The Agent Skills Directory

[COMMAND_EXECUTION]: The skill operates by executing shell scripts (autoresearch.sh and autoresearch.checks.sh) created by the agent at runtime. This capability allows the agent to run any command-line tool available in the environment to perform benchmarks and correctness checks.
[DYNAMIC_EXECUTION]: The agent is tasked with writing and then executing code in a continuous, autonomous loop. This dynamic script generation and execution is a high-privilege activity essential to the skill's purpose but represents a significant capability for local code execution.
[PROMPT_INJECTION]: The instructions include directives that override standard interactive agent behavior for long-running tasks, such as 'LOOP FOREVER' and 'Never ask "should I continue?"'. This increases the risk of unintended or runaway actions if the agent is left unsupervised.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to instructions embedded in the project codebase or benchmark outputs, as it reads all files in the scope to inform its optimization strategy.
Ingestion points: All files in the user-specified project scope and the console output (stdout/stderr) of the generated benchmark scripts.
Boundary markers: No specific delimiters or safety warnings are provided for codebase ingestion; benchmark outputs are parsed using a simple METRIC prefix.
Capability inventory: File system modification, shell command execution, and git history manipulation (commit, revert, branch).
Sanitization: The skill relies on basic regular expressions in helper scripts to extract numeric values from benchmark output, providing minimal protection against crafted strings.

autoresearch