The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a 'runtime selection' mechanism described in references/runtime-selection.md and implemented in scripts/workflow_run.py. This mechanism resolves and executes a Python script (status_runtime.py) located within the repository being managed. This allows a target project to provide its own executable logic that overrides the skill's bundled helpers, creating a risk of arbitrary code execution if the project or branch being managed is untrusted.
[DATA_EXFILTRATION]: The scripts/jev_review.py script transmits project data (success criteria and evidence text) to an external service at api.typesafe.ai. While this feature is optional and requires explicit configuration via environment variables and an API key, it represents a managed path for project data to leave the local environment.
[COMMAND_EXECUTION]: Multiple components of the skill, including scripts/loop_run.py and scripts/_common.py, utilize subprocess.run and os.system equivalents to execute shell commands. The skill is specifically designed to run arbitrary 'verify' and 'trigger' commands defined in project artifacts like LOOP.md and task files, granting it high-privilege access to the local shell.
[INDIRECT_PROMPT_INJECTION]: The skill processes and acts upon a wide variety of files within the managed project, including LOOP.md, task briefs, and discussion records (DIALOGUE.md, ANSWERS.md). This large ingestion surface for untrusted data, combined with the skill's ability to execute both shell commands and project-provided Python scripts, provides a significant vector for indirect prompt injection attacks.

gsd-path-ship