crayonmlx 0.1.0a1__tar.gz

crayonmlx-0.1.0a1/PKG-INFO

@@ -0,0 +1,201 @@
+Metadata-Version: 2.3
+Name: crayonmlx
+Version: 0.1.0a1
+Summary: Local MLX studio for bringing drawings to life on Apple Silicon
+Author: Nikolay Nechaev
+Author-email: Nikolay Nechaev <godlikeguest@gmail.com>
+Requires-Dist: mflux>=0.24,<0.25
+Requires-Dist: fastapi>=0.162.0
+Requires-Dist: huggingface-hub>=1.1.6,<2
+Requires-Dist: pillow>=13.0.0
+Requires-Dist: platformdirs>=4.14.0
+Requires-Dist: python-multipart>=0.0.33
+Requires-Dist: uvicorn[standard]>=0.54.0
+Requires-Python: >=3.12
+Description-Content-Type: text/markdown
+
+# CrayonMLX
+
+A local web studio for transforming children's drawings on Apple Silicon.
+Choose what to change: rendering style, characters, background, framing, and light.
+Keep the child's original ideas and explicitly protect the details that matter.
+Python + FastAPI, plain HTML/CSS/JavaScript, native MLX inference through MFLUX.
+
+**Early alpha.** The UI and orchestration are tested independently of model weights.
+Actual image quality, memory requirements and checkpoint compatibility depend on your model.
+
+## Install and run
+
+Requires an Apple Silicon Mac, macOS compatible with MLX, and [uv](https://docs.astral.sh/uv/).
+From a checkout:
+
+```sh
+uv tool install .
+crayonmlx
+```
+
+Open http://127.0.0.1:8787. Choose another port with `crayonmlx --port 8788`.
+After the package is published, installation will be `uv tool install crayonmlx`.
+The project is not published yet. MFLUX libraries are installed with the app;
+model weights are **not** bundled or downloaded at startup.
+
+For development:
+
+```sh
+uv sync
+uv run crayonmlx
+uv run pytest
+uv run ruff check .
+uv build
+```
+
+## Workflow
+
+1. In **Settings**, choose absolute paths for the model folder and output folder.
+   Any accessible location, including an external disk, is supported.
+2. Put each full Diffusers or MFLUX checkpoint in its own subfolder, or select a
+   checkpoint folder itself. HF cache users can select a concrete snapshot directory.
+   An individual `.safetensors` or `.gguf` file is not a complete checkpoint.
+3. Alternatively download a repository and revision in **Models**. Private/gated
+   repositories need a Hugging Face token in Settings and access granted on HF.
+   Downloads run in the background. Incomplete downloads are hidden and can be resumed
+   by requesting the same repository again. Download status is session-local.
+4. Upload a drawing; start your own combination or select a saved full preset, then choose one local model. Its variant and MFLUX
+   adapter are detected automatically. Unrecognized or ambiguous checkpoints are
+   shown in Models with an explanation and cannot be used for generation.
+5. Assemble independent blocks: **overall style**, **characters/objects**, **background**,
+   **composition**, **lighting**, and **editing/preservation instructions**. Each block
+   offers saved templates, “Do not change”, and editable text. Overall style does not
+   dictate a new character or location. Subject-specific styling takes precedence
+   for the original foreground subjects only.
+6. Check what must remain: identity, clothing, poses, colors, subjects/objects,
+   placement/scale, lettering/decorations, and original background. Choosing a new
+   background unchecks background preservation; choosing a new composition unchecks
+   placement preservation. The UI tells you when this happens. An unchecked option
+   allows adaptation; it does not itself request a particular replacement.
+7. Choose a generation configuration to apply its parameters, then adjust values
+   for this run if needed.
+8. Manage full presets, prompt templates and generation configurations in the
+   **Presets** tab. Each library supports creation, editing and deletion.
+   Quantization includes None (omit the MFLUX quantization flag), 4 bit and 8 bit.
+9. Generate. History keeps the source, result, composed prompt, resolved seed, model
+   path, and all parameters as an immutable snapshot. History is read-only.
+
+## Build a transformation
+
+For example, turn the original characters into anime characters standing in Tokyo:
+
+| Block | Selection |
+| --- | --- |
+| Overall style | Do not change, or choose a treatment for the whole scene |
+| Characters/objects | Anime |
+| Background | Tokyo |
+| Composition | Entire subject |
+| Lighting | Neon |
+| Preservation | Identity, clothing, poses, colors, original subjects, lettering |
+
+The city replaces the surrounding background; a house that is the main subject
+remains that house. Framing and lighting do not introduce new protagonists.
+Use **Show final prompt** to inspect the complete request before generating.
+
+Included treatments: Huntrix, Roblox, anime, Studio Ghibli, realistic painting,
+Disney (including Frozen and Encanto), Pixar (including Soul), and pixel art.
+They are independently available for the whole scene and the existing main subjects.
+Backgrounds include Seoul, Tokyo, Paris, Saint Petersburg, New York, Kyoto, forest,
+mountains, sea, and space. Lighting and composition have their own libraries.
+The interface is currently in Russian; generation instructions are in English.
+
+The three libraries remain independent:
+
+- **Prompt templates** supply one editable block of instructions.
+- **Generation configurations** supply numerical rendering parameters.
