EngineeringTeam is a workflow layer for AI coding agents. Its safety posture is deliberately conservative.

• No unpinned remote execution. The package does not download or execute remote code. All scripts are local and dependency-free.
• No session-start shell behavior. Installing the package does not register any hook that runs shell commands when a session starts. The workflow is invoked manually with engineering-team.
• No hidden persistence. The package does not write outside the repository or a user-chosen install target. Install scripts never overwrite existing files unless you pass --force.
• No automatic production mutation. The workflow never deploys, mutates production systems, or runs destructive commands on its own.
• Human approval for destructive or production-sensitive actions. The skill requires explicit human approval before live-system mutation, irreversible data/file changes, broad generated-file rewrites, or release/rollback decisions.
• Least-privilege tool guidance. Specialist agents declare read-only vs edit boundaries. Investigation and review agents are read-only; only the implementation, test, and documentation roles may edit, and only after the evidence gate.
• Plugin install transparency. Installation copies clearly named files into visible locations (.codex/agents/, .github/agents/) or prints exact manual steps for marketplace-based harnesses.

• scripts/validate-codex-package.py, skills/engineering-team/scripts/validate-package.py — read and validate manifests.
• scripts/generate-agents.py — regenerates agent files from agents-src/.
• scripts/doctor.py — runs the read-only validators and reports health.
• scripts/install.py, skills/engineering-team/scripts/install-custom-agents.py — copy agent files into a chosen target; idempotent unless --force.

Open a private security advisory on the repository, or email the maintainer listed in .claude-plugin/plugin.json. Please do not file public issues for sensitive reports. Include reproduction steps and the affected files.