Version 4.1.0

Deep Research export and element-picker hardening release focused on reliability in maximized ChatGPT Deep Research views, iframe contexts, and saved HTML (srcdoc) exports.

Added

1. Unified Deep Research section collector/assembler across all DR export paths.
2. Recursive nested srcdoc parser and fallback doc scoring.
3. Ranked picker targeting and same-origin nested-iframe descent.
4. Explicit cross-origin isolation guidance for picker limitations.

Changed

1. Bridge request/response schema is richer and backward compatible.
2. Panel extraction now favors semantic ARIA hooks first, then fallback selectors.
3. Export assembly now honors toggles consistently and appends sections in a fixed order.

Removed / Retired

1. Retired effective dead-path behavior where panel extraction logic could exist but not be invoked in final output assembly.

Compatibility / Notes

1. No breaking changes to core user-facing controls (Ctrl+M, R, G, menu commands, modal workflow).
2. incSources compatibility retained for older bridge interactions.
3. Connector scanned section is exported when present, including valid empty-state output when no connector sources are available.

Version 4.0.0

How to use (three methods):

• Method A: Press R with picker active. The parent sends h2m-auto-export to the iframe, which runs local export and shows the modal inside the iframe.
• Method B: Click into the iframe to give it focus, then press Ctrl+M to activate the picker inside the iframe directly. Hover, arrow-navigate, click to export any sub-element.
• Method C: In the parent picker, scroll down on the highlighted iframe to send h2m-start-picker, activating the picker inside the iframe.

Version 3.2.0

Version 3.1.0

1. Removed the early return (line 94): initBridge() is still called inside the iframe to set up the bridge for parent-page automated exports, but execution now continues to initialize the full script -- styles, element picker, modal, h2m converter, and all keyboard shortcuts -- inside the iframe's own document context. This means Ctrl+M inside the iframe activates the picker, hovering highlights sub-elements, clicking exports them, arrow keys navigate the DOM tree, and R triggers the full Deep Research export with sources and activity -- all running natively inside the iframe with direct DOM access.

2. autoExportDR() now handles the iframe-local case first: When IS_DR_IFRAME is true, it extracts the report, clicks the Sources tab, extracts sources, clicks the Activity tab, extracts activity -- all directly from the local DOM without any postMessage round-trip. This is the most reliable path.

3. Cross-frame picker activation via postMessage: Added h2m-start-picker / h2m-stop-picker message handlers. When the parent page's picker highlights an iframe and the user scrolls down into it (or clicks a non-DR iframe), the parent sends a message to activate the picker inside that iframe. This works for any iframe on any website where the script is running, not just Deep Research.

4. sendPickerToIframe() function: New helper that posts h2m-start-picker to an iframe's contentWindow and shows a toast. Called when scrolling down on an iframe element or clicking a non-DR iframe.

5. Smarter mousedown handler: Clicking an iframe now distinguishes between DR iframes (triggers autoExportDR) and generic iframes (sends picker activation message). Clicking any non-iframe element inside the iframe does a normal h2m conversion and shows the modal.

6. Robust fmtOut for iframe context: Uses document.referrer (the ChatGPT URL) as the source URL in YAML frontmatter when running inside the iframe, instead of the sandbox URL.

7. Robust download filename: Falls back to h1 text content when document.title is empty (common inside iframes).

8. Updated tip text: Concise instructions including the new "scroll down on iframe to enter" interaction.

How to use inside the maximized Deep Research panel:

• Method A (recommended): Press Ctrl+M on the parent page, hover the iframe, scroll down to "enter" it. The picker activates inside the iframe. Now hover, click, arrow-navigate, and scroll to select any sub-element of the report, aside panel, individual citations, etc.
• Method B: Click into the iframe first (to give it focus), then press Ctrl+M directly inside the iframe. The picker activates natively.
• Method C: Press R (with picker active, from either parent or iframe) for a full automated Deep Research export including report + sources + activity.

