The Agent Skills Directory

[COMMAND_EXECUTION]: The autonomous execution loop outlined in rules/autonomous-loop.md directs the agent to run shell commands for syntax validation (e.g., python -m py_compile) and execute core functions without user review.
[INDIRECT_PROMPT_INJECTION]: The skill accepts and processes potentially untrusted task descriptions as input for its autonomous development cycle. While it employs an assumption-tagging mechanism ([假设]), it lacks robust sanitization for interpolated data.
Ingestion points: SKILL.md (Phase 1: Autonomous Understanding of user input).
Boundary markers: The skill uses assumption tags but does not provide explicit 'ignore embedded instructions' delimiters for the processed task content.
Capability inventory: Full file system write access, shell command execution for validation, and the ability to define/install dependencies.
Sanitization: Absent; the skill relies on internal budget checks rather than input validation.
[DYNAMIC_EXECUTION]: The skill's primary function is the continuous generation and execution of code across multiple iterations. This automated 'generate-test-fix' cycle operates without human-in-the-loop checkpoints.

power-iterate