The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute shell commands and local scripts to perform its functions. Specifically, it invokes git for diffing and status checks, bun to run scripts/review-receipt.mjs for posting PR comments, and node to run scripts/unit-route.mjs for task routing. These operations are core to the skill's purpose as an automated quality gate.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data in the form of branch diffs and project documentation (e.g., ACCEPTANCE.md).
Ingestion points: Untrusted source code changes are read via git diff and processed during review axes in REVIEW_PROCESS.md and SKILL.md.
Boundary markers: The skill implements an isolation rule by spawning subagents with limited context, but the instructions do not specify the use of robust text-based delimiters (like XML tags or unique markers) to separate the untrusted diff content from the agent's instructions.
Capability inventory: The agent has the capability to execute shell commands (git, bun, node), write to local files (the review-findings.md ledger), and post comments to GitHub Pull Requests.
Sanitization: There is no explicit mention of sanitizing or escaping the content of the diff hunks before they are passed to the reviewer agents, relying instead on the subagent's isolation.
[DYNAMIC_EXECUTION]: To achieve isolated review passes, the skill dynamically orchestrates the creation of subagents (within Claude Code) or parallel headless execution sessions. This runtime generation of independent agent tasks is used to decorrelate blind spots and ensure findings are gathered in clean contexts.

review-change