Deep Research (Gemini)

**Judgement:** A feature-rich, well-engineered deep research skill built around Google Gemini’s deep research agent and file search. It is strong on structured outputs, RAG integration, and agent-focused ergonomics, making it a powerful choice when you *intentionally* want to send queries and selected local files to Gemini. **Key strengths:** - Full deep-research lifecycle: start, async status polling, final report export (md/html/pdf), structured output directories with metadata and sources. - Robust RAG + file search: create/list/query/delete stores; upload with smart-sync and extension filters; ephemeral context stores for one-off grounded runs. - Agent/automation friendly: JSON on stdout, rich logs on stderr; compact JSON summaries for heavy outputs; non-interactive mode for CI/agents; adaptive polling. - Good operational controls: timeouts, research depth, cache toggles, dry-run cost estimation, max-cost guard, and history-based polling to reduce waste. **Main risks / limitations:** - **Data exfiltration to a third party:** Any research question and any files provided via `--context` or uploads are sent to Google’s Gemini APIs. This is expected, but inappropriate for highly sensitive or regulated data unless your org explicitly allows Google as a processor. - **Autonomous agents + filesystem access:** In non-interactive mode prompts are skipped; an agent with broad filesystem access could upload more than intended if misconfigured. Mitigations exist (restrict accessible paths, use `--dry-run`, `--max-cost`, and careful `--context`/upload paths) but require discipline. - **External dependency and latency:** Functionality depends entirely on network access and Gemini availability; deep research runs are multi-minute by design, so it’s not suitable for low-latency, in-request operations. **Recommended scenarios:** - As a delegated research tool inside an orchestrated agent system when you need long-running, high-quality analysis grounded in a specific repo or document set. - For engineering, security, or architecture reviews over codebases where you’re comfortable uploading that code to Gemini. - In CI or scheduled jobs that generate comprehensive reports (with citations and metadata) for later human or machine consumption. **Avoid or restrict use when:** - Data cannot leave your environment or be shared with Google for compliance, contractual, or privacy reasons. - You require strict, human-gatekept review of which files can ever be uploaded; in those cases, wrap this skill with additional policy and path whitelisting before giving it to autonomous agents.