The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes 'Intent', 'Snap', and 'Decision' objects which contain natural language fields such as 'what' and 'why'. These fields are read back during the recovery phase via the itt inspect command. If these fields contain malicious instructions, they could influence the agent's behavior.
Ingestion points: Data entering the agent context via itt inspect from files stored in .intent/.
Boundary markers: None explicitly present in the data format to distinguish instructions from historical text.
Capability inventory: The skill can execute itt commands and perform network operations via itt push (in scripts/itt_argv.py and src/intent_cli/hub/client.py).
Sanitization: No explicit content filtering is performed on the historical text before it is presented to the agent.
[EXTERNAL_DOWNLOADS]: The documentation in README.md provides instructions for users to install the required CLI tool by downloading and executing shell and PowerShell scripts directly from the project's GitHub repository.
[COMMAND_EXECUTION]: The skill and its underlying CLI tool utilize subprocess.run and subprocess.Popen in several locations (e.g., src/intent_cli/store.py, src/intent_cli/hub/credentials.py, and scripts/itt_argv.py) to interact with Git, Docker, and the itt binary itself. The skill instructions specifically require using a bundled Python runner (scripts/itt_argv.py) that disables shell evaluation and uses strict argument encoding to mitigate command injection risks.
[REMOTE_CODE_EXECUTION]: The setup process documented in README.md involves piping remote content to a shell for installation (e.g., curl ... | bash). This behavior is limited to the initial setup by the user and targets scripts hosted within the vendor's own repository.

intent-cli