** Download size in bytes of every engine archive published by native-v0.1.57. */
+/** Download size in bytes of every engine archive published by native-v0.1.62. */
 export const NATIVE_ENGINE_ARCHIVE_BYTES = Object.freeze({
-    'gezel-native-0.1.57-darwin-arm64-metal.tar.gz': 7387963,
-    'gezel-native-0.1.57-darwin-arm64.tar.gz': 42223786,
-    'gezel-native-0.1.57-linux-arm64-cpu.tar.gz': 9223249,
-    'gezel-native-0.1.57-linux-arm64-cuda.tar.gz': 621626283,
-    'gezel-native-0.1.57-linux-arm64.tar.gz': 42040003,
-    'gezel-native-0.1.57-linux-x64-cpu.tar.gz': 13398337,
-    'gezel-native-0.1.57-linux-x64-cuda.tar.gz': 1000325316,
-    'gezel-native-0.1.57-linux-x64-vulkan.tar.gz': 26290440,
-    'gezel-native-0.1.57-linux-x64.tar.gz': 57190295,
-    'gezel-native-0.1.57-win32-arm64-cpu.zip': 6049234,
-    'gezel-native-0.1.57-win32-arm64-cuda.zip': 243595554,
-    'gezel-native-0.1.57-win32-arm64.zip': 42236484,
-    'gezel-native-0.1.57-win32-x64-cpu.zip': 9131460,
-    'gezel-native-0.1.57-win32-x64-cuda.zip': 756439338,
-    'gezel-native-0.1.57-win32-x64-vulkan.zip': 21673895,
-    'gezel-native-0.1.57-win32-x64.zip': 55668381,
+    'gezel-native-0.1.62-darwin-arm64-metal.tar.gz': 7577614,
+    'gezel-native-0.1.62-darwin-arm64.tar.gz': 42917428,
+    'gezel-native-0.1.62-linux-arm64-cpu.tar.gz': 9518882,
+    'gezel-native-0.1.62-linux-arm64-cuda.tar.gz': 624413294,
+    'gezel-native-0.1.62-linux-arm64.tar.gz': 43076642,
+    'gezel-native-0.1.62-linux-x64-cpu.tar.gz': 13762290,
+    'gezel-native-0.1.62-linux-x64-cuda.tar.gz': 1016218245,
+    'gezel-native-0.1.62-linux-x64-rocm.tar.gz': 313875007,
+    'gezel-native-0.1.62-linux-x64-vulkan.tar.gz': 26687129,
+    'gezel-native-0.1.62-linux-x64.tar.gz': 58389984,
+    'gezel-native-0.1.62-win32-arm64-cpu.zip': 6219263,
+    'gezel-native-0.1.62-win32-arm64-cuda.zip': 245764366,
+    'gezel-native-0.1.62-win32-arm64.zip': 42822090,
+    'gezel-native-0.1.62-win32-x64-cpu.zip': 9357113,
+    'gezel-native-0.1.62-win32-x64-cuda.zip': 761458453,
+    'gezel-native-0.1.62-win32-x64-rocm.zip': 241630451,
+    'gezel-native-0.1.62-win32-x64-vulkan.zip': 21898409,
+    'gezel-native-0.1.62-win32-x64.zip': 56194044,
 });
 // END PINNED NATIVE ARCHIVE HASHES
 /**

