diffusers-workflow 0.4.0a12__tar.gz → 0.4.0b1__tar.gz

{diffusers_workflow-0.4.0a12/diffusers_workflow.egg-info → diffusers_workflow-0.4.0b1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffusers-workflow
-Version: 0.4.0a12
+Version: 0.4.0b1
 Summary: Declarative workflow engine and web UI for Hugging Face Diffusers
 Author: dkackman
 License: Apache-2.0
@@ -91,10 +91,11 @@ A workflow JSON file is portable and easy to hand off — but that same flexibil
 ## Features
 
 - **Web UI** — browse and run workflows, edit them in introspection-driven forms, watch jobs stream live progress, manage generated output and the models on disk. `python -m dw.serve` and open a browser. See [Server & Web UI](docs/SERVER.md).
-- **MCP server** — a stdio server that lets an MCP client (Claude Code first) author, validate, save, run and diagnose workflows against a running `dw.serve`. `dw-mcp`. See [MCP Server](docs/MCP.md).
+- **MCP server** — a tool surface that lets an MCP client (Claude Code first) author, validate, save, run and diagnose workflows against a running `dw.serve`: `dw-mcp` over stdio, or mounted at `/mcp` by `dw.serve --mcp` for a client on another machine. See [MCP Server](docs/MCP.md) and [Remote GPU server](docs/REMOTE.md).
+- **Step-output caching** — a step whose resolved arguments and seed are unchanged since the last run in the same process reuses its cached result instead of re-executing. This applies to any `Workflow.run` — REPL iteration and a re-run of a server-submitted job alike, so re-running a fixed-seed workflow from the UI finishes instantly and writes no new output files. The cache holds the most recent 50 steps and is dropped by `memory clear`
 - **Declarative JSON workflows** with variable substitution and cross-step data flow
 - **Multi-step pipelines** — chain text-to-image, image-to-video, inpainting, ControlNet
-- **Reproducible by construction** — outputs embed their full workflow definition and seed; any image in the gallery reopens as the exact workflow that made it (see [Trust model](docs/SECURITY.md#trust-model) before reopening one someone else sent you)
+- **Reproducible by construction** — outputs embed their full workflow definition and seed; any image in the gallery reopens as the exact workflow that made it (see [Trust model](docs/SECURITY.md#trust-model) before reopening one someone else sent you). A rerun never overwrites a prior output — a name collision gets an incrementing suffix instead
 - **Long-video chaining** — run a video pipeline once per segment and stitch the segments into one clip, with audio-driven length and frame-to-frame continuity
 - **Quantization** — BitsAndBytes, TorchAO, GGUF, SDNQ, optimum-quanto
 - **Inference acceleration** — TeaCache, FirstBlockCache, FasterCache, MagCache, TaylorSeerCache
@@ -342,6 +343,7 @@ See [workflows/](workflows/) for more workflow files.
 
 - [Server & Web UI](docs/SERVER.md) — The web UI, jobs API, and introspection service
 - [MCP Server](docs/MCP.md) — Tool surface for MCP clients (Claude Code, Claude Desktop)
+- [Remote GPU server](docs/REMOTE.md) — Using the server, UI and MCP from another machine
 - [Workflow Guide](docs/WORKFLOW_GUIDE.md) — JSON structure, variables, steps, data flow
 - [Quantization](docs/QUANTIZATION.md) — BitsAndBytes, TorchAO, GGUF, SDNQ
 - [Inference Acceleration](docs/ACCELERATION.md) — torch.compile, FirstBlockCache, MagCache, TaylorSeer, TeaCache

{diffusers_workflow-0.4.0a12 → diffusers_workflow-0.4.0b1}/README.md

@@ -16,10 +16,11 @@ A workflow JSON file is portable and easy to hand off — but that same flexibil
 ## Features
 
