The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes bundled Python scripts (pipeline_undo.py, loop_run.py, etc.) to execute system commands, primarily targeting git for repository state management.
Evidence:
scripts/_common.py provides wrappers for subprocess.run (e.g., run_git) used throughout the skill's scripts.
scripts/loop_run.py contains a run_shell function that executes commands defined in a repository's LOOP.md spec file. This is gated by resource limits (wall-clock time) and process group management to prevent orphaned processes.
[DYNAMIC_EXECUTION]: Several scripts use dynamic module loading and argument substitution to handle configuration and model selection.
Evidence:
scripts/path_config.py uses importlib.import_module to dynamically load sibling scripts like pipeline_state and review_panel. This is restricted to internal project helpers.
scripts/model_policy.py performs string substitution ({model_args}, {effort_args}) in command templates based on resolved model policies, facilitating flexible sub-agent dispatching.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository files (e.g., PLAN.md, LOOP.md, STATE.md) which could theoretically contain malicious instructions.
Evidence:
scripts/review_panel.py parses PLAN.md to configure cross-model review panels.
scripts/loop_run.py executes verification commands directly from LOOP.md headers.
Sanitization: These scripts use specific regex patterns and deterministic logic to parse machine-readable fields, reducing the risk of accidental instruction following from free-text sections. The risk is considered low and inherent to development automation tools.

gsd-path-undo