package/dist/engines/resolver.d.ts

@@ -24,7 +24,7 @@ export type EngineResolveEvent = NativeEngineResolveEvent | EngineExtractingEven
 export interface ResolveEngineOptions {
     engine: NativeBinaryName;
     home: string;
-    /** GPU backend for llama-server (`cuda`/`vulkan`/`metal`/`cpu`); omit for single-variant engines. */
+    /** GPU backend for llama-server (`cuda`/`rocm`/`vulkan`/`metal`/`cpu`); omit for single-variant engines. */
     variant?: string;
     /** Native release version; defaults to the source pin / `GEZEL_NATIVE_ENGINE_VERSION`. */
     version?: string;
@@ -100,22 +100,23 @@ export declare function resolveEngine(opts: ResolveEngineOptions): AsyncGenerato
  */
 export declare function resolveEngineToCompletion(opts: ResolveEngineOptions, onEvent?: (e: EngineResolveEvent) => void): Promise<EngineResolveResult>;
 /**
- * The archive variant an engine lives in when the caller names none.
+ * The llama-server build an on-demand install fetches when the probe found
+ * `detected`: the first build down discovery's own fallback chain (ROCm and
+ * CUDA → Vulkan → CPU) that the pinned release publishes for this platform.
+ * Without this, a probe answer the release has no archive for — ROCm before
+ * the release that first carries it, Vulkan on Linux ARM — failed the install
+ * outright instead of taking the next build down.
  *
- * These are facts about *where the archive is*, not backend choices:
- *
- *  - **ds4-server on Linux** ships only in the `-cuda` archive, sharing one
- *    copy of the NVIDIA redistributables with llama-server's CUDA build
- *    instead of shipping them twice (see scripts/native-payload.mjs).
- *  - **llama-server on macOS** ships only in the `-metal` archive; there is no
- *    `darwin-arm64-cpu` asset, and the bare `darwin-arm64` key holds
- *    ds4/sd/whisper/uv but no llama-server. Omitting the variant there fails
- *    the extract with "none of [gezel-llama-server, llama-server] were inside
- *    it" — a loud error, but an avoidable one.
- *
- * Deriving both here is the point: every call site would otherwise have to
- * remember, and one eventually won't. Where the variant is a genuine choice
- * (llama-server on Linux/Windows: cuda / vulkan / cpu) this returns undefined
- * and the decision stays with the caller.
+ * ROCm also has to be worth its several hundred megabytes before anything
+ * downloads: `rocmWorthDownloading` (rocm-devices.ts), which says no on
+ * Windows. An explicit `llamaCppBackendOverride` never comes through here.
  */
+export declare function llamaDownloadVariant(detected: string | undefined, opts?: {
+    platform?: NodeJS.Platform;
+    platformKey?: string | null;
+    version?: string;
+    /** Archive names the release publishes; null when unpinned (every build counts). */
+    archives?: Readonly<Record<string, string>> | null;
+    rocmWorth?: () => boolean;
+}): string | undefined;
 export declare function impliedEngineVariant(engine: NativeBinaryName, platform: NodeJS.Platform): string | undefined;

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, } 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, resolveAvailableLlamaBinary, resolvePlatformKey, rocmWorthDownloading, } 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';
@@ -744,6 +744,41 @@ function stripVer(v) {
  * (llama-server on Linux/Windows: cuda / vulkan / cpu) this returns undefined
  * and the decision stays with the caller.
  */
+const LLAMA_DOWNLOAD_BACKENDS = ['cuda', 'rocm', 'vulkan', 'metal', 'cpu'];
+/**
+ * The llama-server build an on-demand install fetches when the probe found
+ * `detected`: the first build down discovery's own fallback chain (ROCm and
+ * CUDA → Vulkan → CPU) that the pinned release publishes for this platform.
+ * Without this, a probe answer the release has no archive for — ROCm before
+ * the release that first carries it, Vulkan on Linux ARM — failed the install
+ * outright instead of taking the next build down.
+ *
+ * ROCm also has to be worth its several hundred megabytes before anything
+ * downloads: `rocmWorthDownloading` (rocm-devices.ts), which says no on
+ * Windows. An explicit `llamaCppBackendOverride` never comes through here.
+ */
+export function llamaDownloadVariant(detected, opts = {}) {
+    const platform = opts.platform ?? process.platform;
+    if (platform === 'darwin')
+        return impliedEngineVariant('llama-server', platform);
+    const preferred = LLAMA_DOWNLOAD_BACKENDS.find((backend) => backend === detected);
+    if (!preferred)
+        return detected;
+    const platformKey = opts.platformKey === undefined ? resolvePlatformKey() : opts.platformKey;
+    const version = stripVer(opts.version ?? effectiveEngineRelease());
+    const ext = platform === 'win32' ? 'zip' : 'tar.gz';
+    const archives = opts.archives === undefined
+        ? isPlaceholderDigest(SHA256SUMS_DIGEST) || process.env.GEZEL_NATIVE_ENGINE_VERSION
+            ? null
+            : NATIVE_ENGINE_ARCHIVE_SHA256
+        : opts.archives;
+    const rocmWorth = opts.rocmWorth ?? (() => rocmWorthDownloading({ platform }));
+    const resolved = resolveAvailableLlamaBinary(preferred, (backend) => {
+        const name = `gezel-native-${version}-${platformKey}-${backend}.${ext}`;
+        return !archives || archives[name] ? name : null;
+    }, true, (backend) => backend !== 'rocm' || rocmWorth());
+    return resolved?.backend ?? preferred;
+}
 export function impliedEngineVariant(engine, platform) {
     if (engine === 'ds4-server' && platform === 'linux')
         return 'cuda';

