gspdev-eval-changes

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 1.00). The prompt explicitly loads full BEFORE/AFTER file contents into evaluator agent prompts and requires outputs to include "Quoted BEFORE text" (verbatim snippets), so any secrets present in those files would be read by the LLM and must be reproduced in its output, creating an exfiltration risk.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill calls gh pr view --json body (i.e., the GitHub PR endpoint such as https://api.github.com/repos/{owner}/{repo}/pulls/{number}) at runtime to fetch a PR description that is parsed into the "change intent" and injected into evaluator prompts, so remote PR content can directly control the agent's instructions.