Package the repo as an Agent Plugins plugin for ChatGPT and Codex

[Marigold]

ChatGPT and Codex now share one Plugin Directory, and it reads the vendor-neutral Agent Plugins format: a root plugin.json, with skills discovered from skills/. That is already this repo's layout, so a single manifest makes the same commit installable in both ecosystems. This adds that manifest plus a repo-scoped catalog, so codex plugin marketplace add owid/skills resolves this repo as a plugin source. No skill content changes.

Skim — the whole diff is manifests and docs. The one thing worth a considered opinion is the note under "Decisions".

How it works

plugin.json carries portable identity (name, description, author, license) at the root and OpenAI's install-surface copy under extensions.com.openai.interface — display name, categories, starter prompts. Skills need no registration there: the format discovers everything under skills/, and a skills declaration is explicitly ignored for packages that have a root manifest. So adding a skill still means editing exactly one list, the Claude one.

.agents/plugins/marketplace.json is the catalog entry pointing at the repo root (source.path: "./"). Without it the repo is a plugin but not a source, and codex plugin marketplace add owid/skills has nothing to resolve.

Two deliberate omissions:

• No version field. The schema makes it optional and hosts fall back to 1.0.0, so the repo's versionless convention survives intact rather than acquiring a semver counter nobody would remember to bump.
• No public directory submission. That path needs identity verification, five positive and three negative test cases, and per-version human review — which collides head-on with shipping on every commit to main. This PR only enables repo and local installs, which have no gatekeeper.

make validate gains a fourth check: plugin.json against the published schema. Worth its own check because the schema sets additionalProperties: false, so a typo'd key doesn't degrade — it makes the whole manifest unreadable — and neither skills-ref nor claude plugin validate looks at this file.

Verified

Installed from this branch with the Codex CLI and confirmed all four skills land in the plugin cache:

codex plugin marketplace add owid/skills --ref research-distribution-channels
codex plugin add owid@owid-skills
ls ~/.codex/plugins/cache/owid-skills/owid/*/skills/
# fetch-chart-data  joining-data  owid-catalog  search-charts

Also verified the local-path route (a personal ~/.agents/plugins/marketplace.json pointing at a worktree) resolves — its paths turn out to be relative to $HOME, which the docs don't state.

Not verified: the ChatGPT desktop app's own install UI, and category: "Productivity" / capabilities: ["Read"] — those are the only values the docs exemplify, and I found no published enumeration to check them against. Both are cosmetic if wrong.

Decisions worth a second look

Installing this way copies the whole repo into the user's plugin cache, evals/ included — the plugin root is the repo root, same as the existing Claude marketplace entry, so this is not a regression. Eval prose still cannot reach an agent's context (skills are read only from skills/), but it does land on disk. If we'd rather it didn't, the fix is a plugins/owid/ subdirectory with the skills inside, which changes both manifests and every install path in the README. I'd leave it, but it's a real call.

README: one install route per client

A follow-up commit reorganises the install section around the reader rather than the packaging format. It now has a routing table and a section each for Claude Code, Codex, the ChatGPT app and the Claude app.

Two gaps that reorganising exposed, both now covered: someone using the ChatGPT app previously had to infer that a Codex CLI command was the step they needed, and had no pointer to developer mode, the Work switcher, or the workspace-marketplace route that lets an admin reach colleagues without it. And the Claude app had no route at all — its apps don't read plugin marketplaces, they take one skill at a time as a zip. Verified that zip -r x.zip x from inside skills/ produces the layout its docs require, with SKILL.md one level down.

The Claude app runs skills in a sandbox rather than on the reader's machine, so its prerequisites differ from the table above it. That is flagged as unverified rather than asserted.

Context for whoever picks this up

• The format is governed by a TSC spanning Amazon, Cursor, Microsoft, OpenAI and Vercel, so plugin.json is a bet on more than ChatGPT.
• OpenAI docs: package your plugin, submission.
• ChatGPT desktop needs Settings → Security and login → Developer mode before local sources appear, and a restart after a source is added or its files change.
• If a marketplace entry's source can't be resolved, Codex skips that entry rather than failing the whole marketplace — so a broken entry is silent, not loud.

🤖 Generated with Claude Code