The Agent Skills Directory

[DATA_EXFILTRATION]: The skill transmits repository data, including diffs and file contents, to external model providers. Specifically, code is sent to OpenAI via the codex CLI and to GLM via the opencode CLI during the automated review process.
[COMMAND_EXECUTION]: The skill frequently executes shell commands to perform its core functions. These include repository interactions (git diff, git show, git grep), GitHub PR management (gh pr view), and execution of external model runners (codex exec, opencode run). It also involves writing temporary files (diffs) directly into the repository root.
[EXTERNAL_DOWNLOADS]: The skill requires the installation and use of external third-party command-line interfaces, specifically codex (for GPT access) and opencode (for GLM 5.2 access), which are not standard system tools or verified platform utilities.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection.
Ingestion points: Untrusted data enters the agent context via gh pr view, gh pr diff, and git diff commands.
Boundary markers: The skill attempts to enforce boundaries by requiring specific JSON output formats and forbidding narration in sub-agent responses, though these are weak security controls.
Capability inventory: The skill possesses extensive capabilities, including subprocess execution (git, gh, codex, opencode) and local file system writes.
Sanitization: No explicit sanitization or filtering of the ingested PR content is performed before it is passed to the analysis models.

delegate-review