The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands defined in campaign configuration files and researcher inputs to run research evaluations and detect system resources.
Evidence: scripts/orchestrate_explore.py executes evaluation_source["command"] using subprocess.run to anchor exploratory trials.
Evidence: scripts/passes/execution_feasibility.py executes system utility commands like nvidia-smi and rocm-smi to detect GPU hardware availability.
[REMOTE_CODE_EXECUTION]: The skill performs dynamic loading and execution of Python code from the target research repository.
Evidence: scripts/passes/execution_feasibility.py uses importlib.util.spec_from_file_location and spec.loader.exec_module to run modules from the user's repository during import_probe_check and constructor_probe_check passes.
Evidence: scripts/write_outputs.py dynamically loads a shared module from a computed path relative to the skill installation directory.
[EXTERNAL_DOWNLOADS]: The skill fetches metadata and content from external research services and repositories.
Evidence: scripts/lookup/providers/arxiv_provider.py fetches paper metadata from export.arxiv.org.
Evidence: scripts/lookup/providers/doi_provider.py resolves digital object identifiers via doi.org.
Evidence: scripts/lookup/providers/github_provider.py fetches repository metadata and README files from the GitHub API (api.github.com).
[DATA_EXFILTRATION]: The skill collects and logs detailed hardware and system environment information.
Evidence: scripts/passes/execution_feasibility.py gathers OS release details, CPU core counts, memory statistics, and GPU specifications, writing them to a RESOURCE_PLAN.md artifact.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection from external research metadata.
Evidence: Research metadata and README content fetched from external sources are incorporated into the agent's reasoning loop for ranking and grounding ideas without sanitization.
Ingestion points: scripts/lookup/providers/github_provider.py (README content), scripts/lookup/providers/arxiv_provider.py (paper summaries).
Boundary markers: Absent; external content is processed alongside researcher ideas without explicit delimiters.
Capability inventory: Subprocess execution, file system writes (explore_outputs/), and dynamic Python module execution.
Sanitization: Absent; the skill decodes Base64 data from the GitHub API but does not filter for malicious instructions.

ai-research-explore