Virtual Icechunk · IOOS Hackathon 2026

Building a virtual Icechunk store

The skill does the building; you supply what it cannot know

Eli Holmes1
1NOAA Fisheries

Our skill builds up to production

Step You get It must
1 · Plan plan.md: files, layout, decisions Be reviewed by you before any code
2 · Smoke test A notebook that builds a local store from a few files Reopen it read-only and read real values
3 · Production A build script: settings at the top, batches, commits, restart Survive being interrupted; tag the result
4 · Docs A README, a loading example, the tested environment Let a stranger open the store with no help

It stops for your review between steps.

What the agent needs from you

It cannot guess these. Put them in your first prompt:

  • Where the files are, and how to get in: a bucket, a THREDDS catalog, an API; public or which credentials
  • Which files: products, variables, years
  • Where the store goes: a bucket you can write to. Write credentials go in the environment, never in the chat
  • Who will read it: Python users, a browser viewer, users outside one cloud region
  • Built once, or kept up to date as new files arrive

What the skill works out from your files

It will propose these, and ask you to confirm:

  • One store, or several? The goal is one store with all the variables. Files on different grids or encodings force groups or separate stores
  • How the files fit together: joined along time, variables in separate files merged
  • What to load, what to keep virtual: coordinates loaded, science arrays left virtual
  • Metadata fixes to follow CF conventions

The files usually decide these. You can still ask “why?”, and override.

Telling the agent to build it

“Build a virtual Icechunk store of the NDVI files in s3://noaa-cdr-ndvi-pds/data/ (public, us-east-1) for 2000–2024. Publish it to my Source Cooperative repository my-org/ndvi; my write keys are in environment variables. Python users and a browser viewer will read it. Start with the plan.”

Then, one session per step:

“The plan is approved. Write the smoke-test notebook.” (builds a local test store)

“The smoke test passed. Write the production build script.” (writes to your bucket)

“Write the README and a loading example.”

Authenticating to your bucket

The agent writes to your bucket with your credentials. Tell it where they are:

Destination Typical credentials
Amazon S3 An AWS profile, or AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY in the environment
Google Cloud Storage A service-account key file, named in GOOGLE_APPLICATION_CREDENTIALS
Source Cooperative Temporary S3 keys from your repository’s View credentials page. They expire

Not sure? Ask it:

“Help me set up write access to my Source Cooperative repository. Do not print my keys, and do not save them in any file in this repository.”

Worked examples

Next: visualizing

Chapter 6 · Visualizing