Version 3.0.0

Complete rewrite of initBridge() (the code that runs inside the iframe):

• getReportRoot() finds [class*="_reportPage_"] directly in the iframe's document
• getAside() finds the <aside> element containing the tab panel
• clickTab(name) clicks the Sources or Activity tab button (button[role="tab"]) to switch the visible panel content
• extractSourcesFromAside() extracts all a[href][target="_blank"] links from the aside after the Sources tab is active
• extractActivityFromAside() extracts thinking activity entries from the aside after the Activity tab is active
• The message handler is now async so it can await sleep(800) after tab clicks to let React re-render
• The response now includes sources and activity fields alongside the report md

Updated exportViaIframe(): Passes incSources and incActivity flags to the bridge, receives the new response fields, timeout increased to 20s to accommodate tab-switching delays.

Updated autoExportDR(): Assembles the report + sources + activity from the bridge response with proper headings. Falls back to parent-side extractors only if the iframe bridge returns nothing (for the non-maximized case).

Version v2.4.0 Changes

Root Cause: The DOM is Split Across Two Origins

The previous versions had a fundamental architectural misunderstanding. By testing the actual HTML example files, I confirmed:

v2.2.0 and v2.3.0 tried to extract citations/scanned/activity from inside the iframe bridge, but these elements don't exist there. They exist in the parent page's DOM and are directly queryable.

What v2.4.0 Changes

1. Removed bridge extraction functions -- bridgeExtractCitations, bridgeExtractScanned, bridgeExtractActivity were deleted from initBridge(). The bridge now only extracts the report body (which IS inside the iframe).

2. Fixed autoExportDR() -- Now correctly: (a) gets the report body via the iframe bridge, then (b) extracts citations/scanned/activity directly from the parent page DOM using the existing extractDRCitations(), extractDRScanned(), extractDRActivity() functions.

3. Added tryDirectIframeAccess() fallback -- If the postMessage bridge times out (e.g., the userscript hasn't loaded in the iframe yet), attempts direct DOM access via iframe.contentDocument (works if allow-same-origin grants access).

4. Improved initBridge() -- Better content detection: looks for _reportPage_ class specifically, handles nested iframes, more robust wr() readiness check.

5. Broadened hasDROverlay() -- Also detects _reportContainer_ and section[aria-labelledby^="report-"] elements in the parent DOM.

6. Improved exportViaIframe() -- Broadcasts postMessage to all iframes (handles triple-nesting where the userscript might run in a different iframe than expected).

7. Element picker -- Side panel elements (citations, activity) are in the parent DOM and CAN be highlighted/selected normally. Only the iframe content itself is opaque to elementFromPoint. Clicking any element detected as DR-related triggers autoExportDR().

v2.3.0 Changelog

Root Cause

When the ChatGPT Deep Research panel is maximized, the entire report (including citations, scanned sources, and research activity) renders inside a cross-origin iframe at web-sandbox.oaiusercontent.com. This is a fundamental browser security boundary:

• document.elementFromPoint() from the parent page returns the <iframe> element itself, never its children
• document.querySelector(...) on the parent page cannot find section[aria-labelledby="report-references-citations"] or any other DR elements because they exist only inside the iframe's DOM
• The existing initBridge() function ran inside the iframe but only extracted the report body, not citations/scanned/activity

Improvements

1. Enhanced iframe bridge (lines 100-149) - The bridge running inside the iframe now includes full duplicates of the citation, scanned source, and activity extraction logic (bridgeExtractCitations, bridgeExtractScanned, bridgeExtractActivity). When the parent sends a postMessage request with incCite/incScan/incAct flags, the bridge extracts all requested data from within the iframe DOM and sends it back.

2. Fixed exportViaIframe() (lines 473-491) - Now passes the user's citation/scanned/activity preferences to the iframe and receives the extracted data back in the response.

3. Fixed autoExportDR() (lines 504-542) - Now checks for the iframe first (maximized case), assembles report + citations + scanned + activity from the bridge response. Falls back to same-origin overlay extraction only if no iframe is found.

