Workflow Tools

mere-workflow-tools is a shared package that installs six companion commands and graph contract tooling:

• mere-doc-tools
• mere-media-scrub
• mere-dataset-tools
• mere-transcript-tools
• mere-image-compose
• mere-batch-runner
• mere-graph-conformance
• mere-graph-compile

Each command follows the same plugin fashion:

<tool> manifest --json
<tool> doctor
<tool> plan ...
<tool> run ./run.json
<tool> resume ./run.json
<tool> cleanup ./run.json

The one-shot command for each tool writes a manifest and immediately executes it.

Runtime Boundary

The package does not include OCR, ASR, captioning, anonymization, segmentation, or image-generation runtimes. It shells out to mere.run and records the exact planned steps. Override the executable with MERE_WORKFLOW_TOOLS_MERE_RUN or --mere-run-command.

Commands

Document OCR plus redaction:

mere-doc-tools process \
  --input ./scan.png \
  --output-dir ./doc-out

Media OCR scrub over a frame directory:

mere-media-scrub scrub \
  --input ./frames \
  --output-dir ./scrub-out

Dataset captioning for LoRA prep:

mere-dataset-tools caption \
  --input ./dataset \
  --output-dir ./caption-out \
  --trigger-token STYLE \
  --focus "border" \
  --focus "lighting"

Audio transcription plus redaction:

mere-transcript-tools transcribe \
  --input ./meeting.wav \
  --output-dir ./transcript-out

Image generation composition:

mere-image-compose generate \
  --prompt "a sharp product render" \
  --output-dir ./image-out \
  --model image-klein-9b

Batch execution:

mere-batch-runner run-jobs \
  --jobs ./jobs.jsonl \
  --output-dir ./batch-out

Each JSONL batch row must contain an argv string array. Optional outputs entries tell the runner which artifacts to hash after execution:

{"argv":["text","anonymize","--output","./redacted.txt"],"outputs":{"redacted":"./redacted.txt"}}

Portable Graphs

mere-dataset-tools exposes dataset.prepare through the fixed graph-provider protocol. The package SDK centralizes confined invocation decoding, diagnostic records, ordered event streams, and catalog conformance:

mere-graph-conformance --provider mere-dataset-tools --json
mere-dataset-tools graph templates list --json
mere-dataset-tools graph templates export lora-train-sample \
  --output ./workflow.json --json

The native template catalog includes dataset-to-LoRA-to-sample, image-to-video, Creative Prompt Lab, and Describe and Remix workflows. The creative templates compose the portable text.value, text.join, text.template, text.enhance, and image.describe catalog nodes without a Studio-specific graph format. The canonical graph and run schemas are mirrored in contracts/ and validated with every repository gate.

Reusable Composition

Compile higher-order authoring documents into an ordinary portable graph:

mere-graph-compile ./program.json \
  --variables-json ./production-values.json \
  --output ./workflow.json \
  --report-output ./compile.json \
  --json

Workflow programs support inline or confined imported modules, compile-time conditions, and deterministic maps of up to 1,000 instances. Mapped outputs use an explicit zero-based instance index. Compilation resolves every module and branch, rewrites references, detects ID collisions, and records source and graph fingerprints. The resulting mere.run/workflow-graph is what local, SSH, and Relay executors consume; no higher-order behavior enters the worker protocol.

Graph Studio

Visual authoring is provided by the separate mere-run-graph-studio application. This package remains its provider boundary for catalog entries, native templates, workflow compilation, and conservative ComfyUI import. The Studio app invokes these public commands and does not import plugin internals.

ComfyUI Bridge

The bridge is an importer, not a second runtime. It inspects ComfyUI UI or API JSON and imports a supported API prompt subset into a native graph:

mere-dataset-tools graph comfy inspect ./comfy-workflow.json --json
mere-dataset-tools graph comfy import ./comfy-api.json \
  --model image-krea2-turbo \
  --output ./workflow.json \
  --inputs-output ./inputs.json \
  --json

Checkpoint, CLIP text, KSampler, latent-image, optional image input, optional LoRA, VAE decode, and save nodes are recognized. Custom nodes and UI-only exports remain inspectable but block import with explicit diagnostics. ComfyUI sampler and scheduler choices are reported as omitted warnings when the native node contract has no equivalent.

Why These Are Plugins

These are local production workflows, not new model capabilities. Keeping them as companion plugins lets mere.run stay focused on inference while users get repeatable, inspectable pipelines for documents, media, datasets, transcripts, image generation, and batch automation.