The Agent Skills Directory

[PROMPT_INJECTION]: The skill utilizes platform hooks, specifically harness-stop.py and self-reflect-stop.py, to intercept the agent's completion signal and inject mandatory instructions. These include self-reflection checklists and directives to continue working until specific task conditions are met, steering the agent's behavior through hidden prompt modifications.
[COMMAND_EXECUTION]: The framework's core functionality relies on the execution of arbitrary shell commands provided in project configuration files. Specifically, harness-tasks.json contains a validation.command field, and the system is designed to run harness-init.sh at the start of every session, granting the agent broad execution capabilities over local environment scripts.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a vulnerability surface by ingesting and acting upon instructions from externalized state files like harness-tasks.json. If these files are manipulated by untrusted data, the agent could be induced to execute unintended commands or follow malicious logic embedded within the task definitions.
[PERSISTENCE]: The skill implements a task persistence layer that maintains state across agent restarts and context window resets. It uses a Stop hook to block the agent from exiting as long as tasks remain in the queue, effectively enabling the agent to maintain an active presence until its internally defined goals are accomplished.
[DATA_EXPOSURE]: The self-reflect-stop.py hook programmatically accesses the session's transcript_path to extract and process previous user prompts. While intended to ensure goal alignment, this represents automated access to and processing of session metadata by the skill's logic.

harness