package/dist/http/model-inventory-cache.d.ts

@@ -4,5 +4,5 @@ export interface CacheEntry {
     at: number;
     models: ModelInfo[];
 }
-export declare const cache: Map<"copilot" | "openai" | "anthropic" | "anthropic-cli" | "codex-cli" | "ollama" | "llama-cpp" | "mlx" | "ds4" | "apple-foundation-models" | "android-mlkit" | "remote", CacheEntry>;
+export declare const cache: Map<"copilot" | "openai" | "anthropic" | "anthropic-cli" | "codex-cli" | "ollama" | "llama-cpp" | "mlx" | "ds4" | "remote" | "apple-foundation-models" | "android-mlkit", CacheEntry>;
 export declare function invalidateModelsCache(provider?: ProviderName): void;

package/dist/http/routes/engines.js

@@ -40,7 +40,7 @@ export function enginesRoutes(ctx) {
      */
     app.get('/binaries/status', (c) => {
         const backendRaw = process.env.GEZEL_LLAMA_SERVER_BACKEND ?? process.env.GEZEL_LLAMA_DETECTED_BACKEND;
-        const llamaBackend = ['cuda', 'vulkan', 'metal', 'cpu'].find((value) => value === backendRaw);
+        const llamaBackend = ['cuda', 'rocm', 'vulkan', 'metal', 'cpu'].find((value) => value === backendRaw);
         const platformKey = resolvePlatformKey();
         return c.json({
             release: effectiveEngineRelease(),
@@ -64,7 +64,7 @@ export function enginesRoutes(ctx) {
     });
     /**
      * Start (or attach to) a download+verify of `engine` (optionally a GPU
-     * `?variant=cuda|vulkan|metal|cpu`) and stream progress as SSE. The
+     * `?variant=cuda|rocm|vulkan|metal|cpu`) and stream progress as SSE. The
      * manual trigger behind Settings → On-device → "Download engine". The
      * lazy on-device-chat path uses the same registry. Idempotent — a
      * second call while one is in flight attaches to the running job.
@@ -231,7 +231,7 @@ export function enginesRoutes(ctx) {
     app.delete('/llama-cpp/quarantine/:backend', async (c) => {
         const backend = c.req.param('backend');
         if (!isLlamaBackend(backend)) {
-            return c.json({ error: 'backend must be "cuda", "vulkan", "metal", or "cpu"' }, 400);
+            return c.json({ error: 'backend must be "cuda", "rocm", "vulkan", "metal", or "cpu"' }, 400);
         }
         const override = (await ctx.store.readConfig()).llamaCppBackendOverride;
         const { restored } = liftLlamaQuarantine(ctx.home, backend, {

package/dist/http/routes/v1-images.d.ts

@@ -25,7 +25,8 @@ export interface ImagesRouteContext {
  *   GET    /models                   Installed and downloadable image models,
  *                                    with this machine's fit. Never part of
  *                                    `/v1/models`, so chat pickers never
- *                                    offer them.
+ *                                    offer them. Models that only change a
+ *                                    given picture are left out.
  *   POST   /models/ensure            `{model: "sd-cpp:<id>"}` → ready, or a
  *                                    background download (`job_id`).
  *   GET    /models/ensure/:jobId     Snapshot of a download.

