Go back to any turn
Every turn has a name, like 5b0e2c1a.12. braid log lists them with what each did, and /braid:fork .12 gives you the command to continue from that turn in a new window.
v0.4 · for Claude Code and Codex
braid keeps each project’s agent conversations in its own .braid folder, as plain files you can read and search. Go back to any turn, and continue it in Claude Code or Codex.
/plugin marketplace add dvnclabs/braid
/plugin install braid@braid
/braid:init
Needs Node 22 or later, or bun, and git. Without Claude Code, clone the repository and run sh bin/braid init.
Every turn has a name, like 5b0e2c1a.12. braid log lists them with what each did, and /braid:fork .12 gives you the command to continue from that turn in a new window.
A saved session continues in either agent. braid fork --to codex writes it as a Codex session, and --to claude-code goes the other way. Each agent gets the conversation the other's model read.
Sessions live in .braid/sessions/, one compressed file each, in a folder git does not see. Every worktree of a project saves into the main checkout’s folder, so deleting a worktree keeps its sessions.
Each file holds a whole session. A fork, or a copy made with /branch, is saved as a session of its own with the history it came from, ready to read, search or train on.
braid verify checks a session the way git fsck checks a repository, and runs by itself when each session ends: every record is in the store, and a fork written from it holds what the agent reads to resume.
Nothing leaves your machine, and braid calls no model of its own. Sessions are ATXF, an open format for agent transcripts, ready for a web interface or a training pipeline to read.
$ braid verify --all
braid verify 9e36aa15.41: ok
✓ saved: 1211 records of the transcript
✓ loads: 53 lines since the last compaction,
the same in a fork as in the transcript
✓ file: 1211 records, each with its parent