The Agent Skills Directory

[COMMAND_EXECUTION]: The dispatch.py script executes shell commands through the run_gate function using subprocess.run(shell=True). These commands, typically project build or test suites like node --test, are provided via command-line arguments when the AI orchestrator initiates the verification phase.
[DATA_EXFILTRATION]: The dispatch.py script reads the contents of source code files (supporting extensions such as .js, .py, .go, .ts) within the workspace. This is used in the fix function to provide context to the LLM for diagnosing and repairing test failures.
[PROMPT_INJECTION]: The skill processes external files (spec.md and agenda.md) as primary input for its planning phase. These files are interpolated into prompts for the Sonnet model, creating an indirect prompt injection surface where malicious instructions in the specifications could influence the generated build plan or the resulting code.
[EXTERNAL_DOWNLOADS]: The skill relies on the claude CLI tool to perform model calls. The Python script executes this local command to communicate with Anthropic's models using the host's existing authentication, avoiding the use of hardcoded API keys.

thrifty-dispatch