The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands, specifically codex exec and claude -p, to implement its 'Tier 2' capability for stateless persona simulation. This allows the agent to run sub-processes to generate diverse feedback from different model instances or configurations when available.
[SAFE]: The skill adopts a 'files-first' architecture, where all sprint state, decisions, and teammate notebooks are stored in a user-accessible sprints/ directory. This ensures that the agent's work is transparent, persistent across sessions, and avoids hidden state within the AI's internal context.
[SAFE]: Includes a 'Teammate Contract' that is explicitly restated in every sub-agent prompt. This contract forbids persona agents from calling messaging tools, spawning further agents, or writing to files unless specifically instructed for a task, which acts as a robust safeguard against unauthorized tool use or recursive agent loops.
[SAFE]: The design ensures the 'Decider' is always the user. Persona agents are instructed to provide friction and diverse viewpoints rather than converging on autonomous decisions, maintaining human-in-the-loop control for all critical project milestones.

sprint-runner