The Agent Skills Directory

[COMMAND_EXECUTION]: Several Python management scripts (scripts/start_workflow.py, scripts/phase_entry.py, scripts/init_workspace.py) utilize subprocess.run to execute system commands.
scripts/init_workspace.py executes git commands to retrieve repository metadata and branch names for automated report naming.
scripts/phase_entry.py and scripts/start_workflow.py execute bensz_skill_kernel.cli via sys.executable to manage state transitions, verifier gates, and workspace initialization.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process potentially untrusted external content.
Ingestion points: The skill reads user-provided project backgrounds, experimental data, paper drafts, and external literature retrieved from URLs or via the research-literature-search dependency.
Boundary markers: Instructions delegate much of the processing to sub-agents (research-literature-interpretation). While they mention desensitizing summaries, there is a lack of explicit, consistent XML/delimitated boundary markers in the primary instructions for raw data interpolation.
Capability inventory: The skill possesses file system write capabilities (for reports and workspace state) and executes shell commands via management scripts.
Sanitization: scripts/init_workspace.py implements a sanitize function using unicodedata.normalize("NFKC") and regex filters to prevent directory traversal or command injection via repository/PR names used in output paths.

research-idea