The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external referee correspondence and internal feedback (PDF or text formats), as described in Phase 1 and 2 of the protocol. This untrusted data could contain instructions meant to influence the agent's behavior during the analysis or planning phases.
Ingestion points: Phased reading of external PDFs (e.g., {venue}-round{n}-reviews.pdf) and text-based feedback reports in references/phases.md.
Boundary markers: The skill employs a 'Verbatim Gate' in Phase 4 requiring exact transcription and provides LaTeX escaping rules (e.g., & to \&) to prevent command injection into document artifacts.
Capability inventory: The agent can write various Markdown and JSON files, execute a Python script (dag_validator.py) via uv, and compile documents using latexmk.
Sanitization: Standard LaTeX escaping is performed during Phase 4, but there is no specific sanitization for natural language instructions embedded within the feedback.
[COMMAND_EXECUTION]: The skill instructs the agent to perform several shell operations to manage the revision workspace. These include creating directories (mkdir), copying files (cp), and compiling LaTeX documentation (latexmk). Additionally, it executes a provided Python script (dag_validator.py) to validate the revision tasks.
[EXTERNAL_DOWNLOADS]: The skill utilizes the uv tool to execute the dag_validator.py script with the networkx library specified as an ephemeral dependency (uv run --with networkx). This results in the download of the networkx package from the Python Package Index (PyPI). networkx is a well-known and standard library for graph analysis.

strategic-revision