convez.ai
Skill
jeremylongshore

contributing-clanker

jeremylongshore/contributing-clanker

Ver no GitHub
BomPluginScore · de 0 a 100

OSS contribution workspace + the /contribute Claude Code skill — keeps AI-assisted upstream contributions honest before they reach maintainers.

Instalação

$npx convez-ai jeremylongshore/contributing-clanker

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
100%
Último commit
há 3 dias
Versão
v0.3.0
Issues abertas
1
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

contributing-clanker v0.2.0

Make AI-assisted open-source contributions land cleanly — caught by deterministic gates before they reach maintainers as slop.

A local-only Claude Code skill plus workspace for contributing to open-source projects you don't own. It runs 51 deterministic safety gates over every claim comment, design issue, and pull request, so AI-generated work never reaches an upstream maintainer as low-quality "slop." State is plain markdown — greppable, git-trackable, no database, no cloud calls.

License: MIT Release

Links: GitHub · Gist One-Pager

What Is This?

contributing-clanker is a workspace repo plus a vendored Claude Code skill (/contribute). The skill orchestrates discovery, per-repo research, drafting, and testing for upstream OSS contributions; the workspace holds the skill's source, its spec, its tests, and the upstream clones you contribute to. Every externally-visible action (claim, design issue, PR) is routed through a gate-runner that blocks anything failing one of 51 deterministic checks — with the reason written to an append-only log.

The runtime is deliberately boring: Bash + gh + jq, with markdown and JSONL as the only persistence. No server, no database, no cloud — the whole system is inspectable by hand and recoverable from git.

Capabilities

CapabilityDescription
Discovery@scout finds + ranks upstream issues by star-tier; --repos for surgical targeting
Research@researcher builds per-repo dossiers (CLA / DCO / AI-policy / commit format / templates)
Gating51 deterministic gates (phases A–G) block AI-slop before it ships; verdicts logged
Trust laddera contributor's (N+1)th PR scope is governed by N prior merges at that repo (gates A07 + B13)
Lifecycletransition.sh walks each candidate open → shortlist → claimed → working → submitted → merged
Runtime mirrorinstall.sh deploys scripts/ → ~/.contribute-system/bin/; doctor.sh verifies no drift

Key Principles

  • Deterministic over vibes — logged shell verdicts (PASS/WARN/BLOCK), not an LLM saying "looks fine"
  • Markdown-only state — greppable, git-trackable, survives any tool; no DB, no daemon, no cloud
  • Enforcement travels with the code — gates + per-repo dossiers live in the repo / runtime state, never hard-coded
  • Human approval before any external submission — the skill stops and asks; it never auto-posts upstream

Scope

SupportedNot Yet
Filesystem-only (Phase 1), single-userMCP service (Phase 3)
51 gates across phases A–Gthe remaining planned catalog gates (62-mode catalog)
Local Claude Code skillmarketplace plugin packaging (Phase 2, epic 25c)
gh-driven live GitHub statecross-machine / multi-user coordination

Quickstart

Prerequisites

  • Claude Code
  • gh CLI, authenticated (gh auth status)
  • jq on PATH

Install

git clone https://github.com/jeremylongshore/contributing-clanker
cd contributing-clanker
bin/install.sh              # production install (copy)
bin/install.sh --symlink   # dev install (edits land live in the repo)

Both modes also deploy the runtime mirror to ~/.contribute-system/bin/. After install, /contribute is available in Claude Code.

Run

/contribute            # in any Claude Code session — surfaces in-flight PRs/issues + candidates
/contribute <owner>/<repo>   # onboard a new upstream repo (builds dossier, briefs you)

Run Tests

bats tests/unit/gates/            # unit tests (48 cases across the gates)
skills/contribute/scripts/test-known-traps.sh        # regression suite (5 total: test-*.sh)
skills/contribute/scripts/doctor.sh                  # verify the deployed runtime mirror matches the repo
skills/contribute/scripts/lint-bash.sh               # shellcheck the gate scripts

Configuration

The skill is configured by markdown state, not env vars. The two env vars below exist only to retarget the deploy for tests.

Variable / fileDefaultPurpose
~/.contribute-system/profile.md(created on first run)scout preferred languages / star tiers + own-org exclusions
CONTRIBUTE_BIN_DIR~/.contribute-system/binruntime-mirror deploy target (overridable for tests)
CONTRIBUTE_RESEARCH_DIR~/.contribute-system/researchper-repo dossier directory (overridable for tests)

Architecture

Three layers, gates in the middle:

/contribute (SKILL.md) ──spawns──→ subagents: scout · researcher · draft-writer · test-runner
        │
        │ every external action
        ▼
   transition.sh ──→ gate-runner.sh ──→ 51 gates (phases A–G)   ──BLOCK/WARN/PASS (logged)
        │
        ▼
   ~/.contribute-system/  (markdown state: candidates · research/dossiers · log.jsonl · profile.md)
        │  gh CLI (live, never cached)        only after gates PASS + human approval
        ▼
   upstream GitHub repo
  • Layer 1 — Per-repo dossiers (~/.contribute-system/research/<owner>__<repo>.md): what a specific upstream expects (branch convention, CLA/DCO, AI policy, PR/issue templates, draft-first, review bots, etiquette). Built by @researcher; cached + refreshable (smart-refresh preserves engineer-curated sections).
  • Layer 2 — Deterministic gates (scripts/gates/, deployed to ~/.contribute-system/bin/gates/): one small script per failure mode → PASS / WARN / BLOCK / INFORM + a one-line reason. Pluggable: drop a script in, the runner discovers it.
  • Layer 3 — Lifecycle workflow (/contribute): walks each candidate through the states above, running the right gate set per transition. BLOCK refuses the transition; WARN surfaces in the briefing.

Project Structure

contributing-clanker/
├── skills/contribute/        # the skill — single source of truth
│   ├── SKILL.md              # /contribute orchestrator
│   ├── agents/               # 5 subagents
│   ├── scripts/              # 17 top-level + gates/ (51) + gates/lib/
│   ├── references/           # candidate / dossier / workflow specs
│   └── assets/               # 3 generic templates (claim / pr / evidence)
├── bin/install.sh            # deploys skill + runtime mirror (with doctor smoke-check)
├── 000-docs/                 # spec (10 epics) + release AARs
├── tests/                    # bats unit + L4 regression
├── features/                 # Gherkin BDD acceptance
├── <upstream-repo>/          # one subdirectory per upstream clone (own CLAUDE.md each)
├── AGENTS.md                 # non-interactive shell rules
└── CLAUDE.md                 # project conventions for Claude Code
~/.contribute-system/         # runtime state (candidates, dossiers, log) — NOT in this repo

Each upstream clone has its own CLAUDE.md with stack-specific commands — read it before working there.

Documentation

  • Spec + index: 000-docs/000-INDEX.md — 10 epics → spec docs + release AARs
  • Changelog: CHANGELOG.md (Keep a Changelog + SemVer)
  • One-pager + operator audit: the Gist
  • Shell rules: AGENTS.md — always cp -f / rm -f / mv -f / apt-get -y (interactive prompts hang the agent)
  • History: the pre-2026-04-30 SQLite-tracker + monorepo system was deprecated; planning docs in 99-archived-system-docs/, code in git history.

License

MIT — see LICENSE

Ver o original no GitHub

Bom para

Relacionados