The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute repository-defined test commands (e.g., npm test, pytest, make test) and perform Git management tasks. A critical primitive used is git add -A followed by git reset --hard, which is explicitly documented to permanently delete untracked files without recovery if the environment is not clean before execution.\n- [INDIRECT_PROMPT_INJECTION]: The workflow requires the agent to read and interpret the output from external test runners to determine the success or failure of a phase. This ingestion of untrusted data from an external tool execution creates a potential surface for indirect prompt injection.\n
Ingestion points: Standard output and error streams from the <test_command> execution in SKILL.md.\n
Boundary markers: Absent; the agent is instructed to inspect output for exit codes and counts without specific delimiters or isolation.\n
Capability inventory: The agent has the capability to write and modify code, delete files via Git reset, and execute arbitrary repository scripts.\n
Sanitization: Absent; the skill does not define filtering or validation logic for the tool output before it is processed by the agent logic.\n- [DYNAMIC_EXECUTION]: The core TCRDD loop involves the agent dynamically generating implementation code during the GREEN phase and immediately executing it through the project's testing infrastructure to verify behavior.

tcrdd