The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands provided in configuration files (run_cmd and entry). This is core to its functionality as a research runner, allowing it to execute training or evaluation scripts.
[COMMAND_EXECUTION]: The skill implements a 'sandbox' mode where the agent authors its own Python or shell scripts and subsequently executes them. This dynamic script generation and execution pattern occurs within the specified sandbox root.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface. It reads and processes duel_log.md and iteration logs produced by its own previous code runs to 'borrow ideas' and influence future logic. This feedback loop could allow malicious or unexpected outputs from a research run to manipulate the agent's behavior in subsequent iterations.
Ingestion points: Reads iteration results from duel_log.md, results.tsv, and log files in the sandbox.
Boundary markers: No specific delimiters are used to warn the agent to ignore instructions embedded in these logs.
Capability inventory: The skill can execute shell commands, write files, and spawn sub-agents (on supported platforms like Claude Code).
Sanitization: No sanitization or validation of the ingested log content is performed before the agent interprets it.
[PROMPT_INJECTION]: The instructions explicitly tell the agent to 'never pause to ask' and to 'not pause for permission once the loop is running.' This high degree of autonomy reduces user oversight of the commands being executed.

dueling-autoresearch