The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/passes/execution_feasibility.py uses importlib.util.spec_from_file_location and spec.loader.exec_module(module) to dynamically load and execute Python modules from the target repository being analyzed. While it attempts to exclude common entry point scripts (e.g., train.py, main.py), importing any file from an untrusted repository can lead to arbitrary code execution if that file contains top-level logic.
[DYNAMIC_EXECUTION]: The script scripts/write_outputs.py dynamically loads a shared module write_explore_bundle.py from a relative path (../../../shared/scripts/). This relies on a specific external directory structure and represents dynamic loading from a computed path.
[COMMAND_EXECUTION]: scripts/orchestrate_explore.py and scripts/passes/execution_feasibility.py utilize subprocess.run to execute other scripts (such as scan_repo.py and analyze_project.py) and system utilities including git, nvidia-smi, and rocm-smi for environment management and hardware telemetry.
[EXTERNAL_DOWNLOADS]: The skill fetches metadata and documentation from well-known services including arXiv (export.arxiv.org), DOI (doi.org), and GitHub (api.github.com). These network operations are core to the skill's research purpose and target official API endpoints.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources (paper summaries, GitHub READMEs, and target repository code) and uses this data to synthesize research ideas and plans. These synthesized plans are subsequently used by the agent to guide experimental steps.
Ingestion points: Paper summaries from ArXiv (arxiv_provider.py), README content from GitHub (github_provider.py), and file content extracted from the local repository (repo_extractors.py).
Boundary markers: None identified; untrusted text is interpolated into idea cards and campaign reports.
Capability inventory: Subprocess execution for orchestration, dynamic code import for repo probing, and file writing for artifact generation.
Sanitization: External descriptions are used as strings without specific sanitization against instruction overrides.
[PERSISTENCE]: scripts/orchestrate_explore.py uses git worktree add to create persistent experiment workspaces on the local filesystem outside the primary repository directory.

ai-research-explore