ShipSafe: How a LinkedIn Post Sent Me Back Into My Old Field for 10 Hours

Описание и применение

ShipSafe – автоматический агент, проверяющий pull‑request, запускающий тесты в изолированном контейнере и требующий ручного одобрения перед публикацией комментария.

Задача

Pull request reviews are tedious and error‑prone; reviewers often skim diffs and assume tests pass, allowing bugs to slip through.

Результат

ShipSafe completed a PR review in under 10 hours, ran tests in a sandbox, and required explicit human approval before posting, catching a real bug that Qodo also identified.

Рабочий процесс

• Register for Agent Harness Hackathon
• Create empty repo and ship-safe.json agent definition
• Use TrueForge harness to call GitHub MCP for PR data
• Spin up Daytona sandbox to run tests (npm test or test.py)
• If multiple files changed, spawn subagents per file
• Aggregate subagent verdicts into a review summary
• Show review to user and ask for approval to post comment
• Persist session state across page refreshes
• Merge PR after user approves review

Доказательства

ShipSafe talks to GitHub through an MCP server, not a hand‑rolled API client.

Источник

ShipSafe runs tests in a Daytona sandbox, never on my host machine.

Источник