The Agent Skills Directory

[COMMAND_EXECUTION]: The executor subagent (defined in agents/executor.md) utilizes shell commands such as wc, find, and tr to calculate character counts of transcripts and output files as a proxy for token usage metrics.
[DYNAMIC_EXECUTION]: The agents/executor.md instructions include a Python heredoc pattern (python3 << EOF) used to programmatically update metrics.json with execution results. This is a localized technique for data management within the skill's workspace.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it is designed to ingest and process untrusted data (eval prompts and test files) from the evals/ directory to test other skills.
Ingestion points: Evaluation prompts and input files are loaded from evals/evals.json and the evals/files/ directory.
Boundary markers: No explicit boundary markers or specific 'ignore instructions' warnings are defined in the subagent prompts to separate eval data from instructions.
Capability inventory: Subagents have access to standard file operations and shell execution tools (Read, Write, Bash) to perform their tasks.
Sanitization: There is no evidence of sanitization or escaping of the ingested eval content before it is processed by the subagents.
[PRIVILEGE_ESCALATION]: The scripts/init_skill.py script uses chmod(0o755) to ensure that newly created helper scripts are executable. This is a standard and benign file permission adjustment for the skill's initialization process.

skill-creator