OSS Contribution Harness
yurekami/oss-contribution-harness
Pipeline anti-regressão para PRs de IA em OSS com 12 gates de credibilidade
Ver no GitHubSkill para Claude Code que implementa um pipeline seguro de contribuições open source, com 12 regras que aumentam a credibilidade de PRs gerados por IA. Automatiza deduplicação, reprodução, atestação e follow-up para que contribuições tenham chances reais de merge instead de serem descartadas como noise.
Instalação
npx convez-ai yurekami/oss-contribution-harness1 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
Casos de uso
- Automatizar contribuições de qualidade para repos open source
- Validar que PRs geradas por IA entendem o sistema antes de submeter
- Implementar phase gates de reprodução, deduplicação e atestação em fluxo de contribuição
- Rastrear segurança, follow-ups e throttling de cadência de PRs
Veredito

Artefato técnico sólido que resolve um problema real: a perda de credibilidade de contribuições IA em OSS. As 12 regras são bem fundamentadas e o pipeline é rigoroso, mas requer integração profunda com fluxos existentes e expertise em CLI do GitHub.
README
A tradução para o português deste README ainda está na fila.
oss-contribution-harness
A regression-safe pipeline for getting AI-drafted pull requests merged upstream instead of filed as noise. Implements the credibility signals that maintainers look for now that AI agents have changed the contributor pipeline.
"Talk and code is cheap. Show me you really care." — Roger Wang, MLSys 2026
Implemented as a Claude Code skill, but the workflow is tool-agnostic — the gates are the point.
Why this exists
AI coding agents changed open source contribution. Weekly PR volume to major AI infrastructure repos spiked visibly around agent releases in 2025–2026. Generating plausible code is no longer the bottleneck — the bottleneck is reviewer trust. Maintainers now need to evaluate whether the contributor understands the system, whether the change solves the right problem, and whether the person will stay involved after the PR is opened.
LLM-driven contribution at scale fails for predictable reasons:
- PRs that look like fixes but don't reproduce on the target.
- PRs that follow
CONTRIBUTING.mdbut ignore the revealed norm in merge history. - PRs that "fix" load-bearing intentional behavior.
- Drafts in generic LLM-PR voice that maintainers spot instantly.
- Duplicates of open or recently-closed work.
- Bots that pierce CLA / merge / approval clicks they have no authority to make.
- Security-class findings pushed to public PR diffs instead of disclosure channels.
- PRs orphaned at "open" — never followed up after CI / reviews / bots.
- Velocity that platforms read as abuse regardless of per-item quality.
- Hardcoded branches that can't factor per-repo context the design stage missed.
This harness encodes ten rules as hard phase gates with explicit judgment
points, not pre-declared logic. Each rule corresponds to a phase in the pipeline
and writes its evidence to per-candidate state under ~/.omc/ghcontrib/.
The 12 rules
| # | Rule | Where it lives in the harness |
|---|---|---|
| 1 | Reproduction is the merge gate. No repro → candidate dies. | Phase 5 — gate on repro_verdict. |
| 2 | Imitate merge history, not CONTRIBUTING.md. | Phase 3 — fingerprint last 10 merged PRs. |
| 3 | Respect load-bearing behavior. Some "bugs" are intentional. | Phase 2 — blame + linked issues + design discussion. |
| 4 | Few-shot with your own past PRs. | Phase 4 — pull gh search prs --author $ME --merged. |
| 5 | Dedupe before drafting. Open + closed issues + PRs. | Phase 1 — gh search then read every result. |
| 6 | Wrap attestation; never pierce it. | Phase 9 — explicit y/N from a human; never auto-y. |
| 7 | Route security-sensitive findings off the public PR path. | Phase 6 — heuristic classifier → security-queue/. |
| 8 | Plan the follow-up in the pipeline. | Phase 11 — T+24h / T+72h / T+7d schedule. |
| 9 | Throttle on calendar time, not just per-item quality. | Phase 8 — append-only ledger, ≤5 PRs/7d/identity. |
| 10 | Let the agent branch at decision points. | Recorded in state.json.decisions[] with rationale. |
| 11 | Demonstrate system understanding. Plausible code is not a credibility signal. | Phase 2 — understanding.md proves subsystem comprehension. |
| 12 | Make design decisions reviewable, not just the diff. | Phase 7 — "Design context" section in the PR body. |
Pipeline
scout → candidate → dedupe → archaeology → reveal → fewshot → reproduce
→ draft → security-gate → attest → open → followup → close|merged|dead
Forward-only. Each phase writes evidence to
~/.omc/ghcontrib/candidates/<owner>__<repo>__<id>/.
Install
Drop SKILL.md into your Claude Code skills directory:
mkdir -p ~/.claude/skills/omc-learned/ghcontrib
curl -sLo ~/.claude/skills/omc-learned/ghcontrib/SKILL.md \
https://raw.githubusercontent.com/yurekami/oss-contribution-harness/main/SKILL.md
Prerequisites:
gh(authed:gh auth login)gitjq(Windows:winget install jqlang.jq; macOS:brew install jq)
State directories are created on first run.
Invocation
/ghcontrib <owner/repo> --scout
/ghcontrib <owner/repo> --candidate <issue-url>
/ghcontrib --resume <id>
/ghcontrib --followup # sweep all open PRs and compute next actions
Real example
examples/deepseek-aider/ contains the full
state of one live run that produced
deepseek-ai/awesome-deepseek-integration#582
— Aider added to the Applications list, +5 LOC matching the median of recent
merges, drafted in the user's voice, with the closed precedent issue
referenced and the maintainer's explicit invitation cited.
state.json shows every gate decision and rationale; draft.md is the PR
body that went up; reveal-summary.json is the merge-history fingerprint
that drove sizing and tone.
Attestation
This harness never clicks CLA, signs DCO, approves a review, merges a PR,
or pushes as anyone other than the identity gh is authed as. Every
network-visible action requires a fresh in-session human y. If the source
project requires a CLA, the human signs it in their browser. The bot
automates up to the click, never through it.
Throttle
Default budget: ≤5 PRs / 7 days / identity, ≤2 PRs / repo / 14 days.
Edit the constants in Phase 8 of SKILL.md to suit your tolerance for
platform abuse heuristics. Velocity is a semantic signal regardless of
per-item quality.
Security
If a candidate touches auth, crypto, sandbox, deserialization, SQL, shell
exec, path traversal, SSRF, TLS — or if the diff would itself reveal a
vulnerability — the harness redirects to ~/.omc/ghcontrib/security-queue/
and surfaces the project's SECURITY.md disclosure channel. It does not
push.
Attribution
Rules 11–12 and the credibility principle were informed by Roger Wang's talk "Rethinking Open Source Contribution in the Age of AI Agents" at MLSys 2026. Roger is a core maintainer of vLLM and co-founder of Inferact. The talk drew on direct experience maintaining AI infrastructure as agent-generated PRs became a significant share of contributor activity.
License
MIT — see LICENSE.