The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes common local command-line tools including git for branch and tag operations, jq for parsing JSON output, and openssl for generating unique run identifiers.
[COMMAND_EXECUTION]: It manages project tasks through the beads (bd) CLI, which acts as a local graph database for task tracking and state management within the repository.
[PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection by reading user-specified files and goal descriptions to generate plans. 1. Ingestion points: User goal arguments and any files referenced within those arguments. 2. Boundary markers: The generated plan uses specific Markdown headers and a sentinel string (<promise>ALLDONE</promise>) to delimit the end of the run. 3. Capability inventory: Access to git, beads, bash, and jq for local repository manipulation. 4. Sanitization: The skill does not explicitly sanitize the user-provided goals, but it enforces a strict RED/GREEN/VERIFY structure for task descriptions.
[SAFE]: No network exfiltration, remote code downloads, or access to sensitive credential files (e.g., .ssh, .aws) were detected. All generated artifacts are stored in local directories like .llmtmp/ or .beads/.

ralph-builder