 - **Web UI** — browse and run workflows, edit them in introspection-driven forms, watch jobs stream live progress, manage generated output and the models on disk. `python -m dw.serve` and open a browser. See [Server & Web UI](docs/SERVER.md).
-- **MCP server** — a stdio server that lets an MCP client (Claude Code first) author, validate, save, run and diagnose workflows against a running `dw.serve`. `dw-mcp`. See [MCP Server](docs/MCP.md).
+- **MCP server** — a tool surface that lets an MCP client (Claude Code first) author, validate, save, run and diagnose workflows against a running `dw.serve`: `dw-mcp` over stdio, or mounted at `/mcp` by `dw.serve --mcp` for a client on another machine. See [MCP Server](docs/MCP.md) and [Remote GPU server](docs/REMOTE.md).
+- **Step-output caching** — a step whose resolved arguments and seed are unchanged since the last run in the same process reuses its cached result instead of re-executing. This applies to any `Workflow.run` — REPL iteration and a re-run of a server-submitted job alike, so re-running a fixed-seed workflow from the UI finishes instantly and writes no new output files. The cache holds the most recent 50 steps and is dropped by `memory clear`
 - **Declarative JSON workflows** with variable substitution and cross-step data flow
 - **Multi-step pipelines** — chain text-to-image, image-to-video, inpainting, ControlNet
-- **Reproducible by construction** — outputs embed their full workflow definition and seed; any image in the gallery reopens as the exact workflow that made it (see [Trust model](docs/SECURITY.md#trust-model) before reopening one someone else sent you)
+- **Reproducible by construction** — outputs embed their full workflow definition and seed; any image in the gallery reopens as the exact workflow that made it (see [Trust model](docs/SECURITY.md#trust-model) before reopening one someone else sent you). A rerun never overwrites a prior output — a name collision gets an incrementing suffix instead
 - **Long-video chaining** — run a video pipeline once per segment and stitch the segments into one clip, with audio-driven length and frame-to-frame continuity
 - **Quantization** — BitsAndBytes, TorchAO, GGUF, SDNQ, optimum-quanto
 - **Inference acceleration** — TeaCache, FirstBlockCache, FasterCache, MagCache, TaylorSeerCache
@@ -267,6 +268,7 @@ See [workflows/](workflows/) for more workflow files.
 
 - [Server & Web UI](docs/SERVER.md) — The web UI, jobs API, and introspection service
 - [MCP Server](docs/MCP.md) — Tool surface for MCP clients (Claude Code, Claude Desktop)
+- [Remote GPU server](docs/REMOTE.md) — Using the server, UI and MCP from another machine
 - [Workflow Guide](docs/WORKFLOW_GUIDE.md) — JSON structure, variables, steps, data flow
 - [Quantization](docs/QUANTIZATION.md) — BitsAndBytes, TorchAO, GGUF, SDNQ
 - [Inference Acceleration](docs/ACCELERATION.md) — torch.compile, FirstBlockCache, MagCache, TaylorSeer, TeaCache

{diffusers_workflow-0.4.0a12 → diffusers_workflow-0.4.0b1/diffusers_workflow.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffusers-workflow
-Version: 0.4.0a12
+Version: 0.4.0b1
 Summary: Declarative workflow engine and web UI for Hugging Face Diffusers
 Author: dkackman
 License: Apache-2.0
@@ -91,10 +91,11 @@ A workflow JSON file is portable and easy to hand off — but that same flexibil
 ## Features
 
