The Agent Skills Directory

[SAFE]: The skill architecture incorporates mandatory review and verification stages, ensuring that no code is shipped without passing both automated checks and human-in-the-loop gates.
[SAFE]: Stage 4 (Review) serves as a mandatory gate where all code must pass P0/P1 quality checks, preventing the progression of flawed or unauthorized changes to the production environment.
[COMMAND_EXECUTION]: The skill uses standard command-line tools like git, find, and rg for repository management and context gathering, and it constructs controlled execution commands for delegating mechanical tasks to a sandboxed environment.
[SAFE]: The execution logic enforces architect-level verification of all automated build units, explicitly stating that subagent-reported success is not trusted until it is re-verified by the main session through acceptance tests.
[SAFE]: The skill manages potential indirect prompt injection by using durable plan files as read-only decision artifacts and deriving progress from actual git state and acceptance tests rather than trusting metadata.
[SAFE]: Delegation to external runtimes like Codex is protected by a requirement for explicit user opt-in and clean-baseline preflight checks, maintaining host ownership of all sensitive operations.

sprint