Run this task with `hai` (https://github.com/braincore/hai-cli).
Install `hai` from your terminal: curl -LsSf https://hai.superego.ai/hai-installer.sh | sh
Run this task from your terminal: hai task ken/task-safety-checker

---

name = "ken/task-safety-checker"
version = "1.0.0"

description = "Review other tasks for safety concerns"

steps = [
    "/pin This task will help you review another task for safety concerns.",
    "/ask-human Which task would you like to review for safety?",
    "!hai Load the task with /task-view",
    """/prompt Identify all potentially unsafe steps in the task:

- Destructive commands (Examples below)
  - File deletion: `/exec rm ...` or `!!rm ...`
  - Asset deletion: `/asset-remove ...`
- Unknown files and assets that are used in potentially dangerous ways
  - File downloaded and then executed: `!!wget ... -o script.py`, `!!python script.py`
  - Asset downloaded and then executed: `/asset-export ...`, `!!python ...`o
- Destructive tool use
  - These ask the AI to generate potentially dangerous programs and commands
  - "!sh delete home dir", "!py delete my documents"

Any other potential risks or things the user should be aware of.

Output a table of each step that is potentially dangerous and why.
""",
    """!hai Review the next step from the table that hasn't been reviewed (starting from the top).

You should return a list of commands like so:

1*. Load any opaque resources from that step into the conversation for review.
   - e.g. /exec curl <opaque url being used by task step>
   - e.g. /asset-view <opaque asset being used by the task step>
2. /prompt Review the loaded resources
3. !hai Prompt yourself to review the next task step.

* may require multiple commands to load all resources
Step #2 is only necessary if resources were loaded for further review.
Step #3 is only necessary if there is a task step that hasn't been reviewed.
""",
    "Provide a summary of your findings and recommendations for the user.",
]