 - **Web UI** — browse and run workflows, edit them in introspection-driven forms, watch jobs stream live progress, manage generated output and the models on disk. `python -m dw.serve` and open a browser. See [Server & Web UI](docs/SERVER.md).
-- **MCP server** — a stdio server that lets an MCP client (Claude Code first) author, validate, save, run and diagnose workflows against a running `dw.serve`. `dw-mcp`. See [MCP Server](docs/MCP.md).
+- **MCP server** — a tool surface that lets an MCP client (Claude Code first) author, validate, save, run and diagnose workflows against a running `dw.serve`: `dw-mcp` over stdio, or mounted at `/mcp` by `dw.serve --mcp` for a client on another machine. See [MCP Server](docs/MCP.md) and [Remote GPU server](docs/REMOTE.md).
+- **Step-output caching** — a step whose resolved arguments and seed are unchanged since the last run in the same process reuses its cached result instead of re-executing. This applies to any `Workflow.run` — REPL iteration and a re-run of a server-submitted job alike, so re-running a fixed-seed workflow from the UI finishes instantly and writes no new output files. The cache holds the most recent 50 steps and is dropped by `memory clear`
 - **Declarative JSON workflows** with variable substitution and cross-step data flow
 - **Multi-step pipelines** — chain text-to-image, image-to-video, inpainting, ControlNet
-- **Reproducible by construction** — outputs embed their full workflow definition and seed; any image in the gallery reopens as the exact workflow that made it (see [Trust model](docs/SECURITY.md#trust-model) before reopening one someone else sent you)
+- **Reproducible by construction** — outputs embed their full workflow definition and seed; any image in the gallery reopens as the exact workflow that made it (see [Trust model](docs/SECURITY.md#trust-model) before reopening one someone else sent you). A rerun never overwrites a prior output — a name collision gets an incrementing suffix instead
 - **Long-video chaining** — run a video pipeline once per segment and stitch the segments into one clip, with audio-driven length and frame-to-frame continuity
 - **Quantization** — BitsAndBytes, TorchAO, GGUF, SDNQ, optimum-quanto
 - **Inference acceleration** — TeaCache, FirstBlockCache, FasterCache, MagCache, TaylorSeerCache
@@ -342,6 +343,7 @@ See [workflows/](workflows/) for more workflow files.
 
 - [Server & Web UI](docs/SERVER.md) — The web UI, jobs API, and introspection service
 - [MCP Server](docs/MCP.md) — Tool surface for MCP clients (Claude Code, Claude Desktop)
+- [Remote GPU server](docs/REMOTE.md) — Using the server, UI and MCP from another machine
 - [Workflow Guide](docs/WORKFLOW_GUIDE.md) — JSON structure, variables, steps, data flow
 - [Quantization](docs/QUANTIZATION.md) — BitsAndBytes, TorchAO, GGUF, SDNQ
 - [Inference Acceleration](docs/ACCELERATION.md) — torch.compile, FirstBlockCache, MagCache, TaylorSeer, TeaCache

{diffusers_workflow-0.4.0a12 → diffusers_workflow-0.4.0b1}/diffusers_workflow.egg-info/SOURCES.txt

