The Agent Skills Directory

Command Execution: The skill utilizes command-line tools such as svn, gh (GitHub CLI), and atr (ASF Tooling Release) to retrieve metadata and prepare release actions. A key security feature is that every state-changing action, including posting comments and starting the vote process, is presented as a proposal requiring explicit Release Manager confirmation before execution.
Indirect Prompt Injection Surface: The skill ingests data from external planning issues and changelogs (defined in SKILL.md). It contains specific guardrails instructing the agent to treat all external text as untrusted data rather than executable instructions, mitigating the risk of malicious prompts embedded in release metadata.
Ingestion points: Planning issue bodies and external changelogs.
Boundary markers: Explicit instructions to treat external content as data and refer to safety guidelines in the repository's AGENTS.md file.
Capability inventory: Execution of svn, gh, and atr commands.
Sanitization: Logical separation of data from instructions and human-in-the-loop review of final command outputs.
Dynamic Execution: The skill performs a pre-flight check by executing a local Python module (setup_preflight) located in the .apache-magpie-local directory. This is used to verify the project's configuration state before performing drafting tasks.

vote-draft