+- **Full presets** store a copy of all blocks, preservation choices, and parameters.
+
+Updating a template does not rewrite saved full presets or historical jobs. Built-in
+library upgrades replace only unmodified factory entries; edited user entries are
+kept. Deleted entries are not reinstalled on every launch.
+
+## Supported MLX adapters
+
+- **MFLUX / Qwen Image Edit**: `qwen-image-edit` (2509) and `qwen-image-edit-2511`.
+  Qwen uses reference-image editing; it does **not** apply image strength.
+- **MFLUX / FLUX.1 image-to-image**: `dev`, `schnell`, `krea-dev`.
+  FLUX schnell does not apply guidance. Higher FLUX image strength preserves more of the source image.
+
+The server identifies the selected checkpoint independently of presets. It uses
+MFLUX metadata, component configs, and the original repository folder name (including
+HF snapshot paths). Keep the original HF name if the checkpoint has no MFLUX manifest.
+Conflicting or unsupported metadata is rejected before a job is created. Pre-quantized MFLUX models can reduce download size and RAM needs.
+The UI quantization setting is passed to MFLUX; checkpoint quantization may be fixed.
+Generation is explicitly offline: missing components fail rather than download silently.
+Only MLX engines are in scope; no PyTorch/MPS or other inference backends.
+
+## Data
+
+Default application data on macOS: `~/Library/Application Support/crayonmlx/`.
+On first launch you can override it with `CRAYONMLX_DATA_DIR=/absolute/path crayonmlx`.
+Settings → Internal data lets you copy settings, presets, history and the HF token
+into an empty folder and use it after restart. The source is retained; model/image
+paths are unchanged. Transfers require no active jobs or downloads and freeze
+further writes until restart. The chosen location is recorded in `location.json`
+in the platform's CrayonMLX configuration folder and takes precedence over the
+bootstrap environment variable on subsequent launches.
+
+- `crayonmlx.sqlite3`: settings and durable generation history.
+- `presets/*.json`: full presets, initialized once.
+- `prompt-templates/*.json`: reusable categorized prompt fragments.
+- `generation-configs/*.json`: independent render configurations.
+- `hf-token`: credential file with mode `0600`; never returned to the browser.
+- Each output job folder: `source.png`, `result.png`, `generation.json` and execution measurements. Worker logs live in `runtime/<session>/worker.log`.
+
+Changing the output directory affects new jobs. Old history keeps its original paths.
+Moving or deleting those files makes their previews unavailable. Jobs run sequentially
+and snapshot paths/parameters at submission. On restart, unfinished jobs become failed
+and remain visible in read-only history. Graceful shutdown waits for current work to finish.
+One separate MLX process keeps the current model resident between jobs. Models
+are unloaded after 15 idle minutes by default; Settings accepts a custom timeout or
+None to disable automatic unloading. Switching model or requested quantization
+releases the previous process before loading another model. Manual load/unload
+and enable/disable controls are available in Models. Disabling preserves files and
+lets an already running generation finish.
+
+Generation reports preparation, loading, completed denoising steps, decoding, and
+elapsed time. MLX active allocations, cache, and peak are reported separately;
+these are not the full process RSS. Successful history entries record execution
+measurements separately from the requested configuration, including effective
+quantization when reported by MFLUX.
+
+The server binds to loopback only. There is no public API or user/account system.
+HF tokens are intended only for downloading models, not for uploads or inference.
+
+## Code layout
+
+All application sources, including the frontend, are in `src/crayonmlx/`:
+
+```text
+app.py             service composition and lifecycle
+schemas.py         validated request/configuration values
+engines/           render contract, registry, MFLUX adapters
+models/            local catalog and HF downloads
+presets/           atomic JSON storage and starting prompts
+generation/        serialized render jobs
+storage/           SQLite history/settings and local credential
+web/               internal UI routes, HTML, CSS and JS modules
+```
+
+To add an MLX engine, implement `engines/base.py` and register it in `engines/registry.py`.
+Tests use a small injected renderer; they do not claim to test model inference.
+Builds include frontend assets and expose the `crayonmlx` command for `uv tool install`.
+
+## Publishing and contributing
+
+See [PUBLISHING.md](PUBLISHING.md) for the maintainer guide: local build verification,
+TestPyPI rehearsal, PyPI publication, and installation with `uv tool`.
+The package has not been published by this setup process.
+
+Keep application code in `src/crayonmlx/`, use small modules with clear responsibilities,
+and add tests for changes to data compatibility, job snapshots, or engine behavior.
+Run `uv run pytest` and `uv run ruff check src tests` before submitting changes.
+Do not commit downloaded models, generated images, application data, or access tokens.
+A project license has not yet been selected; add the chosen license before presenting
+this repository as a licensed open-source release.

