The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates implementation decisions by ingestings data from research evidence files, which presents a surface for indirect prompt injection.
Ingestion points: Processes INTENT.md, RESEARCH.md, and evidence-*.md as specified in SKILL.md and the decider role brief.
Boundary markers: Employs structural markdown templates (e.g., templates/synthesis.md) and explicit role-based constraints defined in references/decider.md to manage context.
Capability inventory: The execution environment allows for Git mutations and arbitrary shell command execution via bundled scripts.
Sanitization: No specific content filtering or sanitization of ingested markdown evidence is implemented before analysis by the sub-agent.
[COMMAND_EXECUTION]: The skill provides the agent with high-privilege automation capabilities via shell and Git execution.
The scripts/loop_run.py script executes shell commands defined in a LOOP.md spec file using /bin/sh or bash.exe.
Scripts such as scripts/isolation.py and scripts/pipeline_git.py execute various Git commands using subprocess.run to manage project structure and lifecycle.
[EXTERNAL_DOWNLOADS]: The skill performs necessary network synchronization for project integration.
scripts/integration.py uses the GitHub CLI (gh api) to create and manage pull requests on GitHub.com.
Multiple scripts use standard Git commands like fetch, ls-remote, and push to interact with remote origin repositories.
[DYNAMIC_EXECUTION]: The skill uses dynamic code patterns to manage its components and maintain state integrity.
scripts/path_config.py uses importlib.import_module to load sibling runtime scripts.
scripts/detect_project.py utilizes ctypes to call low-level Windows APIs (kernel32.dll and ntdll.dll) to ensure atomic and safe file creation.

gsd-path-decide