@@ -30,6 +30,7 @@ dw/security.py
 dw/serve.py
 dw/settings.py
 dw/step.py
+dw/step_cache.py
 dw/teacache.py
 dw/teacache_models.json
 dw/test.py
@@ -48,6 +49,8 @@ dw/server/__init__.py
 dw/server/app.py
 dw/server/enhancers.py
 dw/server/jobs.py
+dw/server/mcp_mount.py
+dw/server/netinfo.py
 dw/server/updater.py
 dw/server/ui/index.html
 dw/server/ui/assets/abap-08VXUWAP.js
@@ -64,45 +67,45 @@ dw/server/ui/assets/csharp-BKxtCVv1.js
 dw/server/ui/assets/csp-bTuwJoIa.js
 dw/server/ui/assets/css-DIMkf-bt.js
 dw/server/ui/assets/css.worker-B3ciXF_0.js
-dw/server/ui/assets/cssMode-CRm5Fmld.js
+dw/server/ui/assets/cssMode-DIA5LESv.js
 dw/server/ui/assets/cypher-CVaqCwHa.js
 dw/server/ui/assets/dart-onAF5SnQ.js
 dw/server/ui/assets/dockerfile-DZFCIeNp.js
 dw/server/ui/assets/ecl-D05T4iGw.js
 dw/server/ui/assets/editor-jjEx9u7D.css
-dw/server/ui/assets/editor.api-CArwgKCZ.js
+dw/server/ui/assets/editor.api-CuOF-U_A.js
 dw/server/ui/assets/editor.worker-q-txB4vs.js
 dw/server/ui/assets/elixir-6RTg0lbw.js
 dw/server/ui/assets/flow9-C5_-GSwl.js
-dw/server/ui/assets/freemarker2-B42MpbWP.js
+dw/server/ui/assets/freemarker2-xMcLxHzc.js
 dw/server/ui/assets/fsharp-C8Ef5oNN.js
 dw/server/ui/assets/go-C-y9NEjX.js
 dw/server/ui/assets/graphql-fmXr3nnJ.js
-dw/server/ui/assets/handlebars-BhleeD_A.js
+dw/server/ui/assets/handlebars-Dbt47IUh.js
 dw/server/ui/assets/hcl-CpzslTdj.js
-dw/server/ui/assets/html-Bi_v7qEe.js
+dw/server/ui/assets/html-C4QoqgED.js
 dw/server/ui/assets/html.worker-C93Ht9o9.js
-dw/server/ui/assets/htmlMode-c39yjrZR.js
-dw/server/ui/assets/index-CKHZ9ZUD.css
-dw/server/ui/assets/index-Do_ZJ3yS.js
+dw/server/ui/assets/htmlMode-BQ_x-rQl.js
+dw/server/ui/assets/index-BCNGRglb.css
+dw/server/ui/assets/index-C14JKFhY.js
 dw/server/ui/assets/ini-sBoK_t0W.js
 dw/server/ui/assets/java-BEtHBSE6.js
-dw/server/ui/assets/javascript-BTuzJzdO.js
+dw/server/ui/assets/javascript-yDlADdyl.js
 dw/server/ui/assets/json.worker-B2V3pomh.js
-dw/server/ui/assets/jsonMode-DY2FmMCP.js
+dw/server/ui/assets/jsonMode-X4znCY4b.js
 dw/server/ui/assets/julia-Bri6UV-V.js
 dw/server/ui/assets/kotlin-BOotOW0E.js
 dw/server/ui/assets/less-B9JPFI3C.js
 dw/server/ui/assets/lexon-CfSJPG6W.js
-dw/server/ui/assets/liquid-iXwDEnQP.js
-dw/server/ui/assets/lspLanguageFeatures-CVXyTq9V.js
+dw/server/ui/assets/liquid-CAaCQzwu.js
+dw/server/ui/assets/lspLanguageFeatures-CTJIGm94.js
 dw/server/ui/assets/lua-CsQS60Ue.js
 dw/server/ui/assets/m3-D-oSqn_W.js
 dw/server/ui/assets/markdown-Cimd5fb3.js
-dw/server/ui/assets/mdx-QE70XlV0.js
+dw/server/ui/assets/mdx-ByHLng1y.js
 dw/server/ui/assets/mips-CIPQ_RoX.js
 dw/server/ui/assets/monaco--ixms01u.css
-dw/server/ui/assets/monaco-CzCru1EC.js
+dw/server/ui/assets/monaco-LJZc6TiS.js
 dw/server/ui/assets/msdax-DauUninz.js
 dw/server/ui/assets/mysql-SOo6toE5.js
 dw/server/ui/assets/objective-c-FvmIjYaQ.js
@@ -117,10 +120,10 @@ dw/server/ui/assets/powerquery-D3hlyOfw.js
 dw/server/ui/assets/powershell-DmHpPYUd.js
 dw/server/ui/assets/protobuf-C531GsRP.js
 dw/server/ui/assets/pug-Z5eAx3Zn.js
-dw/server/ui/assets/python-DxHS-kHp.js
+dw/server/ui/assets/python-BCT_lvfW.js
 dw/server/ui/assets/qsharp-DkqhCAOL.js
 dw/server/ui/assets/r-BwWrilGY.js
