import-export-archive

name	import-export-archive
description	Import conversations from an official ChatGPT data export zip or conversations.json. Use for bulk import, for old or deleted conversations, or when the browser route is unavailable. Also the only route that carries generated images as real files.

Import from an account export

The export comes from ChatGPT Settings → Data controls → Export data, and arrives by email as a zip. conversations.json at its root is a flat array of conversations in the same shape the backend API returns.

Use this route for bulk work, for anything no longer in the sidebar, and when images matter — the zip bundles them as real files, which the browser route cannot retrieve.

List before extracting

An export holds every conversation the account has ever had. Do not extract all of them because a script can:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/split_export_archive.py" export.zip --list

1483 of 1483 conversation(s)

[   0] 2026-08-24   17u/17 a  Quarterly planning notes
[   1] 2026-08-21    4u/5  a  Sample onboarding conversation

Narrow with --match "<regex over title>", --since YYYY-MM-DD, --id <uuid> or --index N. Add --json for machine-readable output.

Conversation titles are themselves personal. When the user is looking for one thread, filter to it rather than printing the whole listing into the transcript.

Extract, then normalize

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/split_export_archive.py" export.zip \
  --match "onboarding" --out-dir ./raw

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/normalize_conversation.py" \
  ./raw/2026-08-21-sample-onboarding-conversation.json \
  --source-kind export_archive -o ./normalized.json

Files are named <date>-<title-slug>.json. normalize_conversation.py also takes the raw array directly with --index N, which saves a step for a one-off.

Then render as usual — render-markdown, render-typst-pdf, or conversation-to-context.

Differences from the browser route

• current_node is sometimes absent or points at a node that is not in the export. The normalizer falls back to the newest leaf, which is nearly always the same branch; on a heavily edited thread it may not be. Check stats.branch_points.
• Model slugs are sparser on older conversations, so models_used can be short or empty for anything predating that field.
• Images referenced by asset pointers exist as files elsewhere in the zip. Matching them up by pointer id is not automated here; unzip and look for the file id if the user needs them.

Bulk import

For many conversations, loop over the extracted files. Two things to hold onto: write outputs somewhere git-ignored, and do not print titles in bulk. If the user wants hundreds imported into a repo, confirm the destination is ignored before starting rather than after.