The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses node:child_process (specifically spawn and execFileSync) to interact with the local Git repository and to invoke external AI CLI tools like Grok, Claude, and Codex.
[DYNAMIC_EXECUTION]: In scripts/review-loop/adapters.mjs, the resolveSpawn function constructs shell commands to execute functions or binaries defined in the user's interactive shell (zsh or bash). While it applies a regular expression to sanitize the command name (/^[A-Za-z][A-Za-z0-9_.-]{0,63}$/), it still enables the execution of host-resident code based on names provided during the automated loop.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a data ingestion surface as it processes Git diffs and external review findings to drive its logic. This creates a potential vector where malicious code or comments within a reviewed repository could attempt to influence the instructions followed by the AI agents participating in the fix-and-review cycle.
Ingestion points: scripts/review-loop/schema.mjs parses review findings and reassessments from markdown text; scripts/review-loop/evidence.mjs reads Git diffs.
Boundary markers: The skill uses markdown headers (e.g., # Review Findings) and structured table formats to delineate data, though these are primarily for parsing structure rather than instruction isolation.
Capability inventory: The skill can execute shell commands, read/write files within the repository, and read/write configuration files in the user's home directory (scripts/review-loop/repositories.mjs, scripts/review-loop/auto-run.mjs).
Sanitization: The parser includes logic to detect unescaped pipe characters in markdown tables to prevent column-shifting errors during data extraction.

agentic-review-handoff