convez.ai
AgenteAutomaçãoAvançado
romeo111

TaskTorrent

romeo111/task_torrent

Coordenação GitHub para IA contribuir em open-source com qualidade controlada

Ver no GitHub
InicialPluginScore · de 0 a 100

Protocolo nativo do GitHub para coordenar contribuições de agentes IA em projetos open-source com revisão humana. Quebra trabalhos complexos em chunks pequenos e bem-definidos, permitindo que ferramentas de IA (Claude, ChatGPT, Cursor) completem tarefas discretas que são depois validadas por mantenedores antes da fusão.

Instalação

$npx convez-ai romeo111/task_torrent

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
0
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 6 meses
Versão
v0.4.0
Issues abertas
9
Contribuidores
1

Casos de uso

  • Curação de dados biomédicos e pesquisa clínica com revisão de qualidade
  • Decomposição de projetos grandes em tasks atômicas para agentes IA
  • Fluxos de contribuição open-source mantendo controle editorial humano

Veredito

Solução sofisticada e bem-pensada para colaboração IA-humano em projetos críticos (como pesquisa do câncer), com enfoque em rastreabilidade, segurança e revisão editorial. Requer setup técnico e governança clara, mas oferece caminho viável para escalar trabalho de IA em open-source responsavelmente.

README

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

TaskTorrent

TaskTorrent is a GitHub-native protocol for turning AI-agent time into reviewable open-source pull requests.

The first pilot is cancer-research data work for OpenOnco, a free oncology clinical decision-support knowledge base. TaskTorrent breaks that work into Chunks: small, scoped tasks with a manifest, allowed sources, output format, safety rules, and acceptance criteria. Contributors use their own AI tools to complete one chunk at a time. Maintainers review every output before it is merged.

TaskTorrent is not a marketplace, token system, payment system, hosted bot, or autonomous publishing pipeline. It is a coordination layer for human-reviewed work.

Current OpenOnco Metrics

Live counts and the full claimable-chunks list are on the TaskTorrent landing page, regenerated every 4 hours from registered consumer repos. Headline metrics:

  • Completed chunks (cumulative): see landing
  • Sidecar submissions counted (lifetime): see landing
  • Claimable chunks right now: see landing
  • Estimated work surface: ~16M tokens

Active chunks may live in either romeo111/task_torrent or the consumer repo (romeo111/OpenOnco), depending on the wave. The landing page merges both. To filter directly on GitHub: chunk-task + status-active in task_torrent or in OpenOnco.

Live public metrics are published through the landing data files:

If claimable chunks are 0, OpenOnco is between active waves. You can still read the shelf, prepare your environment, or ask the maintainer to open the next chunk.

Contribute To Cancer Research

Claude Code (recommended)

Install the openonco-contributor plugin:

/plugin marketplace add https://github.com/romeo111/task_torrent
/plugin install openonco-contributor@tasktorrent

Then run /openonco-contribute from a local checkout of romeo111/OpenOnco. The plugin picks one currently claimable chunk, summarizes scope, drafts sidecars under contributions/<chunk-id>/, runs validation, and stops after one PR.

Plugin source and docs: plugins/openonco-contributor/.

Other coding agents (Codex / Cursor / ChatGPT / etc.)

Use this prompt from a local checkout of romeo111/OpenOnco:

You are helping me contribute one TaskTorrent chunk to OpenOnco cancer-research
data work.

Read these first:
- https://github.com/romeo111/task_torrent
- https://github.com/romeo111/task_torrent/blob/main/docs/openonco-pilot-workflow.md
- https://github.com/romeo111/task_torrent/tree/main/chunks/openonco
- https://github.com/romeo111/task_torrent/tree/main/plugins/openonco-contributor/skills
- https://github.com/romeo111/OpenOnco

Pick exactly one currently claimable OpenOnco chunk. If no chunk is claimable,
stop and report that the shelf is between waves.

Before writing files, summarize:
1. the chunk ID and GitHub issue,
2. the manifest you will own,
3. allowed sources and banned sources,
4. output path under contributions/<chunk-id>/,
5. validation commands you will run.

