chatgpt-deep-research-pdf-download

name	chatgpt-deep-research-pdf-download
description	Locates Deep Research reports in ChatGPT threads and saves them as PDF via Export → PDF inside the deep_research sandbox (user-playwright MCP). After saving, use scripts/sync_chatgpt_deep_research_pdfs.py record (local helper) to update deep-research-download-log.json.

ChatGPT: Deep Research PDF download

What you are looking for

In a thread that ran Deep Research, the UI typically includes a distinct report or artifact block (long document, citations, headings). Export is usually under a share or overflow menu on that report, with an action labeled like Download as PDF, Export, or PDF.

Labels change; always use the current browser_snapshot text and roles to pick the right control.

High-level click path

1. Open the thread from the project thread list (click ref from latest snapshot).
2. browser_wait_for content settle; browser_snapshot.
3. Scroll the main conversation pane until the Deep Research report is visible (use browser_run_code with page.mouse.wheel or scroll into view if needed).
4. Open the report’s share / more / export menu (icon near the report header or top-right of the card—match snapshot).
5. Click Download as PDF (or equivalent wording).

Iframe detail (important)

The report UI lives under iframe[src*="deep_research"] (MCP snapshots may show iframe[title="internal://deep-research"]). The Export control is usually in that frame’s inner document (often the first child frame). Prefer browser_snapshot and browser_click with refs from the latest snapshot; use browser_run_code when you need waitForEvent('download') and download.saveAs(path).

After Export, a [role="menu"] often lists Copy contents, Export to Markdown, Export to Word, Export to PDF. If the PDF row is localized, match /pdf/i on menu buttons or use the last menu button when the order matches that pattern.

Playwright download (MCP browser_run_code)

Use browser_run_code so the download is captured reliably. Set the save path to an absolute path under the repo project folder, e.g. Windows: C:/proj/chatgpt_pro_deep_research/engineering/My_Report.pdf. Do not rely on Node require('fs') in MCP snippets unless your environment supports it; download.saveAs accepts a string path.

Local Python does not perform downloads; it only helps with inventory, README tables, and log records (AGENTS.md).

Pattern (adapt frame chain and selectors from live snapshot):

async (page) => {
  const base = 'C:/proj/chatgpt_pro_deep_research/<project_folder>/';
  // Open thread URL first (outside this snippet).
  const inner = page.frameLocator('iframe[title="internal://deep-research"]'); // or src*=deep_research
  const root = inner.frameLocator('#root'); // if snapshot shows #root; else use child frame from snapshot
  await root.getByRole('button', { name: 'Export' }).click({ force: true });
  await page.waitForTimeout(800);
  const menu = root.locator('[role="menu"]');
  await menu.waitFor({ state: 'visible', timeout: 35000 });
  let pdf = menu.getByRole('button', { name: /pdf/i });
  if ((await pdf.count()) === 0) pdf = menu.getByRole('button').last();
  const [download] = await Promise.all([
    page.waitForEvent('download', { timeout: 300000 }),
    pdf.click({ force: true, timeout: 180000 }),
  ]);
  const name = download.suggestedFilename().replace(/[\\/:*?"<>|]/g, '-');
  await download.saveAs(base + name);
  return { saved: base + name };
}

Notes:

• If the menu is nested, perform two steps: open menu (click), then waitForEvent('download') + click PDF in a second browser_run_code or use browser_click between steps with fresh snapshots.
• If waitForEvent times out, re-snapshot: labels vary (Export to PDF, Download PDF, localized text).

If MCP download path is restricted

Some environments save to a default downloads directory. If saveAs fails or file appears elsewhere:

1. Search the MCP/browser documentation or default download directory for the newest .pdf.
2. Move the file into <repo>/<project_folder>/ using shell commands in the workspace.

De-duplication

Before saving, list existing *.pdf in the target folder. If a file with the same name exists, compare size or append _2, _3 before .pdf.

After each successful PDF

1. Run the local helper (from repo root):
   uv run python scripts/sync_chatgpt_deep_research_pdfs.py record --conversation-id <uuid> --folder <folder> --outcome downloaded --pdf-filename "<file>.pdf"
   Use conv-id subcommand on the thread URL if you need the UUID.
2. Refresh the folder README table: uv run python scripts/sync_chatgpt_deep_research_pdfs.py update-readmes --only <folder> (or update-readmes for all).
3. Keep human-readable title and project notes for the README index skill if you edit prose manually.