Inspect local PDF links, comments and reading notes

Choose the structure that answers the task

An unresponsive clickable PDF area needs its actual Link declaration. A reviewed report needs original annotation text and reply context. A device log needs complete reading records. These are three different inputs and deliveries; visible URLs, page text and line dedup cannot substitute for the underlying structures.

Read a target without opening it

T294 scans real Link annotations. It resolves direct page targets, legacy Dests and the full Names/Dests tree using the original key bytes, declared Limits and unique page references. URI, GoTo, GoToR, Launch, Named and chained Next actions have explicit states. Other actions retain their type and declaration as unsupported. No action, URI or embedded script is executed, opened or fetched.

The checked three-page source contains 11 Links and two named destinations. Some resolve to pages 1–3, one has a missing target, a missing name remains dangling, and a JavaScript action stays an unsupported declaration. The downloaded CSV quotes full structured targets; a 2,000-character visible cell is only a preview.

• A written URI does not establish safety, network reachability or the reason a viewer ignored it.
• The same source can contain text URLs that are not clickable annotations; they do not increase Link count.
• Page rectangles are raw PDF points, not rotated or cropped screen coordinates.
• Malformed URI types, contradictory ownership, duplicate raw names, invalid NameTree ordering/Limits and cycles reject the whole inventory.

Keep review text and thread meaning

T295 exports the original plain Contents, author, original M/CreationDate strings, raw Rect/QuadPoints and IRT reply relationships. It does not reinterpret date timezones. Missing Contents is JSON null rather than invented text; RC presence is flagged and rich-only text is explicitly unsupported. The Markdown uses literal fenced blocks even when source text contains HTML or backticks.

The checked source has 17 supported records, 11 Links, one Widget and one Popup. It retains a multiline reply, a rich-only record and a dangling reply. Link, Widget and Popup have separately counted excluded roles. The complete source includes every supported subtype, including Redact; exporting a Redact annotation does not apply a redaction.

• Supported review types: Text, FreeText, Highlight, Underline, StrikeOut, Squiggly, Ink, Stamp, Caret, Line, Square, Circle, Polygon, PolyLine and Redact.
• Any other subtype, reused annotation identity, duplicate NM, conflicting P owner, malformed geometry or reply cycle fails the entire export.
• An absent or external reply object remains flagged as dangling; the tool does not fabricate its parent.
• This reads existing plain annotation text. It cannot reconstruct flattened review marks, recover highlighted sentences by OCR or execute rich HTML.

Organize whole clipping records instead of lines

T296 accepts an already copied UTF8 MyClippings.txt with English Highlight, Note and Bookmark metadata. A complete block contains the exact source book title, explicit positive page/location ranges, raw Added on text, body and a separate ========== line. That line is a reserved delimiter. BOM and CRLF/LF are reported; internal body line endings and Unicode remain unchanged.

The six-record BOM/CRLF example retains five records after exact dedup. An empty Bookmark is valid. The two excerpts at locations 40–45 and 42–47 both survive. Whole-record equality includes title, metadata, body and line endings; same location with different content, dates or whitespace does not overwrite anything. Original indexes map to retained IDs. Exact title strings define books without guessing an author split.

Download complete originals and inspect failures

The table is bounded to the first 200 records and 2,000 UTF16 characters per cell. Native copy contains a bounded processing summary. Downloads retain the full JSON records, all quoted CSV rows and literal Markdown bodies. CSV prefixes formula-like values with an apostrophe; JSON and Markdown preserve the originals. Dates remain raw source strings and no timezone is inferred.

The production checks use real 500-page sources with 10,000 Links and 10,000 review records, a separate 100,000-record clipping log, and an exact 10MiB single-note input. An independent pypdf/stdlib reader checks every source record, reference, target, reply, CSV cell and Markdown body. Real core progress cancellation followed by the same input recovers the full delivery; browser interaction acceptance is separately maintained by the controller.

• Each PDF: 30MiB, 500 pages, 200,000 parsed objects, depth 64, 10,000 annotations, 16MiB extracted strings and 4,096 bytes per PDF name. T294 adds 10,000 destinations and 1,000 NameTree nodes. JSON/CSV for links or JSON/CSV/Markdown for reviews total at most 20MiB. Object/XRef streams total 80MiB decoded; plain/Flate/LZW/ASCII85/ASCIIHex/RunLength, up to four filters, Predictor=1 only (LZW EarlyChange=0/1). Both cases of legal #hex names are parsed without dropping records.
• Clippings: one 10MiB UTF8 log and 100,000 complete records; three complete outputs total at most 40MiB. The source and output limits both apply; a short but highly repetitive metadata structure may still exceed the output budget.
• Empty logs and PDFs without matching records yield valid empty lists. Unsupported locale/type, bad UTF8, reversed/repeated ranges and missing delimiters reject clippings with record position.
• Failure or cancellation produces no partial success report. Correct the indicated source structure and retry; originals stay unchanged.
• No device/account access, URL requests, cloud sync, Kindle reimport, missing-log recovery, source-link repair or comment removal is performed.

Tools used in this article