The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/prepare_review_bundle.py executes local git commands using the subprocess module to collect repository information, branch status, and code diffs.
Evidence: The function run_git in scripts/prepare_review_bundle.py calls subprocess.run(["git", "-C", str(repo), *args], ...).
The commands are used for standard developer operations such as git diff, git rev-parse, and git status.
[COMMAND_EXECUTION]: The SKILL.md and references/isolation-protocol.md files provide documentation and examples for running external CLI tools such as codex and claude to perform automated code reviews.
The instructions utilize standard flags such as --ephemeral and --no-session-persistence to maintain execution isolation.
[DATA_EXPOSURE]: The skill aggregates repository diffs, test outputs, and task descriptions into a local .agent/code-reviews/ directory.
This data is stored locally within the project workspace and is intended for consumption by authorized reviewer agents as part of the documented workflow.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (code diffs and task descriptions) which are then bundled for a reviewer agent.
Ingestion points: request.md, writer-summary.md, and diff.patch created in scripts/prepare_review_bundle.py.
Boundary markers: The skill includes specific instructions to the reviewer to ignore previous chat context and focus strictly on the provided bundle files.
Capability inventory: The reviewer agent is granted limited permissions (workspace-write) to create review artifacts.

code-reviewer