The Agent Skills Directory

[COMMAND_EXECUTION]: The toolset uses the subprocess module to execute Git commands and user-defined test commands. Analysis shows that these calls use the safe list-based format (shell=False), which prevents shell injection vulnerabilities. Execution is bounded by timeouts and logs are stored in project-specific directories.
[DATA_EXPOSURE]: The pipeline_tools/core.py file implements a redact function that uses regular expressions to identify and strip common secrets (passwords, tokens, API keys, Bearer tokens) and local user paths from command outputs before they are written to logs or metrics.
[SAFE]: The skill includes a scope_check mechanism that validates changed files against a frozen allow-list in the task contract. It explicitly ignores untracked files that are not permitted and can detect if forbidden files (like .env or secret files) have been modified, returning a drift exit code to block the workflow.
[SAFE]: The workflow defines clear boundaries between agent roles (Main, Executor, Reviewer) and mandates that all findings be backed by machine-readable evidence blocks (JSON inside Markdown). This ensures that agent claims can be mechanically verified against actual command outputs and Git state.
[INDIRECT_PROMPT_INJECTION]: While the skill processes task sheets and reports that could be modified by agents, it utilizes strict JSON schema validation and structural checks (as seen in contract.py and core.py) to ensure only intended fields are processed. The system-level instructions in SKILL.md further enforce role boundaries and verification requirements, mitigating the impact of potential instructions embedded in processed data.

programing-pipeline