crayonmlx-0.1.0a1/README.md

@@ -0,0 +1,185 @@
+# CrayonMLX
+
+A local web studio for transforming children's drawings on Apple Silicon.
+Choose what to change: rendering style, characters, background, framing, and light.
+Keep the child's original ideas and explicitly protect the details that matter.
+Python + FastAPI, plain HTML/CSS/JavaScript, native MLX inference through MFLUX.
+
+**Early alpha.** The UI and orchestration are tested independently of model weights.
+Actual image quality, memory requirements and checkpoint compatibility depend on your model.
+
+## Install and run
+
+Requires an Apple Silicon Mac, macOS compatible with MLX, and [uv](https://docs.astral.sh/uv/).
+From a checkout:
+
+```sh
+uv tool install .
+crayonmlx
+```
+
+Open http://127.0.0.1:8787. Choose another port with `crayonmlx --port 8788`.
+After the package is published, installation will be `uv tool install crayonmlx`.
+The project is not published yet. MFLUX libraries are installed with the app;
+model weights are **not** bundled or downloaded at startup.
+
+For development:
+
+```sh
+uv sync
+uv run crayonmlx
+uv run pytest
+uv run ruff check .
+uv build
+```
+
+## Workflow
+
+1. In **Settings**, choose absolute paths for the model folder and output folder.
+   Any accessible location, including an external disk, is supported.
+2. Put each full Diffusers or MFLUX checkpoint in its own subfolder, or select a
+   checkpoint folder itself. HF cache users can select a concrete snapshot directory.
+   An individual `.safetensors` or `.gguf` file is not a complete checkpoint.
+3. Alternatively download a repository and revision in **Models**. Private/gated
+   repositories need a Hugging Face token in Settings and access granted on HF.
+   Downloads run in the background. Incomplete downloads are hidden and can be resumed
+   by requesting the same repository again. Download status is session-local.
+4. Upload a drawing; start your own combination or select a saved full preset, then choose one local model. Its variant and MFLUX
+   adapter are detected automatically. Unrecognized or ambiguous checkpoints are
+   shown in Models with an explanation and cannot be used for generation.
+5. Assemble independent blocks: **overall style**, **characters/objects**, **background**,
+   **composition**, **lighting**, and **editing/preservation instructions**. Each block
+   offers saved templates, “Do not change”, and editable text. Overall style does not
+   dictate a new character or location. Subject-specific styling takes precedence
+   for the original foreground subjects only.
+6. Check what must remain: identity, clothing, poses, colors, subjects/objects,
+   placement/scale, lettering/decorations, and original background. Choosing a new
+   background unchecks background preservation; choosing a new composition unchecks
+   placement preservation. The UI tells you when this happens. An unchecked option
+   allows adaptation; it does not itself request a particular replacement.
+7. Choose a generation configuration to apply its parameters, then adjust values
+   for this run if needed.
+8. Manage full presets, prompt templates and generation configurations in the
+   **Presets** tab. Each library supports creation, editing and deletion.
+   Quantization includes None (omit the MFLUX quantization flag), 4 bit and 8 bit.
+9. Generate. History keeps the source, result, composed prompt, resolved seed, model
+   path, and all parameters as an immutable snapshot. History is read-only.
+
+## Build a transformation
+
+For example, turn the original characters into anime characters standing in Tokyo:
+
+| Block | Selection |
+| --- | --- |
+| Overall style | Do not change, or choose a treatment for the whole scene |
+| Characters/objects | Anime |
+| Background | Tokyo |
+| Composition | Entire subject |
+| Lighting | Neon |
+| Preservation | Identity, clothing, poses, colors, original subjects, lettering |
+
+The city replaces the surrounding background; a house that is the main subject
+remains that house. Framing and lighting do not introduce new protagonists.
+Use **Show final prompt** to inspect the complete request before generating.
+
+Included treatments: Huntrix, Roblox, anime, Studio Ghibli, realistic painting,
+Disney (including Frozen and Encanto), Pixar (including Soul), and pixel art.
+They are independently available for the whole scene and the existing main subjects.
+Backgrounds include Seoul, Tokyo, Paris, Saint Petersburg, New York, Kyoto, forest,
+mountains, sea, and space. Lighting and composition have their own libraries.
+The interface is currently in Russian; generation instructions are in English.
+
+The three libraries remain independent:
+
+- **Prompt templates** supply one editable block of instructions.
+- **Generation configurations** supply numerical rendering parameters.
+- **Full presets** store a copy of all blocks, preservation choices, and parameters.
+
+Updating a template does not rewrite saved full presets or historical jobs. Built-in
+library upgrades replace only unmodified factory entries; edited user entries are
