P
proxydis/agentic-dev
Run a batch of cards with two AI agents, a developer and a reviewer: a parallel Claude Code launcher and a…
在 GitHub 查看 ↗链接已于 2026-11-12 核查 · HTTP 200 · 打开的是厂商自己的官网
- ★ 数取值
- 1 ★ · proxydis/agentic-dev · 取自 GitHub 搜索 API · (2026-11-13 20:00:50 核对)
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
- 收录与活跃
- 收录于 2026-11-12 · 最近见到 2026-11-14 · 共见到 3 次 · 被 1 个来源列出
- 评分构成
- 综合分 103 · 1 ★ · 1 个来源 · repo
常见问题(答案都来自我们的记录)
它适合谁 / 用来做什么?
我们把它归在这些标签下: 开源项目。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库
proxydis/agentic-dev;我们收录于 2026-11-12;最近见到 2026-11-14。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。# agentic-dev
以下摘录自该仓库的 README(片段) · 在 GitHub 读完整 README →
Version française
Two tools that let two AI agents, a developer and a reviewer, build, review and ship a batch of cards, from
"start batch 1.0.0" to validation in production. The human starts the batch and only comes back for the decisions
that are theirs to make.
run-lot.sh: the development launcher. It splits the batch into non-interactive
Claude Code passes, one card per pass, on two parallel lanes. It
runs the full test suite only once, at the end of the batch, then starts the CI of the batch pull requests.
cycle-lot: the full cycle, from review to production. One state machine per batch, moved forward by the agents
through commands, and a systemd relay that wakes each agent when its turn comes and owns every wait (CI,
deployments).
Both tools were written for OpenClaw agents working in Slack, on two GitHub repositories
(back and front), with a Trello board. Everything specific to an installation lives in unversioned configuration files.
The scripts, their messages and the prompts they build are in French.
## The cycle of a batch
| # | Step | Who | What happens |
| 1 | Start | dev | cycle-lot debut: batch cards to Work in progress, release/ branches cut from develop |
| 2 | Development | dev | run-lot.sh pre, then impl: feat/… branches cut from the release, pull requests into the release |
| 3 | Arbitration | dev | The dev settles questions with the project rules (regles-arbitrage.md) and records each decision |
| 4 | Blocking question | dev → human | A question the rules cannot settle stops the cycle (cycle-lot arret) |
| 5 | End-of-batch CI | run-lot.sh | Unit and e2e tests, quality and security analysis, once, on the final heads of the PRs |
| 6 | Red CI | dev | The dev fixes it before asking for review: cycle-lot passe refuses a CI that is not green |
| 7 | Review | reviewer | Code review, tests, Playwright journeys on the deployed batch; verdict MERGE or REWORK. The first review requires the dev's real-world evidence (passe --reel). A rework review only reads the diff since the reviewed head |
| 8 | Rework | dev | Reviews 1 and 2: every point, blocking or not. From the 3rd review: blocking points only, each non-blocking point is a card of the next batch. The dev fixes the family of each defect, not the single case, and a separate pass re-reads the rework diff (passe --contre-relecture) |
| 9 | 7 ↔ 8 | | Automatic until MERGE; stops at the 5th REWORK |
| 10 | Batch merge | reviewer, relay | The MERGE verdict immediately merges each PR into the release (squash, back first); the relay adds the release commit (CHANGELOG, version) |
| 11 | Delivery | relay | Green release: the relay opens the 4 PRs, release into develop and master, back and front. The dev only steps in while a human action is still pending |
| 12 | Delivery review | relay or reviewer | Green CI and content limited to the reviewed batch: the relay merges the 4 PRs. Otherwise the reviewer checks and merges |
| 13 | Production | relay, then dev | CI of develop and master (deployment included), then the deployment settings declared during the batch, set and proven by the dev, and the list of visual checks to run in production |
Two tools that let two AI agents, a developer and a reviewer, build, review and ship a batch of cards, from
"start batch 1.0.0" to validation in production. The human starts the batch and only comes back for the decisions
that are theirs to make.
run-lot.sh: the development launcher. It splits the batch into non-interactive
Claude Code passes, one card per pass, on two parallel lanes. It
runs the full test suite only once, at the end of the batch, then starts the CI of the batch pull requests.
cycle-lot: the full cycle, from review to production. One state machine per batch, moved forward by the agents
through commands, and a systemd relay that wakes each agent when its turn comes and owns every wait (CI,
deployments).
Both tools were written for OpenClaw agents working in Slack, on two GitHub repositories
(back and front), with a Trello board. Everything specific to an installation lives in unversioned configuration files.
The scripts, their messages and the prompts they build are in French.
## The cycle of a batch
| # | Step | Who | What happens |
| 1 | Start | dev | cycle-lot debut: batch cards to Work in progress, release/ branches cut from develop |
| 2 | Development | dev | run-lot.sh pre, then impl: feat/… branches cut from the release, pull requests into the release |
| 3 | Arbitration | dev | The dev settles questions with the project rules (regles-arbitrage.md) and records each decision |
| 4 | Blocking question | dev → human | A question the rules cannot settle stops the cycle (cycle-lot arret) |
| 5 | End-of-batch CI | run-lot.sh | Unit and e2e tests, quality and security analysis, once, on the final heads of the PRs |
| 6 | Red CI | dev | The dev fixes it before asking for review: cycle-lot passe refuses a CI that is not green |
| 7 | Review | reviewer | Code review, tests, Playwright journeys on the deployed batch; verdict MERGE or REWORK. The first review requires the dev's real-world evidence (passe --reel). A rework review only reads the diff since the reviewed head |
| 8 | Rework | dev | Reviews 1 and 2: every point, blocking or not. From the 3rd review: blocking points only, each non-blocking point is a card of the next batch. The dev fixes the family of each defect, not the single case, and a separate pass re-reads the rework diff (passe --contre-relecture) |
| 9 | 7 ↔ 8 | | Automatic until MERGE; stops at the 5th REWORK |
| 10 | Batch merge | reviewer, relay | The MERGE verdict immediately merges each PR into the release (squash, back first); the relay adds the release commit (CHANGELOG, version) |
| 11 | Delivery | relay | Green release: the relay opens the 4 PRs, release into develop and master, back and front. The dev only steps in while a human action is still pending |
| 12 | Delivery review | relay or reviewer | Green CI and content limited to the reviewed batch: the relay merges the 4 PRs. Otherwise the reviewer checks and merges |
| 13 | Production | relay, then dev | CI of develop and master (deployment included), then the deployment settings declared during the batch, set and proven by the dev, and the list of visual checks to run in production |
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[](https://anclum.com/zh/s/proxydis-agentic-dev/)
- 来源
- github · https://github.com/proxydis/agentic-dev
- 按领域浏览
- 开源项目
- Traction
- 1 ★
- 收录于
- 今天
同类推荐
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。