package/dist/http/routes/v1-images.js

@@ -4,8 +4,9 @@ import { streamSSE } from 'hono/streaming';
 import { z } from 'zod';
 import { imageCatalogFits, imageModelFit, wontRunRefusal } from '../../models/image-fit.js';
 import { ReadOnlyModelError, modelExistsOnlyReadOnly, modelStorageRoots, } from '../../models/storage-roots.js';
+import { resolveImageEditCapability } from '../../providers/image/img2img-support.js';
 import { NoLocalImageModelError } from '../../providers/image/local-factory.js';
-import { ImageCapacityError, ImageGenerationCancelledError, } from '../../providers/image/types.js';
+import { ImageCapacityError, ImageGenerationCancelledError, ImageSourceRequiredError, } from '../../providers/image/types.js';
 import { registerImageEnsureRoutes } from './image-ensure-routes.js';
 import { v1ImageStylesRoutes } from './v1-image-styles.js';
 const log = createLogger('v1-images');
@@ -28,6 +29,17 @@ function catalogIdOf(model) {
 function qualified(catalogId) {
     return `${ENGINE}:${catalogId}`;
 }
+/**
+ * Made only to change a picture it is given. Generations here take no source
+ * picture, so such a model can never render one: it is left out of the list
+ * and refused by name, never handed a blank page.
+ */
+function editsOnly(modelId, declared) {
+    return resolveImageEditCapability({ modelId, editOnly: declared }).editOnly;
+}
+function editOnlyError(name) {
+    return apiError('edit_only_model', `${name} only changes a picture it is given and cannot draw a new one. Pick another image model.`);
+}
 function downloadBytes(manifest) {
     return (manifest.approxSizeBytes + manifest.auxiliaryFiles.reduce((n, f) => n + f.approxSizeBytes, 0));
 }
@@ -60,6 +72,8 @@ function generationError(err) {
         return { code: err.code, message };
     if (err instanceof NoLocalImageModelError)
         return { code: err.code, message };
+    if (err instanceof ImageSourceRequiredError)
+        return { code: err.code, message };
     const code = err?.code;
     if (code === 'capacity-denied')
         return { code: 'capacity_denied', message };
@@ -71,7 +85,8 @@ function generationError(err) {
  *   GET    /models                   Installed and downloadable image models,
  *                                    with this machine's fit. Never part of
  *                                    `/v1/models`, so chat pickers never
- *                                    offer them.
+ *                                    offer them. Models that only change a
+ *                                    given picture are left out.
  *   POST   /models/ensure            `{model: "sd-cpp:<id>"}` → ready, or a
  *                                    background download (`job_id`).
  *   GET    /models/ensure/:jobId     Snapshot of a download.
@@ -117,6 +132,8 @@ export function v1ImagesRoutes(ctx) {
             if (manifest.kind !== 'image-model')
                 continue;
             seen.add(manifest.id);
+            if (editsOnly(manifest.id, manifest.editOnly))
+                continue;
             const fit = fitById.get(manifest.id);
             const isInstalled = installedIds.has(manifest.id);
             data.push({
@@ -137,7 +154,7 @@ export function v1ImagesRoutes(ctx) {
         }
         // Installed models the catalog no longer lists still render.
         for (const model of installed) {
-            if (seen.has(model.id))
+            if (seen.has(model.id) || editsOnly(model.id, model.editOnly))
                 continue;
             data.push({
                 id: qualified(model.id),
@@ -157,8 +174,17 @@ export function v1ImagesRoutes(ctx) {
             return c.json(apiError('unsupported_backend', `Image models are ${ENGINE}:<id>`), 400);
         }
         const provider = await ctx.images.current();
-        const installed = await provider.listInstalledModels();
-        if (installed.some((m) => m.id === id)) {
+        const [installed, item] = await Promise.all([
+            provider.listInstalledModels(),
+            ctx.catalog.get('image-model', id).catch(() => null),
+        ]);
+        const local = installed.find((m) => m.id === id);
+        const listed = item?.manifest.kind === 'image-model' ? item.manifest : undefined;
+        // Downloading one would fetch gigabytes no generation here can use.
+        if (editsOnly(id, local?.editOnly ?? listed?.editOnly)) {
+            return c.json(editOnlyError(local?.name ?? listed?.name ?? model), 409);
+        }
+        if (local) {
             return c.json({ status: 'ready', model_id: qualified(id) });
         }
         const refusal = await wontRunRefusal(() => imageModelFit(fitContext, id));
@@ -209,16 +235,17 @@ export function v1ImagesRoutes(ctx) {
             return c.json(apiError('unsupported_backend', `Image models are ${ENGINE}:<id>`), 400);
         }
         const modelId = requested ?? (await provider.resolveDefaultModelId?.());
-        if (!modelId || !installed.some((m) => m.id === modelId)) {
+        const chosen = modelId ? installed.find((m) => m.id === modelId) : undefined;
+        if (!modelId || !chosen) {
             const error = requested
                 ? apiError('model_not_installed', `${request.model} is not installed; ensure it first`)
                 : apiError('no_local_image_model', new NoLocalImageModelError().message);
             return c.json(error, 409);
         }
+        if (editsOnly(modelId, chosen.editOnly))
+            return c.json(editOnlyError(chosen.name), 409);
         const style = request.style
-            ? installed
-                .find((model) => model.id === modelId)
-                ?.loras?.find((lora) => lora.id === request.style && lora.purpose === 'style')
+            ? chosen.loras?.find((lora) => lora.id === request.style && lora.purpose === 'style')
             : undefined;
         if (request.style && !style) {
             return c.json(apiError('style_not_available', 'That style is not installed for this image model.'), 409);

