ocean-science

Physical oceanography for Open Science Pillars: load the ECCO v4r4 state estimate and SWOT KaRIn sea surface height through a size gate, compute meridional transports, close property budgets on the native grid, identify water masses, attribute sea level, and compare the result with observations, with an uncertainty on every headline number and the knowledge concept behind every choice cited. It is a Hydrosphere capability, discipline Ocean Physics (kind: capability in .osp/repository.yaml); the words used on this page (capability, plugin, sphere, knowledge bundle, runtime) are defined in the glossary.

How it is organized: its knowledge is the PO.DAAC provider bundle, reviewed by its maintainers and confirmed by the provider where a contact has confirmed it, consulted in place, plus the local conventions, connectors and attested computations here; its behavior is the skills, each computation's executor and attester beside the skill that runs them; its verification is the golden notebooks here, which prove the scripts those computations name; its connectors are declared in .osp/package.yaml, from which the Claude package files are rendered.

Install

On Claude Code:

claude plugin marketplace add open-science-pillars/marketplace
claude plugin install ocean-science@open-science-pillars

What comes with it: core (the foundation skills and the start, discover-data and report workflows) and nasa-daac-knowledge (the PO.DAAC provider bundle), both declared as dependencies, so the one install brings them with it. An install stays at the release it was installed from: claude plugin update ocean-science@open-science-pillars moves this plugin and only this plugin; a dependency moves by its own update command (claude plugin update core@open-science-pillars, claude plugin update nasa-daac-knowledge@open-science-pillars), and a release that raises a floor says so in its notes. claude plugin list shows what you have.

On Claude Cowork: add the marketplace by repository (open-science-pillars/marketplace) under Customize > Plugins > Add marketplace, then install the same capability from it; the shell commands on this page are for Claude Code.

Local requirements: uv. The observations connector and every script here (the verification notebooks, the fixture fetch, the figure renderer) declare their dependencies in a PEP 723 header and run as uv run <script>; uv builds the environment on first run, so nothing is installed by hand. Never python script.py: it skips the header and fails at the first import. uv run <nasa-daac-knowledge>/tools/doctor.py --warm . builds every environment ahead of a first or offline run.

Downloading NASA data needs an Earthdata Login; nothing else here does, so you can install the plugin, read the knowledge bundle and search collections before you have an account. earthaccess finds the credential on its own: set EARTHDATA_TOKEN (the easy path in CI and cloud notebooks), or put a username and password in ~/.netrc (machine urs.earthdata.nasa.gov, chmod 600), or let it prompt you. Never commit either to a repository. PO.DAAC, where the ECCO and SWOT products live, needs no further authorization on the account.

Runtimes

Which runtimes this release is qualified on is the table below, rendered from the qualification records; what each word asserts is in the marketplace repository's docs/runtime-distribution.md.

Runtime support for ocean-science 0.9.0 (release lock sha256:de135f0aa0f8), rendered by build-kit's osp.py advertise from .osp/surfaces.yaml and the qualification records; edit those, not this block.

Runtime	Role	Declared status	Qualification
Claude Code	development and runtime, required	tested	Not qualified, waived for this release (Claude Code is not qualified for this release because the release environment could not produce a record that would be evidence. Two runs of the same candidate disagreed with each other: the first was contaminated by a concurrent qualification sharing the same plugin cache, and the second, run alone, had the shell tool denied part way through, so the runtime could not execute the computation it was asked to. Tool denials varied between runs, so neither result is a statement about the capability. What was established, deterministically and by hand from the installed 0.9.0 package rather than from the checkout, is that the attested sea level budget runs and attests at the new script paths: PASS, run sha256:c1031ce03a2eb00c, every check ok. This waiver first recorded one further finding as a defect in the package, that the skill does real work but its output never names the capability. That was wrong and is withdrawn. The expectation was the harness default, which asked a domain skill to print its own package name, and a skill that closes a sea level budget has no reason to write one into output a scientist reads. The same expectation was later found to match a reply from an environment where this capability was not installed at all, so it was not testing the capability. The probe now names what the skill must state to have done the work, the sanctioned computation and the attester whose verdict clears it, and it matched nineteen runs of the installed package across both invocation forms and did not match the uninstalled control. The surface is not advertised until a run passes on a machine that can run it.; human:PaulMRamirez, 2026-09-21)
Claude Cowork	runtime, required	tested	Not qualified, waived for this release (Claude Cowork is not qualified for this release. A Cowork record is a run by a person with Cowork in front of them, installing from the catalog, and no such run has been made for 0.9.0; the runtime cannot be driven headlessly, so the coordinator carrying this release on the maintainer's behalf cannot make one either. This repeats the decision recorded for land-ice 0.1.0 on 2026-09-19 for the same reason. Nothing about the capability is known to fail there: its projection renders and validates in the gate. The surface is not advertised until a run exists.; human:PaulMRamirez, 2026-09-20)
OpenAI Codex	runtime, required	planned	Not qualified, waived for this release (OpenAI Codex is not qualified for this release. No release in this organization has been qualified on Codex: the Agent Plugins projection renders and passes plugin-check in the gate, but the Codex leg has never been exercised, so there is no procedure to run and nothing to record. This repeats the decision recorded for land-ice 0.1.0 on 2026-09-19 for the same reason. The surface is not advertised, and the projection is published as conformant rather than as tested.; human:PaulMRamirez, 2026-09-20)
Claude Science	future runtime	limited-release	Outside the required matrix

