nimbus-worker-scaffold

name	nimbus-worker-scaffold
description	Scaffold a new NimbusImage Docker worker end to end, including entrypoint, Dockerfiles and GPU labels, tests, Compose registration, WORKERNAME.md, and REGISTRY.md. Use when adding, creating, or porting a new annotation, property, image-processing, ML, or test worker. Use nimbus-worker-hardening for existing-worker fixes and nimbus-interface for runtime API details.

Scaffolding a NimbusImage Worker

Adding a worker is the single most common change in this repo, and it is easy to ship one that runs but is missing a label, a test, or a registry entry — gaps that only surface in a production deploy. This skill exists so a new worker lands complete and correct the first time.

A worker is a Docker image that NimbusImage runs as a fresh container per job (docker run --rm <image> --apiUrl ... --token ... --request compute|interface --parameters ...). It pulls images/annotations from a Girder backend and returns either annotations or property values.

This skill covers the structure of a worker. For the runtime API (image access, annotation CRUD, coordinate conventions, local testing), read the nimbus-interface skill — don't re-derive that here.

Step 1 — Decide the worker's shape

Three questions determine every downstream choice. Settle them first (ask the user if the request is ambiguous):

1. What does it output?

   • Creates annotations (segmentation, spot detection, generators) → annotation worker (workers/annotations/<name>/). Label isAnnotationWorker + annotationShape (polygon | point | line).
   • Computes numbers on existing annotations → property worker (workers/properties/{blobs,points,lines,connections}/<name>_worker/). Label isPropertyWorker + annotationShape.
   • Processes images and writes a new image back to Girder (blur, registration, restoration) → annotation-style worker that uploads a TIFF, not annotations.

2. Does it need a GPU? ML models on CUDA → yes. Classical CV / numpy → no. This sets the base image and the mandatory isGPUWorker label (Step 3).

3. What's the interface? Which of the interface types (number, text, select, checkbox, channel, channelCheckboxes, tags, layer, notes) does the user need, and what does each return? The type→return-type table lives in the project AGENTS.md; the tags type returns a plain list of strings (a common crash source), and channelCheckboxes returns a dict. workers/annotations/sample_interface/entrypoint.py demonstrates every type.

Step 2 — Write entrypoint.py

Every entrypoint has the same three parts. Copy a working template from references/templates.md rather than writing from scratch — pick the one that matches Step 1:

• interface(image, apiUrl, token) — builds the UI dict and calls client.setWorkerImageInterface(image, interface). Give every field a displayOrder; add a notes field at displayOrder: 0 describing the tool.
• compute(datasetId, apiUrl, token, params) — the work.
  • Annotation workers should use WorkerClient (worker.process(f, f_annotation=...)), which handles Batch XY/Z/Time iteration and upload for you. Model on workers/annotations/random_squares/entrypoint.py.
  • Property workers build {annotation_id: {prop: value}}, wrap as {datasetId: property_value_dict}, and call add_multiple_annotation_property_values. Model on workers/properties/blobs/blob_intensity_worker/entrypoint.py.
• __main__ — the argparse block with the match args.request dispatch to compute / interface. It is boilerplate; copy it verbatim.

Read the field values out of params['workerInterface'] using the return types from the table above. Validate required inputs (especially tags) early and sendError with an actionable message — a worker that crashes with a bare traceback gives the user nothing to act on.

Step 3 — Write the Dockerfile(s) and the label set

Pick the base image from Step 1:

Worker kind	Base image	ENTRYPOINT
Property / simple annotation (conda)	nimbusimage/worker-base:latest	inherited from base (omit ENTRYPOINT)
Image-processing (needs extra CV libs)	nimbusimage/image-processing-base:latest	inherited from base
GPU / ML	nvidia/cuda:<ver>-cudnn*-devel-ubuntu22.04 + conda (multi-stage)	run_worker.sh /entrypoint.py
Test/demo worker	nimbusimage/test-worker-base:latest (micromamba)	_entrypoint.sh python /entrypoint.py, COPY --chown=$MAMBA_USER

The label block is not optional and not cosmetic — the dispatcher reads it. Every production Dockerfile (and any Dockerfile_M1, in the final stage of a multi-stage build) must set. This valid annotation-worker example is CPU-only:

LABEL isUPennContrastWorker="" \
      isGPUWorker="false" \
      isAnnotationWorker="" \
      annotationShape="polygon" \
      interfaceName="Human-readable name" \
      interfaceCategory="Grouping shown in the UI" \
      description="One line describing what it does"

Use isGPUWorker="true" for a GPU worker. For a property worker, replace isAnnotationWorker with isPropertyWorker. Set annotationShape to polygon, point, or line for property and shaped annotation workers; omit it for image-processing workers that do not emit annotations.

isGPUWorker decides whether the job goes to the gpu or cpu Celery queue. An unlabeled image silently falls back to the gpu queue in production and logs a warning — so a CPU worker missing the label wastes a GPU slot, and this is exactly the gap PRs #148/#153 existed to close across every worker. Never leave it unset.

