The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell commands through scripts/collect_context.sh and Python subprocesses in scripts/review_scope.py. These operations interact with git and the GitHub CLI (gh) to retrieve pull request metadata, diffs, and existing review comments. These actions are restricted to legitimate source control operations required for the review task.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from external sources, specifically pull request descriptions and comments, which could potentially contain malicious instructions intended to influence the AI's review behavior.
Ingestion points: Data is ingested via gh commands in scripts/collect_context.sh and stored in temporary artifacts.
Boundary markers: The system uses structured XML-like delimiters (e.g., <task>, <context>, <grounding_rules>) in its prompt templates to separate instructions from data. The synthesis instructions in SKILL.md explicitly direct the agent to treat existing comments as candidates rather than truth.
Capability inventory: The skill can write to the filesystem (temporary directories), execute shell commands (git, gh, and discovered model runners), and spawn sub-agents using the Agent tool.
Sanitization: A dedicated synthesis phase (Phase 5) uses a fresh model context and a detailed filtering pipeline (references/filtering_pipeline.md) to deduplicate, normalize, and evidence-check findings, which serves as a robust defense against adversarial data in the input.
[DYNAMIC_EXECUTION]: The skill performs dynamic model orchestration by probing for available command-line interfaces (codex-runner, claude-runner, etc.) and invoking them based on host availability. It also utilizes the Agent tool to spawn sub-agents for synthesizing findings. This dynamic behavior is a core architectural feature for cross-model verification and does not involve the execution of arbitrary external code.

full-review