The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions to override the agent's standard interactive behavior and safety checkpoints by directing it to skip confirmation prompts and run indefinitely. Evidence in SKILL.md: 'Never stop. Once running, do not pause to ask "should I continue?" — the loop runs until manually interrupted.' and 'do not pause the loop to ask for direction.'
[COMMAND_EXECUTION]: The skill executes shell commands based on an analysis plan that can be influenced by untrusted sub-agents, creating an indirect prompt injection surface. Evidence in SKILL.md (Step 6): 'Analyse — execute every plan.md row'. Ingestion points: ResearchAgent ideas (specifically the 'note' field used for diagnostics) are ingested from sub-agents. Boundary markers: While JSON schema is used for data transfer, the content is interpolated into executable plans without clear delimiters. Capability inventory: The Judge agent has the capability to execute shell commands to perform analysis and run the training entrypoint. Sanitization: There is no defined sanitization or validation for the diagnostics proposed by sub-agents before they are executed.
[COMMAND_EXECUTION]: The skill dynamically creates and executes shell scripts. Evidence in SKILL.md: Instructions to write and execute a timeout script at '<sandbox_root>/run_with_timeout.sh'.

tournament-autoresearch