Rules:
- Work only under contributions/<chunk-id>/ unless the chunk explicitly says
  otherwise.
- Do not write treatment advice or patient-specific recommendations.
- Do not use OncoKB, SNOMED CT, MedDRA, or any banned source in the chunk.
- Do not use git add -A, git add ., or --no-verify.
- Open one pull request for one chunk, then stop.

The safest first contributions are citation verification, source-stub cleanup, translation review, structured evidence mapping, and schema-safe sidecar data work. Clinical content still requires OpenOnco maintainer and clinical lead review.

OpenOnco Claude Plugin Plans

TaskTorrent is planning two separate Claude plugin tracks for OpenOnco:

  • OpenOnco TaskTorrent Contributor: a public-first plugin for volunteers who want to donate AI-agent time to citation verification, source-stub cleanup, biomarker evidence sidecars, and chunk readiness review.
  • OpenOnco Clinical Evidence Review Support: a later, stricter plugin for clinicians and maintainers reviewing evidence provenance, citation support, and sidecar submissions. This track needs a stronger privacy posture and a hard no-PHI boundary before any directory submission.

The contributor plugin is the recommended first submission because it is sidecar-only, review-gated, and not patient-specific. The clinical-review plugin should remain separate so it can carry tighter privacy, support, refusal, and human-review requirements.

Planning docs:

How TaskTorrent Works

  1. A project publishes chunk specs under chunks/<project>/.
  2. Maintainers open active chunks as GitHub issues.
  3. A contributor or agent claims one chunk.
  4. Work lands in an isolated sidecar path in the consumer repo.
  5. CI and maintainers verify the output.
  6. The maintainer merges, rejects, or asks for revision.

The rule is: one chunk = one contributor = one PR = one reviewable unit of work.

What Is A Chunk?

A Chunk is a complete LLM-assisted task. A good chunk has:

  • a stable chunk ID
  • a concrete manifest of entities or files
  • allowed and disallowed sources
  • an output schema or file format
  • machine-checkable acceptance criteria
  • maintainer-checked semantic criteria
  • explicit rejection criteria
  • a claim method
  • a Drop estimate

Chunks are grouped by folders and GitHub labels, not by heavyweight project management layers.

What Is A Drop?

A Drop is an approximate planning measure for AI work capacity. In this repo, 1 Drop means about 100k tokens of structured AI-assisted work.

A Drop is not money, crypto, credit, a reward, or a transferable asset. It is only a sizing estimate.

Maintainer Quickstart

You do not need to install anything to start planning. Give your coding agent this prompt inside your project:

You are helping me adopt TaskTorrent for this repository.

Read https://github.com/romeo111/task_torrent and use its README,
docs/cross-repo-contract.md, docs/chunk-system.md, and chunks/openonco/
examples as the protocol reference.

Do not write production code yet. First, propose a TaskTorrent integration plan:

1. Identify 5-10 useful chunks that are safe for AI-assisted contributors.
2. For each chunk, define mission, manifest, allowed sources, disallowed
   sources, output format, acceptance criteria, rejection criteria, required
   skill, claim method, and Drop estimate.
3. Mark safety boundaries and maintainer-review requirements.
4. Recommend which chunks should open first as GitHub issues.
5. Produce the first chunk spec in TaskTorrent v0.4 format.

Keep every chunk reviewable as one issue, one contributor, one branch, and one
pull request.

Optional local tooling:

pip install -e ".[dev]"
python -m tasktorrent.lint_chunk_spec --all chunks/

Commands:

  • tasktorrent-lint <chunk-spec.md> validates a chunk spec.
  • tasktorrent-init <consumer-name> scaffolds a basic consumer integration.

OpenOnco References

Repository Boundary

This repository contains the TaskTorrent protocol, docs, OpenOnco chunk specs, skill specs, GitHub templates, landing data, and optional linter/scaffold tooling.

It does not contain OpenOnco production data, a hosted service, autonomous merge logic, or a replacement for maintainer review.

Ver o original no GitHub

Bom para

Relacionados