The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted git diffs and file paths to determine the execution of high-privilege sub-agents.
Ingestion points: Git observable signals, directory structures, and changed file paths are consumed by risk-aware predicates in SKILL.md to select the reviewer set.
Boundary markers: The skill does not implement explicit boundary markers or instructions to sub-agents to ignore potentially malicious content within the diffs being reviewed.
Capability inventory: The skill executes git operations (git add, git commit, git status) and manages a test harness with filesystem write access (run_evals.py).
Sanitization: Input validation for run_id using a restrictive regular expression (^[A-Za-z0-9_][A-Za-z0-9_-]{0,31}$) is implemented in evals/_runid.py to prevent path traversal attacks when creating dispatch directories.
[COMMAND_EXECUTION]: The skill is designed to execute shell commands to manage the git lifecycle during the review process.
Evidence: SKILL.md explicitly defines the use of git add -A && git commit to commit working-tree fixes generated by sub-agents such as temper, delve, and quality-gate.
Context: This capability is intrinsic to the skill's primary function as a review gate that fixes and certifies the codebase before merging.

warden