-dw/server/ui/assets/razor-CT5Qrhwg.js
+dw/server/ui/assets/razor-Bw9oTolL.js
 dw/server/ui/assets/redis-ClamHrr6.js
 dw/server/ui/assets/redshift-DT7zqm-g.js
 dw/server/ui/assets/restructuredtext-BYgofb2h.js
@@ -140,21 +143,22 @@ dw/server/ui/assets/swift-Bxkupp3x.js
 dw/server/ui/assets/systemverilog-Bz4Y3fRF.js
 dw/server/ui/assets/tcl-DISqw1ZD.js
 dw/server/ui/assets/ts.worker-D7T1-Ig5.js
-dw/server/ui/assets/tsMode-qXdDdnF1.js
+dw/server/ui/assets/tsMode-BGvHTpYz.js
 dw/server/ui/assets/twig-De2hgUGE.js
-dw/server/ui/assets/typescript-BWfJagji.js
+dw/server/ui/assets/typescript-mwKKpQXP.js
 dw/server/ui/assets/typespec-B8J7ngcE.js
 dw/server/ui/assets/vb-DV3o63ZY.js
 dw/server/ui/assets/wgsl-DpFanUEy.js
-dw/server/ui/assets/workers-C9vxrwoC.js
-dw/server/ui/assets/xml-DhJEwsDK.js
-dw/server/ui/assets/yaml-CiixMxAq.js
+dw/server/ui/assets/workers-BOhWN9wu.js
+dw/server/ui/assets/xml-DuQ6E6P7.js
+dw/server/ui/assets/yaml-D7o22x0f.js
 dw/tasks/audio_utils.py
 dw/tasks/background_remover.py
 dw/tasks/borders.py
 dw/tasks/concat_videos.py
 dw/tasks/depth_estimator.py
 dw/tasks/diffusion_upscale.py
+dw/tasks/dissolve_videos.py
 dw/tasks/format_messages.py
 dw/tasks/gather.py
 dw/tasks/image_to_text.py
@@ -166,6 +170,7 @@ dw/tasks/qr_code.py
 dw/tasks/restore_faces.py
 dw/tasks/rife_model.py
 dw/tasks/segment.py
+dw/tasks/stabilize.py
 dw/tasks/task.py
 dw/tasks/tensor_image.py
 dw/tasks/text_generation.py
@@ -203,6 +208,7 @@ tests/test_device.py
 tests/test_device_helpers.py
 tests/test_device_portability.py
 tests/test_diffusion_upscale.py
+tests/test_dissolve_videos.py
 tests/test_docs_links.py
 tests/test_events.py
 tests/test_examples.py
@@ -213,15 +219,18 @@ tests/test_image_utils.py
 tests/test_integration.py
 tests/test_interpolate_frames.py
 tests/test_introspection.py
+tests/test_jobs_reused.py
 tests/test_list_images.py
 tests/test_mcp_authoring.py
 tests/test_mcp_catalog.py
 tests/test_mcp_client.py
 tests/test_mcp_diagnose.py
+tests/test_mcp_main.py
 tests/test_mcp_media.py
 tests/test_mcp_models.py
 tests/test_mcp_prompts.py
 tests/test_mcp_server.py
+tests/test_mix_audio.py
 tests/test_model_cache.py
 tests/test_modular_pipeline.py
 tests/test_offload_placement.py
@@ -233,6 +242,7 @@ tests/test_previous_results.py
 tests/test_prompt_parser.py
 tests/test_prompt_weighting.py
 tests/test_prompts.py
+tests/test_recenter_crop.py
 tests/test_repl_commands.py
 tests/test_repl_ergonomics.py
 tests/test_repl_hierarchical.py
@@ -240,12 +250,19 @@ tests/test_repl_reorganization.py
 tests/test_repl_worker.py
 tests/test_resize_bucket.py
 tests/test_result.py
+tests/test_result_output_naming.py
 tests/test_scheduler_config.py
 tests/test_schema.py
 tests/test_security.py
 tests/test_segment.py
+tests/test_serve_main.py
 tests/test_server.py
+tests/test_server_downloads.py
+tests/test_server_info.py
+tests/test_server_mcp.py
+tests/test_stabilize.py
 tests/test_step.py
+tests/test_step_cache.py
 tests/test_strip_exif_and_watermark.py
 tests/test_task.py
 tests/test_task_discovery.py
@@ -260,4 +277,5 @@ tests/test_video_utils.py
 tests/test_worker.py
 tests/test_worker_execute.py
 tests/test_workflow.py
+tests/test_workflow_step_cache.py
 tests/test_workflow_trust.py

