[alukach]

AI agents can now learn when to reach for source-coop, not only how. A new Agent Skill, skills/source-coop/SKILL.md, triggers when you mention Source Cooperative, source.coop or data.source.coop, and sends the agent to source-coop --help for everything else. The repository is also a Claude Code plugin marketplace, so it installs in two lines:

# in Claude Code
/plugin marketplace add source-cooperative/source-coop-cli
/plugin install source-coop@source-cooperative

# any other agent that reads skills: copy the folder where it looks for them
cp -r skills/source-coop ~/.your-agent/skills/

How it behaves

The skill is 25 lines and copies no flags, per AGENTS.md ("A skill … should say when to use the CLI and point to --help"). It tells the agent to:

• run source-coop --help, then <command> --help, since help matches the installed version; if the CLI is missing, point the user to https://cli.source.coop/latest/install/;
• start with source-coop auth status --json; if not signed in, ask the user to run ! source-coop auth login (it opens a browser, so the agent never runs it), or use SOURCE_TOKEN;
• pass --json, parse stdout, read errors as JSON on stderr; exit 3 means fix field_errors and retry, 4 not signed in or not allowed, 5 not found;
• send bodies with -F -, never use -i, ask before deleting, revoking or leaving, then pass --yes;
• use source-coop api for anything without a command;
• reach data at s3://ACCOUNT/PRODUCT/ on https://data.source.coop (public data needs no sign-in), with each tool's settings left to https://docs.source.coop/data-proxy rather than copied into the skill, and look in source-coop --help for credentials (creds, and exec where installed, so it reads right before and after feat: exec, to run S3 tools against the data proxy, signed in or not #43);
• prefer the Portolan CLI, if installed, for publishing geospatial datasets with metadata.

The marketplace is named source-cooperative and lists one plugin, source-coop, whose root is the repository root ("source": "./"). plugin.json has no version, so installs follow the latest commit on main rather than being pinned.

Docs changes

• Scripts and AI agents gains an Agent skill subsection: what the skill does and how to install it. No ## heading changed, so the contents list and site pages are the same. Preview: https://cli.source.coop/pr-preview/pr-98/scripts-and-ai-agents/#agent-skill
• AGENTS.md: under Help is the agents' documentation, the skill must stay in step with FOR_SCRIPTS and the CLI's behaviour, changed in the same commit, and claude plugin validate . checks the plugin.

Not in this PR

• A plugin version. claude plugin validate warns that it's missing. Without one, users get each new commit on main; with one, someone (or release-please) must bump it. Worth adding if the skill ever needs to match released CLI versions.
• An icon, which only matters for a listing in Anthropic's plugin directory.
• llms.txt on the docs site. Could follow the same rule: say when to use the CLI and point to --help.

Testing

• claude plugin validate . (Claude Code 2.1.319): passes, with warnings for the missing version, the repository's CLAUDE.md at the plugin root (not loaded by plugins; harmless), and no icon.
• Installed from a throwaway CLAUDE_CONFIG_DIR, both from the local checkout and from GitHub at this branch (/plugin marketplace add source-cooperative/source-coop-cli#feat/agent-skill): claude plugin details source-coop lists Skills (1) source-coop.
• cargo test and cargo test --features staging: 179 passed (no Rust changes, no tests added).
• Docs site built locally; the skill link points at GitHub.
• Not tried: an agent session actually triggering the skill and following it against a real or staging API. No demos, since no command changed.

🤖 Generated with Claude Code

[claude]

Claude finished @alukach's task in 12s —— View job

---

✅ No blocking issues — safe to merge.

I read the full diff (plugin manifests, SKILL.md, README and AGENTS.md edits). I did not run the plugin validator or an agent session.

• The license in plugin.json (Apache-2.0) matches LICENSE and Cargo.toml.
• The skill says to run auth login only through the user, which fits the README's note that login opens a browser. The --yes rule, -F - rule and exit codes 3/4/5 match the AGENTS.md table.
• The skill copies no flags, as AGENTS.md requires. The README examples put comments above commands, and the new ### heading adds no site page.

Notes (non-blocking)

• SKILL.md says "exec where installed". The README doesn't document exec yet, so the wording is right only while it is unreleased. Update the line once feat: exec, to run S3 tools against the data proxy, signed in or not #43 ships.
• plugin.json has no version, so every commit on main is a new install. You already listed this under "Not in this PR".

Simplify (ponytail)

• .claude-plugin/plugin.json:6-11: homepage, keywords, and the duplicate description and author are optional. Cut them, since the marketplace entry already carries a description, and add them back if the plugin goes into a directory listing.
• .claude-plugin/marketplace.json:3: the top-level description and owner.url are optional metadata. Keep only name, owner.name and plugins if you want it minimal.
  · branch feat/agent-skill

---

💰 Estimated review cost: $0.12 · 0m12s · 5 turns

[github-actions]

Docs preview: https://cli.source.coop/pr-preview/pr-98/

Built from 9f4262a. It goes live a minute or two after this run, once GitHub Pages deploys it, and is removed when the PR closes.