4. Fixed isDRElement() (lines 458-471) - Now recognizes the iframe element itself (by title or src attributes) and any element within the iframe's parent container as a DR element, so clicking it triggers autoExportDR().

5. Fixed element picker (lines 670-701) - findDRIframeAt() detects when the mouse is over the DR iframe and redirects highlighting to the iframe element. Both mousemove and mousedown handlers use this, so hovering over the maximized panel highlights the iframe and clicking it triggers the full Deep Research export.

6. Broader iframe detection - hasDRIframe() and getDRIframe() now also match by src*="web-sandbox.oaiusercontent.com" as a fallback, in case the title attribute changes.

Universal Markdown Exporter v2.2.0

Critical Bugs Fixed

1. extractDRCitations - Typo and broken selector

The v2.1.0 code had a fatal typo on line 242: urlUrl.getAttribute('href') instead of urlEl.getAttribute('href'). This would cause a crash during citation extraction attempts. Additionally, the selector strategy was restructured as follows:

• Citation groups (div.flex.flex-col.gap-0) are now queried first, falling back to direct button[aria-label^="Open source"] queries as necessary.
• Each citation button contains its own a[href] with the individual source URL.
• Confirmed via DOM API testing that React's createElement-based DOM allows nested buttons to work correctly (unlike HTML parser behavior in JSDOM).

2. extractDRScanned - Broken structure assumption

The v2.1.0 code assumed scanned items had individual URLs. Audit of the real HTML revealed:

• Scanned sources are grouped by domain (206 domain groups for 554 items)
• Each domain group has ONE a[href] at the domain level
• Individual scanned items have only titles and snippets, no individual URLs
• Rewrote to: iterate domain groups -> get domain href -> list each item's title linked to domain URL, with snippet as sub-item

3. extractDRActivity - Broken selectors and wrong entry type handling

• Title selector used '.text-token-text-primary.text-\\[14px\\]' -- the escaped bracket syntax doesn't work in querySelector. Changed to .text-token-text-primary (sufficient and correct).
• Body selector .text-token-text-secondary.mt-2 was wrong for "Searching" entries which use .mt-1 for their link container.
• The code assumed all "Searching" entries have links. Audit revealed two types:
  • Bare "Searching" (19 entries): have links in div.mt-1 > div.flex-wrap > a[href], no body text.
  • "Searching for X" (6 entries): have body text in .text-token-text-secondary.mt-2, no links.
• Rewrote to handle both cases correctly, with proper link extraction from the .mt-1 container.

4. extractDRDuration - Numbers were garbled from animated spans

The duration div uses animated CSS number roller spans (span[role="img"]). Calling .textContent on these returns all digits 0-9 concatenated. Fixed to use aria-label attributes which contain the clean numeric values (verified: aria-label="n" for citations, aria-label="n" for searches).

Element Picker Fixes

1. mousemove handler - elementFromPoint accuracy

The v2.1.0 code hid only the tip element before calling elementFromPoint. But the .h2m-sel outline on the currently selected element could also interfere with hit-testing. Now temporarily removes the .h2m-sel class from the current element before elementFromPoint and restores it after.

2. mousedown handler - DR-aware click behavior

Previously, clicking any element inside the DR overlay would just convert that single element to markdown (often a single paragraph). Now detects if the clicked element is inside a DR overlay via isDRElement() and automatically triggers the full autoExportDR() flow with all panel extractors.

Enhancements

1. Gemini extraction - Added extractGemini() function for gemini.google.com conversations and canvases, with G keyboard shortcut.

2. Obsidian export - Added purple "Obsidian" button to the modal toolbar with obsidian://advanced-uri integration, plus a configuration dialog for vault name and folder.

3. stripUtm helper - Centralized UTM parameter stripping to avoid repetition across extractors.

Output Format Alignment

The extractors now produce markdown matching the following template:

Citations [x Sources]

1. citation 1 ...

Scanned [y Sources]

1. source 1
   1. content 1 if present ...