{diffusers_workflow-0.4.0a12 → diffusers_workflow-0.4.0b1}/dw/pipeline_processors/chain.py

@@ -35,8 +35,7 @@ import torch
 from diffusers.utils import encode_video, is_av_available
 
 from .. import empty_device_cache
-from ..result import AudioVideo, get_artifact_list
-from ..security import validate_output_path
+from ..result import AudioVideo, get_artifact_list, output_file_path
 from ..tasks.audio_utils import (
     as_channels_samples,
     equal_power_crossfade_join,
@@ -165,6 +164,11 @@ class SegmentedFrames:
         self.paths = list(paths)
         self.total_frames = total_frames
         self.keep_files = keep_files
+        # True once cleanup() has actually removed the files (not when
+        # keep_files skipped that) - a cached Result pointing at a cleaned-up
+        # SegmentedFrames must not be served to a later miss, since the files
+        # a replay would open are gone. See Result.retainable.
+        self.cleaned = False
 
     def __len__(self):
         """Chunk count - one per segment file."""
@@ -191,6 +195,7 @@ class SegmentedFrames:
                 os.remove(path)
             except FileNotFoundError:
                 pass
+        self.cleaned = True
 
 
 class SegmentSpill:
@@ -240,12 +245,14 @@ class SegmentSpill:
                 chain leaves fully playable segments
             sample_rate: Sample rate of that audio
         """
-        path = validate_output_path(
-            os.path.join(
-                self.output_dir,
-                f"{self.base_name}.segment-{len(self.paths):03d}.mp4",
-            ),
+        # A crashed or completed chain leaves segment files behind on disk,
+        # ready to salvage - and the per-wrapper iteration counter (base_name)
+        # restarts at 0 in a fresh process, so a rerun must dedupe through the
+        # same '-N' convention every other output goes through rather than
+        # silently overwriting them.
+        path = output_file_path(
             self.output_dir,
+            f"{self.base_name}.segment-{len(self.paths):03d}.mp4",
         )
 
         audio_track = None

{diffusers_workflow-0.4.0a12 → diffusers_workflow-0.4.0b1}/dw/previous_results.py

@@ -6,6 +6,7 @@ from .arguments import (
     PREVIOUS_RESULT_PREFIX,
     build_objects,
 )
+from .step_cache import reference_resolves_to
 
 logger = logging.getLogger("dw")
 
@@ -126,13 +127,14 @@ def get_previous_results(previous_results, previous_result_name):
             f"Previous result '{previous_result_name}' not found. Available results: {list(previous_results.keys())}"
         )
 
-    # Find the longest known step name that is a prefix of the reference
-    # followed by ".", and treat the remainder as the property name.
+    # Find the longest known step name the reference resolves to, and treat
+    # the remainder as the property name. The exact-match case returned
+    # above, so every name reaching here is a strict '<name>.' prefix.
     result_name = max(
         (
             name
             for name in previous_results
-            if previous_result_name.startswith(name + ".")
+            if reference_resolves_to(previous_result_name, name)
         ),
         key=len,
         default=None,

{diffusers_workflow-0.4.0a12 → diffusers_workflow-0.4.0b1}/dw/repl_commands.py

@@ -279,7 +279,10 @@ class MemoryCommands:
         if not arg or arg == "?":
             print("\nMemory commands:")
             print("  memory show  - Show the worker's GPU memory usage")
-            print("  memory clear - Release cached models and empty the device cache")
+            print(
+                "  memory clear - Release cached models, cached step results, "
+                "and empty the device cache"
+            )
             print()
             print("The worker process starts on the first 'workflow run', so there")
             print("is nothing to show before that. 'workflow restart' frees")

{diffusers_workflow-0.4.0a12 → diffusers_workflow-0.4.0b1}/dw/result.py

@@ -30,8 +30,28 @@ def output_file_path(output_dir, file_name):
     they are concatenated into a file name. Without this a name carrying a
     path separator would write outside the output directory, so the joined
     path goes through the same validator every other path in the engine does.
+
+    A rerun that would otherwise produce a name already on disk gets a
+    '-2', '-3', ... counter instead of silently overwriting it - the same
+    guarantee ComfyUI's SaveImage node makes by scanning its output
+    directory before every write.
     """
-    return validate_output_path(os.path.join(output_dir, file_name), output_dir)
+    candidate = validate_output_path(os.path.join(output_dir, file_name), output_dir)
+    return _dedupe_existing_path(candidate)
+
+
+def _dedupe_existing_path(path):
+    """Append an incrementing counter before the extension until `path` is free."""
+    if not os.path.exists(path):
+        return path
+
+    base, ext = os.path.splitext(path)
+    counter = 2
+    while True:
+        candidate = f"{base}-{counter}{ext}"
+        if not os.path.exists(candidate):
+            return candidate
+        counter += 1
 
