The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is granted Bash(*) tool access and utilizes codex exec to perform autonomous tasks, including running research experiments, managing remote sessions (screen/tmux), and modifying repository files.
[EXTERNAL_DOWNLOADS]: Fetches citation and metadata from dblp.org and doi.org. These are well-known research information services used for validating references and represent a low risk.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the local repository (such as code, experiment logs, and result files) and uses this data to drive its autonomous actions, which include file modifications and command execution.
Ingestion points: Uses Read, Grep, and Glob tools to ingest repository content, and codex exec allows a remote model to read the entire repository autonomously.
Boundary markers: Utilizes Markdown headers (## Round N), horizontal dividers (---), and HTML <details> tags to structure and separate round-specific content in the AUTO_REVIEW.md cumulative log.
Capability inventory: Possesses full Bash(*) access, Write, Edit, Task tools, and the ability to trigger codex exec for autonomous repository exploration.
Sanitization: Includes a specific instruction to avoid splicing repository paths into shell source code, but lacks comprehensive sanitization for other data ingested from the repository before it influences agent actions.
[DYNAMIC_EXECUTION]: Resolves and executes local Python helper scripts (e.g., review_gate.py and copilot_native_evidence.py) from computed paths within the project directory to manage loop transitions and verify reviewer evidence.

auto-review-loop