Claude Project Foundation (cpf)

A Claude Code plugin for spec-driven projects: guided spec authoring (/cpf:specforge) and one set of quality checks enforced by Claude Code hooks, git hooks, and CI.

Status: maintenance

cpf is being sunset. Its successor is spec-gates, a Spec Kit extension with the same one-policy, every-boundary design.

• Existing projects can stay on the latest cpf release; only critical and security fixes land here.
• New projects should start with spec-gates.

Install

Requires Claude Code, git, and jq.

/plugin marketplace add schwichtgit/claude-project-foundation
/plugin install cpf@specforge

Update later with claude plugin update cpf@specforge, then /reload-plugins.

Use

In a project, run /cpf:specforge init once. It projects the quality gates and asks for the CI platform (GitHub, GitLab, or Jenkins). Then write the spec:

Sub-command	Produces
constitution	.specify/memory/constitution.md (principles)
spec	.specify/specs/spec.md (features, criteria)
clarify	resolved ambiguities in spec.md
plan	.specify/specs/plan.md (architecture)
features	feature_list.json (machine-readable features)
analyze	readiness score (target 80+)

Also available: setup (platform checklist), upgrade, doctor (prerequisites), and help. The spec artifacts drive autonomous execution with the bundled initializer and coder agents (.claude-plugin/agents/) or any two-agent harness.

Quality gates

Every check lives in one checks runtime that init projects into the project at .cpf/runtime/:

• verify.sh --boundary agent|git|ci runs the checks.
• commit-check.sh holds the commit and PR rules.

Three boundaries call the same code, so a change that passes at one passes at the next:

Boundary	Caller
agent	Claude Code Stop hook; PR hook on gh pr create
git	pre-commit (staged content; blocks commits to main), commit-msg
ci	GitHub ci-base.yml, GitLab gitlab-ci-base.yml, Jenkinsfile; require summary

What runs:

• Static linters: prettier, markdownlint, and shellcheck, scoped by .cpf/policy.json.
• Staged files: per-file lint (eslint, ruff, gofmt/go vet, YAML).
• Agent project checks: the policy's orchestrator: a built-in walk per language, task lint / task test, or a custom command.
• Commit rules: conventional commits; no emoji, AI-isms, or Co-Authored-By trailers.

Tools come from the project's pins (package-lock.json, .venv or uv.lock, .tool-versions). CI fails if a linter the policy needs is missing.

Other hooks block destructive shell commands and edits to sensitive files, format files on save, and point out a pending upgrade.

Project layout

Path	Owner	Purpose
.cpf/policy.json	project	per-tool include/exclude/severity; verify-quality orchestrator
.cpf/runtime/	cpf	the checks runtime
.cpf/scripts/	cpf	git hooks, install-hooks.sh, doctor.sh
.prettierignore, .markdownlint-cli2.yaml, .cpf/shellcheck-excludes.txt	generated	from the policy (markdownlint: only ignores:)
.cpf/overrides/<path>	project	replaces a plugin-provided template or prompt
.cpf/upstream-cache/, .cpf/pending/	cpf	upgrade baselines; upstream versions of files you edited

Put project-specific CI jobs in the host file (ci.yml, .gitlab-ci.yml, or the Jenkinsfile below its project marker), not in the managed base file.

Upgrade

claude plugin update cpf@specforge   # then, on a branch:
/cpf:specforge upgrade

Each project keeps running the runtime it committed until it upgrades. Upgrade replaces cpf-owned files only if they are unchanged, or match a released version. It keeps files you edited and writes the new version to .cpf/pending/<path>: merge by hand, re-run .cpf/scripts/install-hooks.sh if a hook changed, then delete .cpf/pending/. CHANGELOG.md lists behavior changes per release.

Troubleshooting

• Hooks do nothing. They need jq; run /cpf:specforge doctor.
• Stop is blocked by existing lint debt. Set "severity": "warning" on that tool's section in .cpf/policy.json while you fix it.
• A plugin update changed nothing. Expected: the project's committed runtime is used until /cpf:specforge upgrade.

Developing cpf

See CONTRIBUTING.md. In short: npm ci, then npm run lint, which runs the same runtime at the ci boundary. Also run scripts/test-*.sh (CI runs every one).

License

Copyright 2026 Frank Schwichtenberg. MIT