The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/proof_builder.py executes a benchmark_cmd argument via subprocess.run(shell=True), allowing for arbitrary shell command execution during the proof phase.
[COMMAND_EXECUTION]: scripts/branch_guard.py executes Git commands using subprocess.run(shell=True). While it validates the slug argument with a regex to prevent command injection, the reliance on shell execution remains a minor security risk.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to instructions embedded in the project files it scans.
Ingestion points: Entire codebase scanned via glob and read operations in SKILL.md (Step 1).
Boundary markers: No explicit delimiters are used to isolate ingested code from the agent's instructions.
Capability inventory: Includes Git operations and arbitrary shell command execution in scripts/branch_guard.py and scripts/proof_builder.py.
Sanitization: Ingested content is not sanitized or escaped before being used in the agent's brainstorming and implementation phases.

next-best-improvement