The Agent Skills Directory

[COMMAND_EXECUTION]: The skill defines a mission-validator-script mode that relies on executing a mission_validator_command stored in a local state file (.omw/autoresearch/<slug>/autoresearch-state.json). The agent is instructed to repeatedly execute this command until a specific JSON status is returned, providing a mechanism for autonomous, repeated shell command execution.
[PROMPT_INJECTION]: The skill utilizes a deep-interview process to clarify the research mission and the evaluator (validator). There is a risk of indirect prompt injection where untrusted data processed during the intake phase could define a malicious mission_validator_command, leading the agent to execute arbitrary code under the guise of validation.
Ingestion points: Untrusted data enters the agent context through the deep-interview intake process used to materialize the research mission.
Boundary markers: None mentioned; the instructions do not specify delimiters or warnings to ignore instructions embedded in the research target.
Capability inventory: The skill has the capability to execute shell commands via the mission_validator_command and read/write state files in the .omw/ directory.
Sanitization: The skill does not describe any validation or sanitization of the mission_validator_command before it is saved to the state file and executed.
[PERSISTENCE]: The skill is explicitly designed as a stateful, autonomous loop that "keeps nudging" and "does not stop because the model says 'done'". This persistent nature ensures that any injected malicious validation command will be executed repeatedly across sessions until the agent is manually stopped or the condition is met.

autoresearch