The Agent Skills Directory

[COMMAND_EXECUTION]: The harness utilizes node:child_process and node-pty to automate interactions with the agent's CLI and TUI. It facilitates complex testing scenarios by spawning multiple child processes (via tsx and node), managing tmux sessions, and executing headless Google Chrome instances to render visual evidence of terminal states.- [CREDENTIALS_UNSAFE]: The skill includes logic to monitor the agent's primary authentication file (~/.senpi/agent/auth.json) to ensure its integrity is maintained during tests. Specialized scripts are also provided to verify the agent's OAuth account management by seeding sandbox environments with temporary tokens, ensuring the auth system functions correctly under test conditions.- [EXTERNAL_DOWNLOADS]: The project declares a dependency on node-pty, which requires local compilation of native components during installation. It also utilizes well-known developer tools such as tsx for running TypeScript source files and requires a browser to be present for screenshot generation.- [INDIRECT_PROMPT_INJECTION]: The harness allows the agent to process arbitrary prompts and tool arguments to verify its logic, creating a surface for potential indirect prompt injection during QA cycles. Ingestion points occur through the mock-loop.mjs command-line arguments. The agent under test has access to shell and filesystem capabilities, which the harness restricts to an isolated sandbox directory. Sanitization is not performed by the harness, as its purpose is to evaluate the agent's raw performance in a controlled environment.

senpi-qa