+kept. Deleted entries are not reinstalled on every launch.
+
+## Supported MLX adapters
+
+- **MFLUX / Qwen Image Edit**: `qwen-image-edit` (2509) and `qwen-image-edit-2511`.
+  Qwen uses reference-image editing; it does **not** apply image strength.
+- **MFLUX / FLUX.1 image-to-image**: `dev`, `schnell`, `krea-dev`.
+  FLUX schnell does not apply guidance. Higher FLUX image strength preserves more of the source image.
+
+The server identifies the selected checkpoint independently of presets. It uses
+MFLUX metadata, component configs, and the original repository folder name (including
+HF snapshot paths). Keep the original HF name if the checkpoint has no MFLUX manifest.
+Conflicting or unsupported metadata is rejected before a job is created. Pre-quantized MFLUX models can reduce download size and RAM needs.
+The UI quantization setting is passed to MFLUX; checkpoint quantization may be fixed.
+Generation is explicitly offline: missing components fail rather than download silently.
+Only MLX engines are in scope; no PyTorch/MPS or other inference backends.
+
+## Data
+
+Default application data on macOS: `~/Library/Application Support/crayonmlx/`.
+On first launch you can override it with `CRAYONMLX_DATA_DIR=/absolute/path crayonmlx`.
+Settings → Internal data lets you copy settings, presets, history and the HF token
+into an empty folder and use it after restart. The source is retained; model/image
+paths are unchanged. Transfers require no active jobs or downloads and freeze
+further writes until restart. The chosen location is recorded in `location.json`
+in the platform's CrayonMLX configuration folder and takes precedence over the
+bootstrap environment variable on subsequent launches.
+
+- `crayonmlx.sqlite3`: settings and durable generation history.
+- `presets/*.json`: full presets, initialized once.
+- `prompt-templates/*.json`: reusable categorized prompt fragments.
+- `generation-configs/*.json`: independent render configurations.
+- `hf-token`: credential file with mode `0600`; never returned to the browser.
+- Each output job folder: `source.png`, `result.png`, `generation.json` and execution measurements. Worker logs live in `runtime/<session>/worker.log`.
+
+Changing the output directory affects new jobs. Old history keeps its original paths.
+Moving or deleting those files makes their previews unavailable. Jobs run sequentially
+and snapshot paths/parameters at submission. On restart, unfinished jobs become failed
+and remain visible in read-only history. Graceful shutdown waits for current work to finish.
+One separate MLX process keeps the current model resident between jobs. Models
+are unloaded after 15 idle minutes by default; Settings accepts a custom timeout or
+None to disable automatic unloading. Switching model or requested quantization
+releases the previous process before loading another model. Manual load/unload
+and enable/disable controls are available in Models. Disabling preserves files and
+lets an already running generation finish.
+
+Generation reports preparation, loading, completed denoising steps, decoding, and
+elapsed time. MLX active allocations, cache, and peak are reported separately;
+these are not the full process RSS. Successful history entries record execution
+measurements separately from the requested configuration, including effective
+quantization when reported by MFLUX.
+
+The server binds to loopback only. There is no public API or user/account system.
+HF tokens are intended only for downloading models, not for uploads or inference.
+
+## Code layout
+
+All application sources, including the frontend, are in `src/crayonmlx/`:
+
+```text
+app.py             service composition and lifecycle
+schemas.py         validated request/configuration values
+engines/           render contract, registry, MFLUX adapters
+models/            local catalog and HF downloads
+presets/           atomic JSON storage and starting prompts
+generation/        serialized render jobs
+storage/           SQLite history/settings and local credential
+web/               internal UI routes, HTML, CSS and JS modules
+```
+
+To add an MLX engine, implement `engines/base.py` and register it in `engines/registry.py`.
+Tests use a small injected renderer; they do not claim to test model inference.
+Builds include frontend assets and expose the `crayonmlx` command for `uv tool install`.
+
+## Publishing and contributing
+
+See [PUBLISHING.md](PUBLISHING.md) for the maintainer guide: local build verification,
+TestPyPI rehearsal, PyPI publication, and installation with `uv tool`.
+The package has not been published by this setup process.
+
+Keep application code in `src/crayonmlx/`, use small modules with clear responsibilities,
+and add tests for changes to data compatibility, job snapshots, or engine behavior.
+Run `uv run pytest` and `uv run ruff check src tests` before submitting changes.
+Do not commit downloaded models, generated images, application data, or access tokens.
+A project license has not yet been selected; add the chosen license before presenting
+this repository as a licensed open-source release.

