The Agent Skills Directory

[COMMAND_EXECUTION]: The skill discovers and executes shell commands for testing, building, and linting directly from repository configuration files like package.json, Makefile, and Cargo.toml during the discovery and static analysis phases.
[EXTERNAL_DOWNLOADS]: It performs network operations via git fetch to synchronize with remote repositories when resolving branch diffs and identifying the scope of changes.
[INDIRECT_PROMPT_INJECTION]: The skill represents an indirect prompt injection surface as it ingests and processes untrusted data from the codebase.
Ingestion points: The agent reads git diffs, source file contents, and linter/test outputs (Phase 1, 4, 6).
Boundary markers: It uses a structured CONTEXT_BUNDLE to separate metadata from content for sub-agents, though it depends on the sub-agents' own safety filters for the content itself.
Capability inventory: The pipeline possesses high capabilities, including arbitrary shell command execution (Phase 3, 4, 7), network access in the exerciser (Phase 7c), and file-system write access for automated fixes (Phase 9).
Sanitization: The skill parses tool outputs into structured findings but does not explicitly mention sanitization of the code changes before they are passed to the LLM for review.
[COMMAND_EXECUTION]: The skill interpolates user-supplied arguments from $ARGUMENTS (such as --base and --files) directly into shell command templates, which could be exploited for command injection if the executing agent fails to sanitize the input.

verify