 
 # Audio content types soundfile can write, mapped to their file extension and to any
@@ -142,6 +162,25 @@ class Result:
             logger.debug("Adding single result to result list")
             self.result_list.append(result)
 
+    @property
+    def retainable(self):
+        """Whether the step cache may keep this Result's result_list.
+
+        A chain step's save_segments spills each segment to disk and
+        replays it lazily through a SegmentedFrames; save() removes those
+        files once the video is written (SegmentedFrames.cleanup(), called
+        from save_audio_video below). A Result cached after that point would
+        point a later cache hit at files that are already gone, so a
+        result_list holding such an artifact is not retainable - checked by
+        duck-typed attribute rather than isinstance, so this module need not
+        import pipeline_processors.chain.
+        """
+        for result in self.result_list:
+            frames = getattr(result, "frames", None)
+            if getattr(frames, "cleaned", False):
+                return False
+        return True
+
     def get_artifacts(self):
         """Retrieve all artifacts from stored results.
 

{diffusers_workflow-0.4.0a12 → diffusers_workflow-0.4.0b1}/dw/serve.py

@@ -10,6 +10,7 @@ not to the network. Interactive API docs at http://127.0.0.1:8765/docs
 import argparse
 import multiprocessing
 import os
+import sys
 
 # Spawn start method before anything touches multiprocessing (CUDA/MPS)
 if multiprocessing.get_start_method(allow_none=True) != "spawn":
@@ -64,10 +65,36 @@ def main():
         "anything can submit jobs to (including an MCP client) should "
         "stay untrusted.",
     )
+    parser.add_argument(
+        "--mcp",
+        action="store_true",
+        default=False,
+        help="Also serve the MCP tool surface at /mcp over Streamable HTTP, "
+        "behind the same token as /api, so an agent on another machine "
+        "needs no local install (`claude mcp add --transport http dw "
+        "http://<host>:<port>/mcp --header 'Authorization: Bearer <token>'`). "
+        "Refused on a non-loopback --host without a token.",
+    )
     args = parser.parse_args()
 
     token = args.token or os.environ.get("DW_API_TOKEN") or None
 
+    from .server.app import LOOPBACK_HOSTS
+
+    # A hard error, where the REST-only case below is a warning: an MCP
+    # endpoint can author and run workflows, and unlike the web UI there is
+    # no page to paste a token into. Raised before startup() and before the
+    # worker subprocess is ever spawned.
+    if args.mcp and args.host not in LOOPBACK_HOSTS and not token:
+        print(
+            f"dw-serve: --mcp on {args.host} needs a token. An MCP endpoint "
+            "can author and run workflows, and unlike the web UI there is "
+            "no page to type a token into - pass --token or set "
+            "DW_API_TOKEN, or bind to 127.0.0.1.",
+            file=sys.stderr,
+        )
+        raise SystemExit(2)
+
     # Set before create_app / before the worker subprocess is ever spawned -
     # 'spawn' launches a fresh interpreter that inherits this environment
     # variable, so the job runner sees the same trust choice the API does
@@ -102,8 +129,6 @@ def main():
 
     logger = logging.getLogger("dw")
 
-    from .server.app import LOOPBACK_HOSTS
-
     if args.host not in LOOPBACK_HOSTS and not token:
         logger.warning(
             "Binding to %s with no API token configured (--token or "
@@ -126,10 +151,17 @@ def main():
         prompt_dir=prompt_dir,
         host=args.host,
         token=token,
+        mcp=args.mcp,
+        port=args.port,
     )
     ui = " - UI at /" if default_ui_dir() else ""
+    mcp = " - MCP at /mcp" if args.mcp else ""
     print(
-        f"diffusers-workflow server on http://{args.host}:{args.port}  (docs at /docs{ui})"
+        f"diffusers-workflow server on http://{args.host}:{args.port}"
+        f"  (docs at /docs{ui}{mcp})",
+        # stdout is a pipe under systemd or nohup, where block buffering
+        # would otherwise hold this line back until shutdown
+        flush=True,
     )
     uvicorn.run(app, host=args.host, port=args.port, log_level=args.log_level.lower())