Thinking Activity [Research completed in z]

1. thought subtitle 1**
   1. thought summary 1 ...

zx. Searching

1. citation title 1
2. citation title 2
3. citation title 3 ... zx. citation title zx ...

Universal Markdown Exporter v2.1.0

1. Element picker now works INSIDE the maximized Deep Research overlay

The root cause was that ChatGPT's React synthetic event system calls stopPropagation() on pointer events within the overlay, preventing pointermove from reaching document-level listeners.

Fix: Replaced pointermove with native mousemove + document.elementFromPoint(x, y). This bypasses React's event interception entirely -- elementFromPoint queries the rendering tree directly regardless of what React does with events. The tip overlay is temporarily hidden during the query so it does not intercept the hit test. Combined with mousedown (instead of click) with stopImmediatePropagation(), this ensures the export fires before React's handlers can mask the event.

You can now hover over individual paragraphs, headings, list items, and citations inside the maximized deep research panel, and click to export just that element.

2. Deep Research panel extractors (Citations, Sources Scanned, Thinking Activity)

Three new dedicated extractors parse the same-origin DOM panels directly:

• extractDRCitations() -- Parses section[aria-labelledby="report-references-citations"], extracts each numbered citation with title and URL into the format 1. [title](url)
• extractDRScanned() -- Parses section[aria-labelledby="report-references-sources-scanned"], extracts grouped sources with snippets
• extractDRActivity() -- Parses section[aria-labelledby="report-activity-title"], extracts thinking steps with titles, summaries, and "Searching" entries with their source links

3. Toggleable export options in the modal toolbar

The preview modal now has checkboxes for Citations, Scanned, and Activity that persist via GM_setValue. When pressing R or using the menu command, the export automatically appends whichever panels are toggled on.

Universal Markdown Exporter v2.0.0 -- Trusted-Types Safe Rewrite

Architecture: 100% Trusted-Types Safe

The script was rewritten from scratch with zero innerHTML, zero jQuery, zero Turndown, and zero insertAdjacentHTML. Every UI element is constructed via the DOM API using a compact el() helper function. This means it can function normally, even on strict Content Security Policy sites like gemini.google.com without requiring a separate "Trusted Types safe" variant.

Removed Dependencies:

1. jQuery (3 files: jquery, jqueryui, turndown-plugin-gfm)
2. Turndown.js
3. All $(), .html(), .append($(...)) patterns

Retained Dependency:

• marked.js -- however, the preview renderer now uses DOMParser().parseFromString() + document.adoptNode() instead of innerHTML, in full compliance with Trusted Types Content Security Policies.

Universal Element Selection

The element picker (Ctrl+M or Tampermonkey menu) uses native pointermove/click/wheel/keydown event listeners with the capture: true phase, enabling versatile and consistent performance across diverse web pages.

• Hover highlights elements with a red dashed outline
• Click exports the highlighted element to markdown
• Arrow keys navigate parent/child/sibling
• Scroll wheel zooms parent/child
• R key auto-detects and exports Deep Research or Gemini content
• Esc cancels

When clicking an element that contains or IS a Deep Research iframe, the script automatically routes to the postMessage bridge pathway instead of trying to convert the opaque iframe container.

Deep Research Support

• Live iframe bridge: Inside web-sandbox.oaiusercontent.com, the script runs as a bridge listener, extracting content via the custom converter and passing it back to the parent via postMessage
• React fiber traversal: Recovers citation URLs from __reactFiber nodes (props.item.reference.safe_urls)
• Srcdoc parsing: Handles saved/exported HTML pages with nested entity-encoded srcdoc content

Gemini Support

Detects gemini.google.com, extracts conversation turns (message-content elements, [data-message-id]), canvas content, and routes through the same markdown converter.

Preview Modal

Built entirely from DOM nodes -- dark theme, split-pane with editable textarea on the left and live rendered preview on the right, synced scrolling, toolbar with Copy/Download/GitHub Issue/Obsidian export, citation style selector, frontmatter toggle, and title-as-H1 toggle.