opencode-session-transcript

Fixes /export's truncation bug. Writes Markdown for reading, not JSON for re-import — it complements /export rather than replacing it.

This is a stopgap for bugs in /export (#50732, #50014) that are being fixed upstream. Once those land, this should be retired in favour of a correct /export.

Not a backup. /export writes JSON, which opencode import can read back. This writes Markdown, which cannot be re-imported, and it drops assistant reasoning to do it. If you want a restorable archive, keep using /export.

The problem

OpenCode can export a session, but /export has three separate shortcomings:

It silently truncates. #50732 and #50014 both report that /export writes only the messages currently loaded in the TUI, not the full transcript. In a long session, early prompts and responses are simply absent — and nothing warns you, so the file looks complete. This is the worst possible failure mode for an export tool, because you only find out when you go looking for something that isn't there.

It means different things in different clients. #48105 documents that /export in the TUI produces human-readable Markdown while opencode export on the CLI produces JSON for re-import. The desktop app's /export follows the JSON behaviour, so there is no Markdown export there at all.

There is no CLI equivalent that reads well. opencode session export gives you a multi-megabyte JSON file — correct and complete, but not something you would ever read end to end.

What this does

One command. It reads every message for the current session straight out of OpenCode's SQLite store and writes them to a Markdown file you can actually read.

/save

/home/steven/Documents/archiving-older-ai-conversations.md

835 messages · 234 KB

The output is the same conversation you'd expect to read, with tool calls kept inline and assistant reasoning dropped:

# Archiving older AI conversations

**Started:** 2026-10-13 11:43
**Directory:** `/home/steven/Documents/Default Project`
**Session id:** `ses_ec96bff81ffdMzlHLBZmD0r2Nj`

### You · 11:43
can I archive older AI conversations to free up session space...

### Assistant · 11:43
I'll check the OpenCode docs on sessions and CLI settings.

**[tool]** `webfetch` — completed

Because it reads the database rather than the UI's loaded window, the output is complete regardless of how much of the thread you have scrolled.

Install

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": ["opencode-session-transcript"]
}

Or point at a local checkout:

{ "plugins": ["/absolute/path/to/opencode-session-transcript"] }

No dependencies are installed at runtime. The plugin uses whichever SQLite driver the host provides (bun:sqlite under Bun, node:sqlite under Node 22+).

Usage

Command	Effect
/save	Write the current session to ~/Documents/<session title>.md
/save ~/Desktop/today.md	Write wherever you say
/unarchive	List archived sessions
/unarchive <n | id | title>	Restore one

Set OPENCODE_TRANSCRIPT_DIR to change the default folder.

Filenames come from the session's own title, read at run time. Two sessions sharing a title will overwrite each other — pass an explicit path to keep both.

About /unarchive

This plugin started as a session-archiving tool. /archive and /unarchive were built to declutter the session list, then dropped when the project view — the 4-squares toggle in the top-left of the desktop app — turned out to cover that already. What remained was wanting a session out as readable text, which is what /save does.

Restoring stayed because it is the one half with no equivalent in the UI: OpenCode can archive a session but offers no way to undo it.

Planned deprecation: /unarchive will be removed in a future major version. Restoring an archived session belongs in OpenCode core, not in a stopgap for /export, and keeping it here muddies what this package is for. Expect it to move to a separate plugin rather than disappear.

There is deliberately no matching /archive. Archiving hides a session without removing it, and the project view already makes everything findable again, so on offer it solves nothing.

An ambiguous term restores nothing rather than guessing: a title fragment that matches two sessions will ask you to be more specific rather than pick one.

Safety

• /save is read-only. It only reads session_v2 and session_message and writes one Markdown file of its own.
• /unarchive does write, and takes a VACUUM INTO snapshot first, which is consistent while OpenCode holds the database open in WAL mode. It only ever sets time_archived and time_updated — never a message. It also keeps the three most recent snapshots and prunes older ones.
• Schema-agnostic: discovers session_v2 (v2.x) or session (v1.x) and introspects its columns rather than hardcoding either.

Development

npm test          # node --test, no install required
npm run check     # confirms both modules load
npm run typecheck # requires typescript

The test suite builds throwaway databases in both schemas and runs the shared suite against each, so v1 and v2 support is exercised rather than assumed.

Upstream

Once /export is correct — desktop parity plus the truncation fix — this plugin should be retired. The goal is for /export to work properly, not for this to become infrastructure anyone depends on.

Licence

MIT