The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text from GitLab merge requests and ClickUp/GitLab tasks, creating an attack surface where malicious instructions could influence the agent's review logic.
Ingestion points: Untrusted data enters the agent context from merge request descriptions, source branch names, and external task/issue comments via the glab mr view, glab issue view, and clickup task/comments commands.
Boundary markers: The instructions do not define strict delimiters for external content, although the agent is instructed to prioritize established project standards over task-specific instructions if they conflict.
Capability inventory: The agent has access to powerful tools including Bash, git, and glab, enabling it to read local files, perform network requests, and modify merge request discussions.
Sanitization: While the skill uses pattern matching to extract task IDs, it does not explicitly sanitize the broader natural language descriptions or comments before processing them.
[COMMAND_EXECUTION]: The skill utilizes complex shell command sequences to perform its duties, including git operations and API interactions.
It uses git worktree to isolate the code being reviewed from the developer's working directory, which is a secure practice to prevent accidental local changes.
It generates and manages temporary files using mktemp for GitLab API payloads, ensuring that potentially sensitive review comments are not stored in persistent shared locations.
It performs glab api calls with arguments derived from merge request metadata, though it includes logic to extract IDs that follow specific alphanumeric patterns.
[EXTERNAL_DOWNLOADS]: The skill depends on several external CLI tools and update mechanisms.
It requires glab (GitLab CLI) and clickup CLI to be pre-installed and authenticated on the host machine.
It utilizes an official Figma MCP to fetch design screenshots and tokens for UI verification.
The metadata specifies an update path npx skills update, which involves fetching and executing a package from a public registry.

review-mr