The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted LaTeX proof files provided by the user, which presents a surface for indirect prompt injection where instructions could be hidden in comments or complex math structures.
Ingestion points: The skill reads .tex files and reference materials from the user's project directory.
Boundary markers: The skill uses delimiters such as [FULL PROOF CONTENT HERE] and [FIXED SECTION ONLY] in its prompts to the reviewer models, which provides basic separation but may not prevent sophisticated injection.
Capability inventory: The agent has extensive capabilities including Bash for command execution, Edit/Write for file modification, Agent for spawning sub-agents, and several MCP tools for external model interaction.
Sanitization: There is no evidence of sanitization or filtering of the LaTeX source code before it is passed to the models for reasoning.
[COMMAND_EXECUTION]: The skill executes shell commands via the Bash tool to compile LaTeX documents and run a local Python script for wiki management.
Evidence: The skill calls pdflatex -interaction=nonstopmode <file>.tex in Phase 2e and python3 "$WIKI_SCRIPT" add_claim ... in Phase 5.5.

proof-checker