The Agent Skills Directory

Local Command Execution: The skill performs file management and state tracking using a set of whitelisted local shell commands (mkdir, cp, mv, cat, etc.) and a dedicated Python script (ledger.py). These tools are used for project organization and to enforce the mathematical proof protocol. The tool usage is constrained by the allowed-tools configuration, which limits execution to specific commands and the skill's own internal scripts.
Indirect Prompt Injection Surface: The skill is designed to process external mathematical problems and work produced by reasoning subagents. While this introduces an indirect prompt injection surface, the skill implements several mitigations: (1) Ingestion points: External data enters via problem.md and *.answer.md files. (2) Boundary markers: The skill uses explicit headers like 'Problem statement (verbatim, for reference):' to delimit user content. (3) Capability inventory: The orchestrator utilizes Bash and Agent tools to manage the workflow. (4) Sanitization: The orchestrator is instructed to treat worker content as data and not to read or summarize it directly, delegating the content evaluation to a specialized judge subagent with a 'screening' phase to detect degenerate output.

siege