Reusable baseline files for repositories that should be easy for Codex agents to inspect, modify, build, review, and release.
The kit has two modes:
-`agent-quickstart.md`: shortest prompts and decision tree for Codex agents.
-`new-repository.md`: agent workflow for a fresh repository.
-`existing-project.md`: agent workflow for upgrading an existing repository.
The files are intentionally universal. They do not assume Electron, React, Node, or any single stack. Node examples are included only because many Codex projects use them.
## Agent Goal
Give every repository the same predictable anchor points:
```text
.
|-- AGENTS.md
|-- .codex/
| `-- project.md
|-- .gitea/
| `-- workflows/
| `-- build.yml
|-- docs/
| |-- release-checklist.md
| `-- security-review.md
|-- blueprint.md
|-- blueprint.json
|-- README.md
|-- SECURITY.md
|-- CHANGELOG.md
`-- .gitignore
```
Use only the files that fit the project. For a tiny script repo, `AGENTS.md`, `README.md`, `SECURITY.md`, and `CHANGELOG.md` may be enough. For an app or releasable tool, add the runner, release checklist, and README blueprint workflow.
When working with private repositories on `git.wilkensxl.de`, Codex agents may find a local `GITEA_TOKEN` environment variable on the machine.
Use `GITEA_TOKEN` only for read-oriented Gitea API checks unless the user explicitly asks for a write action. Typical safe checks include repository metadata, workflow run status, and package-read visibility. Never print the token, commit it, or copy it into workflow files.
Example status endpoint:
```text
GET https://git.wilkensxl.de/api/v1/repos/REPOSITORY_OWNER/REPOSITORY_NAME/actions/runs
Authorization: token GITEA_TOKEN
```
`REGISTRY_TOKEN` is still the intended secret name for CI package publishing inside `.gitea/workflows/build.yml`.
After pushing commits that trigger a Gitea workflow, Codex agents must keep checking the resulting workflow run before finishing. Poll the run until it reaches a terminal state. If it succeeds, report the successful run. If it fails or is cancelled, inspect the failing job/logs, fix the issue when it is in scope, commit, push, and repeat the check loop for the next run. A fixed failure is not a stopping point; the loop continues after the follow-up push until a workflow succeeds or a concrete out-of-scope blocker is reached. Do not stop after a single in-progress status when the user asked the agent to continue the loop.