crayonmlx-0.1.0a1/pyproject.toml

@@ -0,0 +1,49 @@
+[project]
+name = "crayonmlx"
+version = "0.1.0a1"
+description = "Local MLX studio for bringing drawings to life on Apple Silicon"
+readme = "README.md"
+requires-python = ">=3.12"
+dependencies = [
+    "mflux>=0.24,<0.25",
+    "fastapi>=0.162.0",
+    "huggingface-hub>=1.1.6,<2",
+    "pillow>=13.0.0",
+    "platformdirs>=4.14.0",
+    "python-multipart>=0.0.33",
+    "uvicorn[standard]>=0.54.0",
+]
+
+[[project.authors]]
+name = "Nikolay Nechaev"
+email = "godlikeguest@gmail.com"
+
+[project.scripts]
+crayonmlx = "crayonmlx.cli:main"
+
+[build-system]
+requires = ["uv_build>=0.12.19,<0.13.0"]
+build-backend = "uv_build"
+
+[dependency-groups]
+dev = [
+    "httpx>=0.28.1",
+    "pytest>=9.2.0",
+    "ruff>=0.17.3",
+]
+
+[tool.ruff]
+line-length = 110
+target-version = "py312"
+
+[tool.ruff.lint]
+select = [
+    "E4",
+    "E7",
+    "E9",
+    "F",
+    "I",
+]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]

