RepoPilot 1.3.0
Verification-driven AI software iteration, powered by the OpenAI Codex SDK and hosted on your own worker. RepoPilot takes goals, GitHub Issues and PR feedback through planning, testing, implementation and independent verification, while maintainers retain merge and deployment decisions.
What's new
- Goals and acceptance criteria: declare the objective, allowed paths and budgets; generate dependency-ordered plans and execute resumable steps.
- Feature implementation: freeze acceptance tests before implementation. Already satisfied features can add verified tests without unnecessary production changes; bug fixes still require reproducible failure.
- Bounded iteration: verification feedback drives retries and replanning. Per-criterion attempt limits survive step renaming; repeated failed patches stop work.
- GitHub automation: opt-in Issue queues enforce configured labels and trusted authors. PR follow-up processes trusted reviewer feedback, required CI checks and compatible target-branch changes without force-pushing or merging.
- Recovery and accounting: restore verified checkpoints and completed publications, validate interrupted PR updates against full Git trees, preserve unknown model-usage reservations and account for known overages.
- Experience and proposals: reuse repository/commit/configuration-scoped history, and preview evidence-based improvement Issues before explicit publication.
- Isolated preview checks: verify candidate and baseline health before publishing. These checks do not deploy production or prove live rollback safety.
Issue reproduction, Node/Vitest/pytest/Go/JUnit evidence, repository rule checks and owned-resource recovery remain available.
Downloads
Portable packages: Linux x64/ARM64, Windows x64, macOS Intel/Apple Silicon. Each archive includes Node.js, application dependencies, examples and documentation. Extract the entire directory and verify its checksum against SHA256SUMS.txt.
Run ./repopilot on Linux/macOS or .\repopilot.cmd on Windows. Git and Docker with Linux containers are external prerequisites. Packages are CLI tools, not single-file installers, and are not code-signed/notarized.
Agent image: ghcr.io/indada/repopilot-agent:1.3.0 for Linux amd64/arm64. Model calls require OPENAI_API_KEY; GitHub credentials remain in the controller.
Upgrade
Stop the old controller and extract this version separately. Preserve configuration, reports and snapshots; update agent.image to the 1.3.0 image and pull it. Keep the configured data path unchanged. init does not overwrite existing configuration. Run --version and doctor before restarting.
Goal automation requires explicit iteration configuration. Existing check, watch and fix commands remain available. Configuration changes invalidate old goal execution; preserve its evidence and create a new goal. Legacy task reruns retain pinned commits. Version 1.0.0 locks and unlabeled Docker resources still need manual inspection.
中文说明
1.3.0 正式加入目标驱动的 AI 自动化迭代：验收标准、分步规划、功能实现、反馈重试、Issue 领取、PR 持续跟进、经验记录和改进建议。每次迭代受测试证据、修改范围与预算约束，支持暂停及恢复，最终仍由维护者决定合并和上线。
五个平台下载包与配套 Agent 镜像同步提供。升级后请按照 docs/ITERATION.md 配置新功能。