The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted local data from manuscripts and state files to orchestrate agent workflows, creating a surface for indirect instructions to influence agent behavior.
Ingestion points: In phase 01, the skill reads paper.json, citations.json, rebuttals.json, and manuscript text files (.tex, etc.) to build a snapshot for planning.
Boundary markers: No explicit delimiters or boundary markers are documented to isolate untrusted content interpolated into engine task briefs.
Capability inventory: Dispatches sub-engines (which may have network or file access), performs git version control operations, and writes to project log files.
Sanitization: The skill does not document sanitization or escaping of ingested data before it is used to generate topics or task briefs for sub-engines.
[COMMAND_EXECUTION]: The skill executes local version control commands to persist project state between iterations.
Evidence: In phase 05, the skill performs git add -A and git commit operations. There is a potential command injection risk if ingested data (such as a topic derived from a manuscript hypothesis) is not properly escaped before being interpolated into the git commit message string.

supervisor-drive