The Agent Skills Directory

[COMMAND_EXECUTION]: The orchestration engine (harness-engine.sh) and its helper scripts execute various shell commands including git, gh (GitHub CLI), find, and date. These are used for managing task state, discovering project conventions, and automating Git workflows such as committing code and creating pull requests.
[COMMAND_EXECUTION]: The claude-agent-invoke.sh script invokes the claude CLI using the --dangerously-skip-permissions flag. This allows sub-agents in the pipeline to execute their designated tools without interactive confirmation, which is a documented requirement for the skill's autonomous execution model.
[DYNAMIC_EXECUTION]: Several scripts, including harness-engine.sh and claude-agent-invoke.sh, use inline Python one-liners (python3 -c) and heredocs for structured data processing (JSON parsing and Markdown extraction). These are localized utilities that do not execute unsanitized external code.
[INDIRECT_PROMPT_INJECTION]: The skill manages a multi-agent pipeline where one agent's output serves as input for another. To mitigate the risk of instruction leakage or context poisoning, the framework enforces an 'anti-drift' protocol, spawning fresh sub-agent sessions for each checkpoint to maintain clean execution boundaries.

harness