Security and agents
Keep security reporting private and give coding agents durable operating limits.
Security controls must remain understandable without a specific coding agent. The repository is the operational record.
Report vulnerabilities privately
SECURITY.md names the supported versions and the current private reporting
method. Do not route vulnerabilities through a public issue. Give the reporter
the response expectations and the information needed for a useful report.
Give agents exact local instructions
AGENTS.md records:
- source and package ownership,
- architectural boundaries,
- public contracts and invariants,
- exact verification commands,
- release restrictions,
- prohibited actions, and
- active migrations when applicable.
Do not tell agents to use codex/*, claude/*, or another tool-specific branch
name. Use the repository's normal branch and pull request policy.
CLAUDE.md contains a short instruction to read AGENTS.md. It does not fork
the policy.
Keep the skill thin
The Lupinum OSS Codex skill reads this handbook and the target repository. It can create, audit, and prepare work. It must not contain a second copy of the fleet policy.
The skill must not silently publish a package, bypass quarantine, add a token, rename a trusted workflow, or claim that an external setting is correct without checking it.
Preserve evidence
Keep release manifests, hashes, CI links, and external-setting audit results in their defined locations. Do not put credentials, authentication URLs, or local tokens into logs or repository files.