The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands via the Bash tool and specific internal commands like /run-experiment. This is part of its core functionality for deploying and monitoring research experiments on local and remote servers.
[DYNAMIC_EXECUTION]: Stage 2 (Implementation) and Stage 4 (Auto Review Loop) involve the autonomous generation and modification of code by the agent (e.g., 'extend pilot code to full scale', 'Claude Code implements fixes'). This code is then directly executed as part of the research workflow.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources which then influences code generation and execution.
Ingestion points: Data is pulled from the arXiv API, WebSearch results, and WebFetch content during the '/research-lit' phase (SKILL.md).
Boundary markers: No explicit boundary markers or 'ignore embedded instructions' warnings are defined for the interpolation of this external data into the implementation logic.
Capability inventory: The skill has access to Bash, Write, Edit, Agent, and Skill tools, allowing it to modify the filesystem and execute arbitrary code (SKILL.md).
Sanitization: There is no evidence of sanitization or validation of the fetched research content before it is used to generate experiment scripts.
[AUTONOMY]: The 'AUTO_PROCEED' constant (set to true by default) allows the agent to auto-select research ideas and move to the implementation/execution phases after a 10-second timeout. This reduces human oversight for what could be expensive (GPU compute) or security-sensitive (code execution) operations.

research-pipeline