---

name: workflow-processing description: "Route Roboflow Inference workflow CLI commands for processing images, image directories, and videos, including parameter merging, output layout, aggregation, overwrite controls, rate limits, and API versus local execution." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

Workflow Processing

Use this sub-skill when a user wants to run Roboflow Workflows from the inference CLI over a single image, an image directory, or a video file, or when they need to understand the output files produced by those commands.

Route here when

• The user mentions inference workflows process-image, process-images-directory, or process-video.
• They need to load a Workflow from --workflow_spec or from --workspace_name plus --workflow_id / --workflow_version_id.
• They need to pass Workflow parameters via --workflow_params and extra CLI arguments, choose api versus inference_package, or tune --max_fps, --max-failures, --threads, --allow_override, or --force_reprocessing.
• They ask what files to expect in an output directory, why aggregation is CSV or JSONL, or why a previous run was skipped or refused to overwrite output.

Do not route here when

• The user is asking how to start, configure, or expose the Inference server. Use CLI operations instead.
• The user is asking for Python InferenceHTTPClient, HTTP request/response shapes, WebRTC streams, or long-running pipeline API calls. Use SDK and WebRTC instead.
• The user is asking about model backend/package selection outside Workflow CLI execution; keep that with the model-runtime sub-skill when available.

First decisions to make

1. Input kind: one image, an image directory, or one video file.
2. Workflow source: either a local JSON Workflow specification via --workflow_spec, or a Roboflow-hosted Workflow via --workspace_name and --workflow_id with optional --workflow_version_id. Do not tell users to supply both sources for the same run.
3. Processing target:
   • Image and directory commands default to --processing_target api and can also run through --processing_target inference_package.
   • Video processing has no --processing_target; it always requires the local inference package because it uses the local video pipeline.
4. Output safety: choose a clean --output_dir unless the user explicitly wants to reuse or merge into prior outputs with --allow_override.
5. Parameters: combine --workflow_params params.json with trailing Workflow parameters only when needed; trailing CLI parameters override matching file keys.
6. Result format: use directory --aggregation_format csv|jsonl for process-images-directory; use video --output_file_type csv|jsonl for process-video.

Read these bundled references

• Workflow command patterns and output behavior for command recipes, parameter merging, output layout, aggregation, target choice, and resume/overwrite behavior.
• CLI reference for the workflow command family and option-level routing.
• Troubleshooting for non-empty directories, invalid JSON paths, missing local packages, target confusion, --max_fps, --max-failures, aggregation format surprises, and debug-mode advice.

A small optional helper is bundled at scripts/inspect_workflow_outputs.py. Use it only to validate JSON config files or inspect an already-produced output folder; it is not a replacement for the Roboflow CLI.

A separate helper, scripts/provision_workflow_fonts.py, provisions the approved workflow font assets when a workflow visualization path needs the bundled font set outside the original repository checkout.

Answer pattern

When responding to a workflow-processing request, provide:

1. The exact inference workflows ... command for the input kind.
2. The Workflow source choice (--workflow_spec or hosted identifiers).
3. The processing target and any package/API preconditions.
4. The expected output layout and which files are safe to consume downstream.
5. The overwrite/resume/rate/failure controls that matter for the user's run.
6. A troubleshooting branch, with --debug_mode only when a stack trace is useful.