The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection via project-specific configuration files.
Ingestion points: The skill reads configuration from <git-commit-config> within the CLAUDE.md file located in the project repository.
Boundary markers: Absent. There are no explicit delimiters or warnings to prevent the agent from treating instructions in CLAUDE.md as system-level overrides.
Capability inventory: The skill can execute git commands, a node validation script with interpolated arguments, and arbitrary shell commands for quality checks (linting, testing).
Sanitization: Absent. Configuration flags and 'extra-instructions' from CLAUDE.md are used directly in the workflow without validation or escaping.
[COMMAND_EXECUTION]: Potential for arbitrary command execution during the automated quality gate process.
Ingestion points: The agent is instructed to 'determine what checks apply' by inspecting the codebase, which includes reading potentially malicious scripts in package.json, Makefile, or other build configurations.
Boundary markers: Absent. The agent relies on its own assessment of the untrusted project files to decide which commands to run.
Capability inventory: Full shell access to execute build, lint, and test tools as part of the pipeline.
Sanitization: Absent. If a repository contains a malicious test script, the agent may execute it as part of the 'Quality Gate' step.

commit