The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a collection of bundled Python scripts to orchestrate complex Git workflows. Scripts such as isolation.py, pipeline_git.py, and integration.py frequently invoke subprocess.run to execute git and gh (GitHub CLI) commands.
scripts/loop_run.py contains a run_shell function that executes arbitrary strings provided in a LOOP.md specification using /bin/sh -c on POSIX systems or bash.exe -c on Windows. This is a core functional feature used to verify code fixes within a controlled loop.
scripts/detect_project.py uses low-level system calls (ctypes.WinDLL, ntdll.NtCreateFile) to perform secure, anchored file operations on Windows, specifically designed to prevent symlink-based attacks during initialization.
[INDIRECT_PROMPT_INJECTION]: The skill has a defined attack surface for indirect prompt injection due to its core function of gathering research from external sources.
Ingestion points: references/researcher.md instructs the agent to "Search broadly" and check URLs or repository references. External content from these sources enters the agent's context during the research phase.
Boundary markers: The system uses standard Markdown fences and specific templates (templates/evidence.md) to delimit findings.
Capability inventory: The framework has extensive capabilities to write to the filesystem, manage Git branches, and execute shell commands through the loop_run.py helper.
Sanitization: The scripts perform structural validation (parsing frontmatter, checking Git history ancestry) but rely on the agent's internal safety filters for interpreting the natural language content retrieved from the web.

gsd-path-research