The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script (scripts/ledger.py) to handle the mechanical aspects of the mathematical ledger and protocol state. It also utilizes standard shell utilities such as mkdir, cp, cat, and grep for file management within a designated run directory. These operations are restricted to the local environment and are essential for the protocol's functionality.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external input in the form of mathematical problem statements and answers generated by worker sub-agents. These are passed to judge sub-agents for evaluation.
Ingestion points: The orchestrator reads the problem statement from user arguments or a file and collects worker answers from DIR/*.answer.md files.
Boundary markers: The protocol employs explicit, query-specific end-of-answer markers (=== END OF ANSWER {Q} ===) to delimit worker output, and the bookkeeping script strictly verifies these markers before a result is accepted.
Capability inventory: The orchestrator can write files, execute a local Python script, and spawn sub-agents. It does not perform network operations.
Sanitization: The skill contains explicit instructions for the orchestrator to avoid interpreting or paraphrasing the mathematics itself, delegating all judgment to fresh sub-agents and mechanical checks to the Python script, which reduces the risk of the orchestrator being manipulated by the content it processes.
[EXTERNAL_DOWNLOADS]: No remote scripts or external dependencies are downloaded during the execution of this skill. The bookkeeping script relies solely on the Python standard library.
[DATA_EXFILTRATION]: No unauthorized network operations or exfiltration patterns were detected. Data remains within the user-specified run directory and the platform's standard agent communication channels.

siege