gemini-deep-research（Gemini deep 研究）

Gemini Deep Research

Run long-form research queries through Google's Gemini Deep Research via browser automation (Playwright CDP). Produces 40k+ char markdown reports with source citations.

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

Prerequisites

1. Chrome with debug port: Must be running with --remote-debugging-port=9222
2. Gemini Advanced subscription: Logged into gemini.google.com in the debug Chrome
3. playwright-core: bun add -g playwright-core (or project-local)
4. Runtime: Use npx tsx (not bun run) — Bun's CDP connectOverCDP times out; Node.js connects in <1s

Launch Chrome (if not running)

/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 \
  --user-data-dir="/tmp/gemini-research-profile" \
  "https://gemini.google.com/app" &

Then log in manually with a Gemini Advanced account.

Usage

CLI (direct)

# Health check — verify Chrome CDP + Gemini login
npx tsx {{skill_dir}}/scripts/research.ts --health

# Basic research (runs preflight automatically)
npx tsx {{skill_dir}}/scripts/research.ts "your research query"

# Save to specific file
npx tsx {{skill_dir}}/scripts/research.ts \
  --output /tmp/report.md \
  --timeout 45 \
  "comprehensive analysis of quantum computing error correction 2025-2026"

# Auto-save to directory (creates {date}-{slug}.md)
npx tsx {{skill_dir}}/scripts/research.ts \
  --output-dir ~/.claude/automation/gemini-deep-research/output \
  "your query"

# Without auto-confirming plan (lets you review first)
npx tsx {{skill_dir}}/scripts/research.ts --no-confirm "query"

Programmatic (import)

import { GeminiDeepResearchClient } from "{{skill_dir}}/scripts/client.js";

const client = new GeminiDeepResearchClient({
  cdpUrl: "http://127.0.0.1:9222",
  maxResearchTimeMs: 30 * 60 * 1000,
  autoConfirm: true,
  onProgress: (msg) => console.log(msg),
});

await client.init();
const result = await client.research("your query");
// result.report — full markdown report (40k+ chars)
// result.plan — research plan text
// result.completed — boolean
// result.durationMs — execution time
// result.shareLink — Gemini share URL (if Firecrawl enabled)
await client.close();

Preflight

Every research run starts with an automatic preflight health check that verifies:

1. Chrome CDP reachable on configured port
2. Browser connection via WebSocket succeeds
3. Gemini page open at gemini.google.com
4. Login state OK (not showing sign-in wall)

If any check fails, research aborts with a clear error message. Use --no-preflight to skip.

Automation Flow

Preflight (CDP + login check) → abort if unhealthy
    ↓
Chrome CDP:9222 → Navigate gemini.google.com/app
    ↓
Tools button → Deep Research drawer item → Active chip verification
    ↓
Type query (30ms/char) → Send button (or Enter fallback)
    ↓
Wait for research plan (~18-120s) → Extract plan text
    ↓
Auto-confirm "Start research" (or manual)
    ↓
Poll completion: mic button + text stability (5s intervals, 30min max)
    ↓
Extract report (longest .markdown element) → Optional share link + Firecrawl

Debug Probes

When selectors break (Google updates Gemini UI), use the probe scripts:

# Check Chrome connectivity
bun run {{skill_dir}}/scripts/probes/dom-inspector.ts status

# Test all selectors against live DOM
bun run {{skill_dir}}/scripts/probes/dom-inspector.ts selectors

# Full DOM inspection
bun run {{skill_dir}}/scripts/probes/dom-inspector.ts probe

# Monitor active research execution
bun run {{skill_dir}}/scripts/probes/research-monitor.ts confirm-and-monitor

# Check research completion + extract share link
bun run {{skill_dir}}/scripts/probes/share-link.ts status
bun run {{skill_dir}}/scripts/probes/share-link.ts extract

Selector Registry

All CSS selectors live in scripts/selectors.ts. When Google updates the Gemini UI:

1. Run dom-inspector.ts selectors to identify broken selectors
2. Run dom-inspector.ts probe to inspect current DOM
3. Update selectors.ts with new selectors
4. Re-test with dom-inspector.ts selectors

Selectors last verified: 2026-03-13 (Tools button: now button.toolbox-drawer-button, aria-label removed)

Key Files

Options Reference

Completion Detection

Research completion is detected via three signals:

1. Mic button visible — button[data-node-type="speech_dictation_mic_button"] reappears
2. Report text > 500 chars — longest .markdown.markdown-main-panel element
3. Text stability — 3 consecutive identical text lengths (15s total)

The spinner may remain visible as a stale artifact after completion — the mic button is the primary signal.

Post-Execution Reflection

After this skill completes, check before closing:

1. Did the command succeed? — If not, fix the instruction or error table that caused the failure.
2. Did parameters or output change? — If the underlying tool's interface drifted, update Usage examples and Parameters table to match.
3. Was a workaround needed? — If you had to improvise (different flags, extra steps), update this SKILL.md so the next invocation doesn't need the same workaround.

Only update if the issue is real and reproducible — not speculative.