convez.ai
AgenteAutomaçãoAvançado
gregce

DeadReckon

gregce/deadreckon

Supervisor confiável para agentes de código com sandbox isolado e assinatura criptográfica

Ver no GitHub
BomPluginScore · de 0 a 100

Harness de supervisão para agentes de IA que executa código de forma isolada e auditável. Implementa um watchdog separado que valida o trabalho do agente, garante durabilidade de estado entre turnos e gera artefatos assinados em vez de transcrições. Compatível com Claude, OpenAI, Gemini e outros provedores.

Instalação

$npx convez-ai gregce/deadreckon

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
15
Forks
5
Acompanhando
0
Saúde
71%
Último commit
há 3 meses
Versão
v0.8.8
Issues abertas
1
Contribuidores
2

Casos de uso

  • Executar agentes de IA sem supervisão humana contínua
  • Auditar e rastrear mudanças de código geradas por IA
  • Automatizar tarefas de desenvolvimento com garantia de integridade

Veredito

Ferramenta robusta para quem precisa delegar trabalho de código a agentes IA com garantias de segurança e auditoria. Abordagem diferenciada com watchdog separado e assinatura criptográfica reduz riscos de manipulação, mas exige fluxo de trabalho e conceitos mais complexos.

README

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


Get started

Install the latest release (macOS / Linux):

curl -fsSL https://deadreckon.sh/install.sh | sh

This resolves the newest release (stable preferred, release candidates included), verifies it against SHA256SUMS, and installs deadreckon. Pin any release tag with DEADRECKON_TAG=<tag>.

Or build from source:

cargo build --release          # binary at ./target/release/deadreckon

Then the whole tool is five commands:

CommandWhat it does
deadreckon start "build the app"Kick off a supervised run. Walk away.
deadreckon attach latestWatch it work live. Ctrl-D leaves it running.
deadreckon statusWhat happened, and the one thing to do next.
deadreckon listFind recent runs and plans.
deadreckon finish latestApply it to your branch, or export it.

For a higher-level read while it runs, deadreckon attach latest --view narrative shows cited prose plus an evidence-backed visual map.

Everything else (budgets, undo, multi-step chains, provider routing) is optional power you reach for later. No API keys? deadreckon run "hello" --smoke --sandbox none exercises the whole harness against a faked provider.

[!TIP] DeadReckon is on the production-release track. The core lifecycle (isolated runs, signed gates, durable state, undo, docs, orchestration, and apply) is implemented and tested; release notes call out any compatibility migrations.

Why it's different

  • It can't fake the finish. A separate watchdog (dr-gate) holds a secret the agent process can't read and signs the result with it. No valid signature, no "done." The agent literally cannot mark its own work accepted.
  • Walk away for real. Every turn is saved to disk. Close your laptop, lose the network, kill the model, then attach from another terminal and resume from the last completed turn.
  • Evidence, not a transcript. Each accepted run promotes to a reviewable artifact: what changed, why, which prompt touched which file, what it spent.

→ The full story (the loop, the mental model, and how it compares) is in Concepts & How It Works.

Documentation

DocWhat's inside
Concepts & How It WorksThe loop, the watchdog, the mental model, supported agents, how it compares
HOWTOPractical usage: setup, providers, sandboxes, configuration, every command
DESIGNProduct and architecture intent
As-Built ArchitectureDetailed implementation reference
Resume Semantics · Multi-RunCrash/resume behavior and concurrency rules
V1 CandidatesDeferred features and roadmap

DeadReckon is maintained as production-release software. The core lifecycle is implemented and tested, alongside multi-agent orchestration (plan / fork / merge), autonomous chains, the provider flight recorder with rewind, and a local self-improvement loop (learn / improve).

Ver o original no GitHub

Bom para

Relacionados