The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to autonomously execute shell commands to verify progress during its '自主迭代' (Autonomous Iteration) phase. It specifically instructs the agent to '用 fresh 命令或证据验证' (verify with fresh commands or evidence) and log the output as verification evidence in iteration reports.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests high-level goals and acceptance criteria from external inputs to drive its autonomous code-writing and command-execution loops. Malicious instructions could be embedded in these goals to manipulate the agent's behavior during the implementation phase.
Ingestion points: Processes 'owner' input (dialogue) and files within the project's .codestable/ directory (e.g., goal.md, state.yaml).
Boundary markers: Utilizes structured YAML and Markdown templates to separate state from instructions, though it lacks explicit guidance to ignore embedded instructions within processed goal content.
Capability inventory: High; includes broad file system access for code editing, version control operations (commit/merge), and arbitrary command execution for verification as defined in SKILL.md and reference.md.
Sanitization: No specific sanitization or validation of the input goals or criteria is documented before the agent acts on them.

cs-goal