crayonmlx-0.1.0a1/pyproject.toml.orig

@@ -0,0 +1,42 @@
+[project]
+name = "crayonmlx"
+version = "0.1.0a1"
+description = "Local MLX studio for bringing drawings to life on Apple Silicon"
+readme = "README.md"
+authors = [
+    { name = "Nikolay Nechaev", email = "godlikeguest@gmail.com" }
+]
+requires-python = ">=3.12"
+dependencies = [
+    "mflux>=0.24,<0.25",
+    "fastapi>=0.162.0",
+    "huggingface-hub>=1.1.6,<2",
+    "pillow>=13.0.0",
+    "platformdirs>=4.14.0",
+    "python-multipart>=0.0.33",
+    "uvicorn[standard]>=0.54.0",
+]
+
+[project.scripts]
+crayonmlx = "crayonmlx.cli:main"
+
+[build-system]
+requires = ["uv_build>=0.12.19,<0.13.0"]
+build-backend = "uv_build"
+
+[dependency-groups]
+dev = [
+    "httpx>=0.28.1",
+    "pytest>=9.2.0",
+    "ruff>=0.17.3",
+]
+
+[tool.ruff]
+line-length = 110
+target-version = "py312"
+
+[tool.ruff.lint]
+select = ["E4", "E7", "E9", "F", "I"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]

crayonmlx-0.1.0a1/src/crayonmlx/__init__.py

@@ -0,0 +1,3 @@
+"""CrayonMLX — a local drawing transformation studio for Apple Silicon."""
+
+__version__ = "0.1.0a1"

crayonmlx-0.1.0a1/src/crayonmlx/__main__.py

@@ -0,0 +1,3 @@
+from crayonmlx.cli import main
+
+main()