A Dockerfile_M1 is optional. It's the CPU/arm64 variant that build_workers.sh selects when MAC_DEVELOPMENT_MODE=true (or on an arm64 host) — the two are the same mechanism, not alternatives — and it's the only way to build the worker on a Mac. Add one only if you want Mac buildability; plenty of GPU workers (including cellposesam) ship without a Dockerfile_M1 and simply can't be built on a Mac. Note this is separate from a runtime CPU fallback (try CUDA, fall back to CPU), which lets the production x86 image run on a GPU-less host. See references/templates.md for the multi-stage CUDA + label-in-final-stage pattern.

Step 4 — Tests

Create tests/ with __init__.py, test_<name>.py, and Dockerfile_Test. Tests run against a mocked annotation_client — no live server. Cover at minimum: (a) interface() produces the expected fields/types, and (b) compute() produces the expected annotations/properties given mocked tiles. Mock UPennContrastWorkerPreviewClient, tiles.UPennContrastDataset, and annotations.UPennContrastAnnotationClient. Template in references/templates.md; model on workers/annotations/random_squares/tests/.

Per the project's test-first convention, write these tests before the implementation when practical, and make sure they fail before they pass.

Step 5 — Register and document

A worker the agent forgot to register won't build; one it forgot to document drifts. Do all of these:

1. docker-compose.yml — add a build service and a _test service. Match the profiles to the worker kind:

   • annotation/property production worker → profiles: ["worker", "annotations"] or ["worker", "properties"]; test service → profiles: ["test"].
   • test/demo worker → profiles: ["testworker"]; its test → ["testworkertest"].

     my_worker:
       build:
         context: .
         dockerfile: ./workers/annotations/my_worker/${DOCKERFILE:-Dockerfile}
       image: annotations/my_worker:latest
       depends_on: [worker-base]
       profiles: ["worker", "annotations"]
     my_worker_test:
       build:
         context: .
         dockerfile: ./workers/annotations/my_worker/tests/Dockerfile_Test
       image: annotations/my_worker:test
       depends_on: [my_worker]
       profiles: ["test"]
   

2. WORKERNAME.md — a hand-written doc in the worker directory (uppercase, e.g. MY_WORKER.md). Cover: title/description, how it works, an interface parameter table, computed properties (property workers), implementation notes, and GPU/limitations. Model on random_squares/RANDOM_SQUARES.md (concise) or deconwolf/DECONWOLF.md (comprehensive). Keep the labels in the Dockerfile and the doc consistent.

3. REGISTRY.md — add the worker to the correct category table. This is the master index and must be updated on every add/remove/rename.

Do not run generate_worker_docs.py --force — it overwrites hand-written docs with stubs. The auto-doc hooks are disabled on purpose; docs are manual.

Step 6 — Build and verify

./build_workers.sh <name>                          # build just this worker
./build_workers.sh --build-and-run-tests <name>    # build + run its tests
# ML/GPU worker:
./build_machine_learning_workers.sh
# test/demo worker:
./build_test_workers.sh --build-and-run-tests <name>
# Mac dev of a GPU worker (builds Dockerfile_M1 — the worker must have one):
MAC_DEVELOPMENT_MODE=true ./build_workers.sh <name>

Don't claim the worker is done until the build succeeds and its tests pass — show the output.

Completion checklist

Before calling a new worker complete, confirm every item — a missing one is the usual cause of a worker that builds locally but misbehaves in deploy:

• entrypoint.py with interface(), compute(), and the __main__ dispatch
• Dockerfile with the full label set, including isGPUWorker
• Dockerfile_M1 only if the worker should be Mac-buildable (optional; it's the variant MAC_DEVELOPMENT_MODE=true selects)
• tests/ with __init__.py, test_<name>.py, Dockerfile_Test
• docker-compose.yml: worker service and _test service, correct profiles
• WORKERNAME.md in the worker directory
• REGISTRY.md updated
• ./build_workers.sh --build-and-run-tests <name> passes