chatgpt-deep-research

ChatGPT Deep Research — Platform Knowledge

Standalone Invocation

If invoked directly via /chatgpt-deep-research, spawn the browser-researcher agent:

Agent(
  subagent_type="browser-researcher",
  name="chatgpt-researcher",
  prompt="skill_path: <skill-path>/SKILL.md\nresearch_prompt: <from user>\noutput_path: <from user or .work/chatgpt-report.md>\ncaller: <your name>",
  run_in_background=true
)

Then relay messages between user and agent via SendMessage. If loaded by browser-researcher via Read, ignore this section.

Platform Configuration

• url: https://chatgpt.com/deep-research
• backend: playwright-cli
• submit_method: click_send_button
• plan_review: informational # research auto-starts; approve-plan is a no-op
• hard_timeout: 90
• completion_signal: text "Research completed in" inside nested root iframe (sandboxed)

Playwright-CLI Backend

All browser interaction is driven by scripts/pw-driver.py via Bash. The agent calls subcommands; the driver handles Chrome lifecycle, CDP attach, and all DOM interaction internally.

Phase	Subcommand
Login check	python3 scripts/pw-driver.py status (exit 10 = login required)
One-time login	python3 scripts/pw-driver.py login (user runs interactively)
Research activation + submit	python3 scripts/pw-driver.py start-research --prompt-file=<path>
Plan capture	python3 scripts/pw-driver.py wait-plan --timeout=120
Plan approval	python3 scripts/pw-driver.py approve-plan
Research monitoring	python3 scripts/pw-driver.py wait-complete --timeout=5400
Report extraction	python3 scripts/pw-driver.py extract --output=<path>
Session close	python3 scripts/pw-driver.py close

Exit codes: 0 success · 10 login required · 20 timeout · 30 extraction failure · 40 DOM mismatch.

DOM Reference (verified March 2026)

Key elements

• button "Deep research" — version dropdown chip in input bar
• textbox with placeholder "Get a detailed report" — main input
• button "Copy response" — copy assistant message to clipboard

Completion indicators

• Stats: "Research completed in Xm · Y citations · Z searches"
• Report card with title and icons appears
• Response action buttons (Copy, Thumbs up/down, Share, Refresh) appear

Platform Quirks

• Obfuscated CSS: ChatGPT uses obfuscated class names — driver uses accessible names, not CSS selectors.
• Deep Research already active: Navigating to /deep-research pre-selects the mode. Driver verifies but doesn't need to toggle.
• Max plan iterations: 5 iterations before asking caller to approve or abort.
• Follow-up mode: After research completes, follow-ups go through standard ChatGPT chat (NOT Deep Research) to save quota.
• Report container: [data-message-author-role="assistant"] — select last element for the research response.