crayonmlx-0.1.0a1/src/crayonmlx/app.py

@@ -0,0 +1,134 @@
+"""Application composition: construct services once and close workers on shutdown."""
+
+import asyncio
+import logging
+from contextlib import asynccontextmanager
+from dataclasses import dataclass
+from pathlib import Path
+
+from fastapi import FastAPI, Request
+from fastapi.responses import FileResponse, JSONResponse
+from fastapi.staticfiles import StaticFiles
+from starlette.middleware.trustedhost import TrustedHostMiddleware
+
+from crayonmlx.engines.registry import default_engines
+from crayonmlx.generation.queue import GenerationQueue
+from crayonmlx.models.downloads import Downloads
+from crayonmlx.models.runtime import ModelRuntime
+from crayonmlx.paths import data_directory
+from crayonmlx.presets.library import ConfigRepository, PromptRepository
+from crayonmlx.presets.repository import PresetRepository
+from crayonmlx.storage.database import Database
+from crayonmlx.storage.secrets import TokenStore
+from crayonmlx.web.folders import router as folders_router
+from crayonmlx.web.library_routes import router as library_router
+from crayonmlx.web.model_routes import router as model_router
+from crayonmlx.web.routes import router
+from crayonmlx.web.storage_routes import router as storage_router
+
+
+@dataclass
+class Services:
+    root: Path
+    database: Database
+    presets: PresetRepository
+    prompts: PromptRepository
+    configs: ConfigRepository
+    tokens: TokenStore
+    engines: dict
+    queue: GenerationQueue
+    downloads: Downloads
+    runtime: ModelRuntime
+
+
+def create_app(data_dir: Path | None = None, engines: dict | None = None) -> FastAPI:
+    @asynccontextmanager
+    async def lifespan(app):
+        root = (data_dir or data_directory()).resolve()
+        database = Database(root)
+        runtime = ModelRuntime(root / "runtime", database.settings)
+        registry = engines if engines is not None else default_engines(runtime)
+        tokens = TokenStore(root)
+        queue = GenerationQueue(database, registry)
+        downloads = Downloads(tokens)
+        app.state.services = Services(
+            root,
+            database,
+            PresetRepository(root),
+            PromptRepository(root),
+            ConfigRepository(root),
+            tokens,
+            registry,
+            queue,
+            downloads,
+            runtime,
+        )
+        try:
+            yield
+        finally:
+            queue.close()
+            runtime.close()
+            downloads.close()
+
+    app = FastAPI(title="CrayonMLX", lifespan=lifespan, docs_url=None, redoc_url=None, openapi_url=None)
+    app.add_middleware(TrustedHostMiddleware, allowed_hosts=["localhost", "127.0.0.1", "[::1]", "testserver"])
+
+    write_lock = asyncio.Lock()
+    app.state.pending_data_dir = None
+
+    @app.middleware("http")
+    async def local_requests(request: Request, call_next):
+        # Local disk operations must not be triggered by a form on a remote website.
+        if request.method not in {"GET", "HEAD", "OPTIONS"}:
+            origin = request.headers.get("origin")
+            if (origin and origin != str(request.base_url).rstrip("/")) or request.headers.get(
+                "sec-fetch-site"
+            ) == "cross-site":
+                return JSONResponse(
+                    {"detail": "Разрешены только запросы из локального интерфейса"}, status_code=403
+                )
+        if request.method not in {"GET", "HEAD", "OPTIONS"}:
+            async with write_lock:
+                if app.state.pending_data_dir:
+                    return JSONResponse(
+                        {
+                            "detail": "Папка данных перенесена. Перезапустите CrayonMLX, чтобы продолжить работу."
+                        },
+                        status_code=409,
+                    )
+                response = await call_next(request)
+        else:
+            response = await call_next(request)
+        response.headers["X-Content-Type-Options"] = "nosniff"
+        response.headers["Content-Security-Policy"] = (
+            "default-src 'self'; img-src 'self' blob:; style-src 'self'; script-src 'self'; frame-ancestors 'none'"
+        )
+        return response
+
+    @app.exception_handler(ValueError)
+    async def invalid_value(request, error):
+        return JSONResponse({"detail": str(error)}, status_code=422)
+
+    @app.exception_handler(OSError)
+    async def filesystem_error(request, error):
+        logging.getLogger(__name__).exception("Filesystem operation failed")
+        return JSONResponse(
+            {
+                "detail": "Не удалось прочитать или записать файл. Проверьте путь, права доступа и свободное место."
+            },
+            status_code=400,
+        )
+
+    web = Path(__file__).parent / "web"
+    app.include_router(router)
+    app.include_router(library_router)
+    app.include_router(folders_router)
+    app.include_router(model_router)
+    app.include_router(storage_router)
+    app.mount("/static", StaticFiles(directory=web / "static"), name="static")
+
+    @app.get("/")
+    def index():
+        return FileResponse(web / "index.html")
+
+    return app

crayonmlx-0.1.0a1/src/crayonmlx/cli.py

@@ -0,0 +1,12 @@
+"""Local-only application entry point."""
+
+import argparse
+
+import uvicorn
+
+
+def main():
+    parser = argparse.ArgumentParser(description="CrayonMLX — локальная студия для рисунков")
+    parser.add_argument("--port", type=int, default=8787)
+    arguments = parser.parse_args()
+    uvicorn.run("crayonmlx.app:create_app", factory=True, host="127.0.0.1", port=arguments.port)

crayonmlx-0.1.0a1/src/crayonmlx/engines/__init__.py

@@ -0,0 +1 @@
+"""CrayonMLX engines components."""

crayonmlx-0.1.0a1/src/crayonmlx/engines/base.py

@@ -0,0 +1,17 @@
+"""Render contract. Adapters render through the application-owned model runtime."""
+
+from pathlib import Path
+from typing import Protocol
+
+from crayonmlx.schemas import Preset
+
+
+class RenderEngine(Protocol):
+    id: str
+    name: str
+    base_models: tuple[str, ...]
+    supports_strength: bool
+
+    def available(self) -> bool: ...
+
+    def render(self, preset: Preset, model: Path, source: Path, output: Path) -> dict | None: ...