8000 GitHub - gominimal/minimal-skills: Claude Code skills for the Minimal developer system, with the eval pipeline that keeps them honest · GitHub
[go: up one dir, main page]

Skip to content

About

Claude Code skills for the Minimal developer system, with the eval pipeline that keeps them honest

Resources

Stars

6 stars

Watchers

0 watching

Forks

Latest commit

 

History

217 Commits

Folders and files

Repository files navigation

Minimal logo

minimal-skills

Claude Code skills for the Minimal developer system, packaged as a plugin, plus the eval pipeline that keeps them honest. Each skill teaches Claude Code how to drive one slice of Minimal (setup, configuration, networking, diagnostics), and every skill is continuously tested for trigger accuracy, answer quality, and drift against the published docs.

Install

claude plugin marketplace add gominimal/minimal-skills
claude plugin install minimal@gominimal

Or run /plugin inside the Claude Code REPL and install minimal from the gominimal marketplace.

Skills

Skill What it does
minimal-setup Get a repo running in the Minimal developer system: install, min init, session creation and activation.
minimal-sandbox Work correctly from inside a sandbox (session or task sandbox): the in-sandbox min helper, installing and persisting packages, task limits, what survives where.
minimal-config Author and validate minimal.toml: packages, stacks, upstream pinning, mip check/min check.
minimal-loadouts Personalize sessions with per-developer loadouts: editors, dotfiles, env vars, patches, composition rules.
minimal-networking Session networking: --network modes, egress rules, .min.internal names, publishing and forwarding ports, min finalize-install.
minimal-diag Diagnose broken sessions and builds: min bug support bundles, uploading one for diagnosis with min bug --upload or min diag upload, deleting it with min diag delete.

Eval pipeline

Two tiers, two canaries. The binding contract for cases, checks, and the runner CLI is evals/SCHEMA.md; the methodology follows Testing Agent Skills.

  • PR text tier: every pull request runs a no-secrets URL lint, then the text-tier suite (does the right skill trigger, does no skill trigger on negative cases, does the response pass its checks). One trial per case; regression cases gate the merge.
  • Nightly functional tier: the full suite (3 trials) on runners with a real Minimal installed from both the stable and nightly channels, so a Minimal release that breaks a skill surfaces within a day.
  • Obsolescence canary (nightly, advisory): the text tier re-run with the skills not installed. When the bare model passes everywhere, a skill has stopped paying rent and should be slimmed or retired.
  • Docs-drift canary (nightly): the published minimal.dev pages the skills link are hashed and compared to evals/docs-snapshots.json; a changed page fails the run so the skill gets re-checked against it.

Failures of blocking nightly jobs open or update a single tracking issue.

Dev loop

Needs the claude CLI on PATH, authenticated (locally via claude login; CI passes a CLAUDE_CODE_OAUTH_TOKEN secret generated with claude setup-token) (npm install -g @anthropic-ai/claude-code); lint-urls and lint-parity need neither.

just lint-urls        # liveness-check every docs URL in the skills
just lint-parity      # assert the minimal README's three-command block matches minimal-setup
just eval             # text tier, all skills, 1 trial
just eval-skill minimal-setup   # text tier, one skill
just triggers         # fast trigger-accuracy pass
just functional       # functional tier; needs a local Minimal install
just obsolescence     # text tier with the skills not installed
just report           # open the latest JSON report

Repo layout

.claude-plugin/       plugin and marketplace manifests
skills/               one directory per skill, each with a SKILL.md
sandbox/              hook assets for the in-sandbox build (gominimal/pkgs)
evals/                eval harness: runner.py, checks.py, cases/, SCHEMA.md
.github/workflows/    pr-evals.yml (PR gate), nightly-evals.yml (nightly tier)
justfile              local dev-loop recipes

Notes

  • Skills link the published docs at minimal.dev instead of restating them; the URL contract lives in evals/SCHEMA.md.
  • Behavior claims are checked against a real install, and anything the reference documents is linked rather than restated. The 0.7.0 pass was verified on 2026-10-14 against min 0.7.0-rc2 (macOS arm64): network modes, egress defaults, .min.internal names, min net forward, static and dynamic ingress, min bug --upload and min diag delete, and the min session run environment trap. The 0.6.0 pass was verified on 2026-09-25 against min 0.6.0 staged from minimal 9b763dd5 (macOS arm64). The 0.4.0 pass was verified on 2026-09-03 against min 0.5.4-dev.62.g30a3031d (macOS arm64, minvmd + gvproxy); reference gaps found on that pass are filed as #622, #623 and #624 in gominimal/inbox.
  • minimal-networking is link-first since 0.7.0, when the cli-min and tasks references gained the networking surface. Where the running 0.7.0-rc2 and the published docs part ways, the skill records the running behaviour inline. Each gap is filed in gominimal/inbox:
    • --dynamic-ingress allow works on a macOS host, though cli-min says a VM-backed host refuses it: #1093.
    • The sandbox-operations page does not list min net expose: #1094.
    • An own_ip session on macOS does not reach host.min.internal or the host-side proxy, though the release notes say it does: #1095.
  • Sandboxed Claude Code sessions get this content through the claude-code-minimal-plugin package in gominimal/pkgs, which fetches a tagged snapshot of this repo and assembles skills/minimal-sandbox/ plus the sandbox/ hooks into an in-sandbox plugin. A marketplace install on a developer host ships the skills only, never the hooks.

License

MIT, see LICENSE.

Note: the optional LLM style judge (--judge) uses the Anthropic SDK and needs ANTHROPIC_API_KEY; the default pipeline authenticates the claude CLI alone and never requires it.

About

Claude Code skills for the Minimal developer system, with the eval pipeline that keeps them honest

Resources

Stars

6 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages

0