The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses local Python scripts to manage the code review workflow and interact with the git repository.
scripts/build_manifest.py, scripts/run_jobs.py, and scripts/_common.py execute shell commands such as git and gh (GitHub CLI) using the subprocess.run() function.
These command executions are implemented using list-based arguments and shell=False, which prevents command injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: As a code review tool, the skill naturally ingests untrusted data from repository files and pull request diffs.
The system mitigates the risk of adversarial instructions within the reviewed code by employing sub-agents with clear operational boundaries.
Ingestion points: Repository source code and pull request hunks are read by sub-agents as part of their assigned review cohorts.
Boundary markers: Prompt templates in assets/PROMPT.md use distinct headers (e.g., FILES, CONTEXT, REPO RULES) to isolate untrusted code content from agent instructions.
Capability inventory: Sub-agents are limited to read-only access of source files and are restricted to writing findings to specific JSON output files managed by the orchestrator.
Sanitization: The orchestrator scripts strictly validate sub-agent outputs against a JSON schema (assets/findings.schema.json), preventing the injection of malicious payloads or unintended instructions into the final report.
[EXTERNAL_DOWNLOADS]: The skill retrieves necessary metadata from GitHub to support pull request reviews.
scripts/build_manifest.py utilizes the official GitHub CLI tool to fetch pull request titles, bodies, and URLs.
These network operations are restricted to the official GitHub API and represent standard, non-malicious functionality for a development tool.

deep-review