package/dist/knowledge/manager.d.ts

@@ -8,7 +8,7 @@
  * class never opens a catalog database itself.
  */
 import type { GezelConfig, IncompleteKnowledgeDownload, KnowledgeActiveInstall, KnowledgeAvailableCatalog, KnowledgeCatalogRef, KnowledgeCatalogStatus, KnowledgeEmbeddingProfile, KnowledgeInstallEvent, KnowledgeInstallJob, KnowledgeInstallRequest, KnowledgePassageRead, KnowledgeSemanticSearchMode, KnowledgeUpdateCandidate, KnowledgeUri, ProjectKnowledgeCatalogs, UnifiedSearchResult } from '@bendyline/gezel';
-import { type KnowledgeNearbyRequest, type KnowledgeNearbyResponse, type KnowledgeRadius } from '@bendyline/gezel';
+import { type KnowledgeNearbyRequest, type KnowledgeNearbyResponse } from '@bendyline/gezel';
 import type { HistoryEvent } from '@bendyline/gezel';
 import type { CatalogService } from '@bendyline/gezel-catalog';
 import { type ModelDownloadProgress } from '@bendyline/gezel-knowledge';
@@ -22,6 +22,7 @@ import { ChatModelInstallRegistry } from '../models/install-registry.js';
 import type { KnowledgeCatalogHost, KnowledgeChunk } from './catalog-host.js';
 import { KnowledgeRegistry } from './registry.js';
 import type { SharedKnowledgeInstaller } from './shared-install-contract.js';
+import { type UnifiedSearchOptions } from './unified-search.js';
 import { type KnowledgeVectorFloors } from './vector-floors.js';
 export type { KnowledgeCatalogStatus, KnowledgeInstallJob, KnowledgeSemanticSearchMode };
 export type KnowledgeInstallRequestSource = KnowledgeInstallRequest['source'];
@@ -245,31 +246,7 @@ export declare class KnowledgeManager {
      * Search each catalog in its embedding space and fuse lexical/vector ranks.
      * Spatial eligibility applies inside workers before candidate selection.
      */
-    searchUnified(query: string, opts: {
-        vector: number[] | null;
-        maxResults: number;
-        projectId?: string;
-        spatial?: KnowledgeRadius;
-        catalogs?: readonly string[];
-        /** Search a profile group keyword-only when its query model is not ready by then. */
-        queryEmbedBudgetMs?: number;
-        /** Query cached profile models only, falling back to keywords when absent. */
-        localModelsOnly?: boolean;
-        /**
-         * Include media rows (images, audio and video windows) as results of
-         * their own. Explicit search asks; proactive injection does not.
-         */
-        media?: boolean;
-        /**
-         * Also return up to this many of a document's other passages that the
-         * search found, after its best one. Fusion otherwise keeps one passage
-         * per document, chosen by vector similarity alone, so an article's name-
-         * dense infobox can stand in for the section a reranker would choose.
-         */
-        chunksPerDocument?: number;
-    }): Promise<UnifiedSearchResult[]>;
-    /** A document the lexical arms named but no chunk carries: cite it with its lead paragraph. */
-    private toDocumentResult;
+    searchUnified(query: string, opts: UnifiedSearchOptions): Promise<UnifiedSearchResult[]>;
     /** A catalog's vector floor, or null when its hits are keyword-only or its scale unmeasured. */
     private vectorFloorFor;
     /**
@@ -285,7 +262,6 @@ export declare class KnowledgeManager {
      * next query gets vectors.
      */
     private embedForProfile;
-    private toResult;
     /** Enabled+mounted catalogs, intersected with the project policy. */
     private activeCatalogKeys;
 }