A runtime is advertised as supported only on a qualified record for this exact release; a release stays valid when a runtime is not qualified, and that runtime is simply not advertised.

First result

tutorials/quickstart.md in this repository is the short form: configure the project, plan the data, load ECCO through the gate, compute the Atlantic meridional heat transport at 26.5N against the recipe, compare it with RAPID and report. It assumes core and ocean-science are installed and an Earthdata Login reachable by earthaccess, since it downloads data. The long form is Tutorial 2, ECCO Heat Transport in the tutorials repository (measured at 15.0 minutes on a fresh install including downloads, about 1.3 GB cached).

What's inside

Skills (skills/, one SKILL.md each): background on ecco, swot, ocean-grids, budget-closure, meridional-transport, water-masses, mixed-layer, sea-level and ocean-indices; the gated loaders load-ecco and load-swot; the workflows transport-analysis, ocean-budget, water-mass-analysis, mixed-layer-analysis, sea-level-analysis, compare-obs, cite-ecco, receipt-figures, briefing-generator, sweep, and the attested runs argo-ohc (this plugin's own computation) and sea-level-budget (the provider bundle's). sweep runs one attested computation once per value of a parameter its concept declares and tables the attested receipts, computing nothing of its own. Each attested computation in the provider bundle is wrapped by the workflow skill that covers it, which names the executor it runs and the attester to run on the receipt before a number is quoted. The loaders show the download size and ask before fetching; ocean-budget refuses to compute on regridded data (it never closes).
Agents (agents/): ecco-scout recommends datasets and cites the knowledge it relied on, never downloading on its own; budget-auditor checks that a computed budget closed and diagnoses it if not.
Knowledge (knowledge/): the plugin's own conventions (the climate indices, the mixed layer criteria), its three observational connector concepts and its attested computations under knowledge/computations/, each beside the skill whose scripts run it. The PO.DAAC provider bundle (dataset concepts with uncertainty structure, gotchas with evidence, recipes with expected numbers) lives at knowledge/podaac/ in nasa-daac-knowledge and is installed as the dependency; skills cite it by bundle path and nothing is copied here.
Verification (verification/): load_ecco.py, load_swot.py, transport_analysis.py, ocean_budget.py, salt_budget.py and volume_budget.py, golden notebooks that re-check each workflow on small cached data. The heat-budget closure is additionally attested against the sanctioned computation and steward-signed tolerances by the attester in this package, skills/ocean-budget/scripts/budget_residual.py. sea_level_budget.py runs this package's attested sea level budget closure on its synthetic fixture (no download, no NASA host) and proves the gap refusal; it is the golden and the prove step of the release qualification.
Evals: the agent-judgment cases (does an agent with this plugin avoid the documented ECCO and SWOT traps?) have one home, the ECCO set of agent-evals (ecco/cases/); each release is run against the tagged case set.

Configuration

Copy ocean-science.local.md.template into your project as ocean-science.local.md. Region, period, the ECCO release and data directories, the compute label, the SWOT defaults and the download gate (2 GB unless you change it) live there, and every loader reads them at the start of ocean work.

Connectors and credentials

The NASA Earthdata connector serves data discovery when it is reachable, with knowledge-based discovery as the fallback; the observations connector fetches tide gauge, Argo and PSMSL records from public agency sources. What leaves your machine, which credential is read where, and what happens when a connector is unavailable is in CONNECTORS.md.

Contributing

Start with the marketplace repository's CONTRIBUTING.md and the guides under its docs/ (contributing a skill, contributing knowledge, testing, the package authoring guide).

License and citation

Apache-2.0. Cite via CITATION.cff.