The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands defined in the 'done-criteria' during the verification phase (Phase 4). While this allows for code execution (e.g., running pytest or gh commands), it is the core functionality intended for automated verification and is constrained by the agent's permission model.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it processes user-supplied objectives and criteria which guide subsequent agent actions and command execution.
Ingestion points: Objective statement and criteria fields provided during the Phase 1 specification.
Boundary markers: The skill utilizes a structured state file (.objective/<slug>/state.md) to organize its working memory, providing structural separation between objective data and execution logic.
Capability inventory: The skill leverages Bash, ScheduleWakeup, and the /do router to perform work and manage the execution loop.
Sanitization: Includes a mandatory 'Criteria-gaming guard' that prevents the loop from satisfying criteria by weakening tests or security controls, and 'NOT-DONE-YET' guardrails to enforce constraints across all iterations.
[DATA_EXFILTRATION]: The skill reads from .claude/settings.json to retrieve the orchestration.token_budget. This is used exclusively for internal resource management (Phase 1) and does not involve transmitting sensitive data to external domains.

objective-loop