braids

README

braids

Branch your Claude Code conversations. Resume any of them from any turn.

braids.chat  ·  Docs  ·  Install

---

Claude Code works best on one linear thread doing one thing. Humans don't work that way. A task forks, doubles back, and spawns three side quests, and the only way to cope today is to run N terminals and hold the map in your head.

braids is that map. One terminal window alongside your N others. It shows every conversation and every branch as a graph, searches all of them in about a millisecond, and turns any message into the start of a new branch.

It never talks to Claude. It arranges the conversations you have with Claude.

_{A real frame, colours and all. Every screenshot here is captured by scripts/demo.py against a fake ~/.claude, never drawn by hand.}

Why

A conversation with an agent is a single line of context. That is a good fit for the model and a bad fit for the work: you want to try an idea without poisoning the thread, come back to a decision made an hour ago and go the other way, or run four attempts at once and keep the one that worked.

braids keeps the graph for you, and gives the model a clean linear path. Every root-to-leaf route through the graph is one ordinary Claude Code session: no wrapper, no proxy, no protocol of its own.

Install

curl -fsSL https://braids.chat/install.sh | sh

Run it again to update: it asks the binary it would replace what version it is, says so and stops when there is nothing to do, and replaces braids where it already lives rather than leaving a second copy elsewhere on your PATH.

That works out the release build for your machine, checks it against the published checksums, and puts one binary on your PATH. Nothing else: no daemon, no configuration file, and nothing to uninstall but the file.

Or with Go:

go install github.com/Ashes47/braids/cmd/braids@latest

Or from source:

git clone https://github.com/Ashes47/braids && cd braids
make install          # -> $(go env GOPATH)/bin/braids

First run

braids index          # read every transcript under ~/.claude (a few seconds)
braids                # open the map

Then, on the map:

j k	move
↵	open a conversation's spine: its landmarks, not every line
b	branch from the turn under the cursor
y	copy the claude --resume command for it
/	search every conversation, memory and work product
w	browse what a session wrote to disk
M	what the project remembers

Nothing here is a command you have to remember: the header lists every binding the screen has, and a mistyped one names the key you meant.

Optionally, braids hooks --install asks Claude Code to report when a session is blocked on you, which is the one thing the files cannot say. It is opt-in, merges with whatever hooks you already have, and --remove takes back only what braids added.

What it does

Map. Every conversation as a tree, with branches shown under the conversation they were cut from, and how far behind each one is.

Spine. One conversation collapsed to its landmarks: what you asked, what came back, where it failed, where it was compacted. A 25,000-turn conversation becomes a few hundred lines you can actually read.

Search. Full text over every message and tool call, plus your memories and the names of every work product, in a few milliseconds. Each result says whether it is a conversation, a memory or a work product, and ↵ opens the screen that can act on it. Search is the front door; the graph is the confirmation.

Branch. Put the cursor on any turn and press b. braids writes a new session file containing that turn's ancestry, and hands you the claude --resume command. Add --workspace and the branch gets a git worktree of its own, so two branches can edit the same file without touching each other.

Merge. Join a branch back as a new conversation, splicing the real turns from both. braids refuses when one side already contains the other, rather than producing a duplicate wearing a new name.

Promote. A subagent's transcript is a conversation too. Turn one into a lane of its own and carry on from where it left off.

Work products. A session's scratch usually dwarfs its transcript: 3.3 GB against 363 MB here, with nine files holding 1.3 GB of it. w opens it as a size browser: heaviest first, directories weighed by what is under them, ↵ to descend. So you can bin one 231 MB dump and keep the rest. ↵ on a file shows its head, y copies its path, and a binary is named rather than spewed.

The harness's own record of a job is shown and refused. braids work --orphans finds the sets whose conversation is gone.

Memories. What a project has told the harness to remember, with the two things you cannot see from inside a session: a memory the index omits, which is therefore loaded by nothing, and a link pointing at a memory that is gone. ↵ reads it as markdown, c opens the conversation that wrote it.

d deletes one to the bin, r renames it and follows the name through every link, and i repairs the index. Each of them changes the index in the same breath as the file, and is refused while a session in that project is running.

Bin. Deleting moves files aside with a manifest and a 14-day retention, so nothing you delete is gone the moment you regret it.

Driving it from an agent

Every command that reports something takes --json, so braids is usable by the thing it is watching. Claude Code can search its own past conversations, find where a decision was made, and branch from that exact turn.

braids search "why did we drop the retry" --json | jq -r '.hits[0] | .lane, .turn'
braids branch --lane <id> --at <turn> --json | jq -r .resume

Two properties make this work, and both are deliberate:

• IDs are whole in JSON. The tables shorten them with an ellipsis to fit a terminal; --json never does. (Pasting a shortened one back works too, since braids accepts the ellipsis rather than refusing an ID it printed itself.)
• Empty is [], not a sentence. Nothing should have to tell "no matches" apart from a parse failure.

Commands

braids                                  open the map
braids index [--full]                   index new and changed transcripts
braids search QUERY [--type T] [--kind K] [--limit N]
braids lanes                            list indexed conversations
braids agents  --lane ID                list the subagents a conversation spawned
braids work    --lane ID [--path SUB]   browse a session's work products
braids work    --orphans [--reclaim]    find, and reclaim, ownerless ones
braids memories [--project NAME]        what a project remembers, and whether
                                        the index still agrees with the files
braids branch  --lane ID --at TURN [--workspace]
braids promote --lane ID --agent ID     turn a subagent into its own conversation
braids merge   --lane ID --from ID [--plan]
braids hooks [--install|--remove]       let sessions report when they block
braids version

Every command takes --help for its own flags and --json if it reports something. A mistyped command names the one you meant.

Hooks are optional

Files can tell you a session is mid-tool-call. They cannot tell you whether it is running or waiting for your approval. Both look identical on disk. One hook can. braids hooks --install asks Claude Code to report it.

It is opt-in and separate from installing the binary: a tool that edits your settings file as a side effect of being installed is not one to trust with it. Installing merges rather than writes: every hook already there is kept, a timestamped copy of the previous file is left beside it, --remove takes back only what braids added, and a settings file that cannot be parsed is refused rather than replaced by a guess.

Everything else works without them. braids hooks says which mode you are in, and so does the map.

Privacy

braids reads your transcripts, so this matters more than usual:

• It makes no network calls. There is no HTTP client and no listener in it. go list -deps ./cmd/braids | grep net/http comes back empty. Check it yourself.
• Nothing leaves your machine. The index is a local SQLite file at ~/.braids/index.db.
• What it writes stays private to you. ~/.braids is 0700 and the index is 0600, because it holds the full text of every message it has read and Claude Code keeps the transcripts that came from at 0700. An index left loose by an older build is tightened the next time braids opens it.
• It never writes to a transcript Claude Code owns. Branching writes a new session file; the source is opened read-only. One writer per file, always.
• Delete braids and you lose a view, never a conversation. Everything is derived from ~/.claude/.

Numbers

Measured on one laptop against a real corpus of 28 conversations, ~62,000 messages, ~360 MB of transcripts. Rounded, because a corpus you are still working in never sits still:

First index, from nothing	~8 s
Re-index, a turn added to a 145 MB conversation	~40 ms
Re-index, nothing changed	under 1 ms
Search, across ~62,000 indexed units	1–6 ms
Index on disk	~190 MB

The last row is the honest cost: the index is roughly half the size of the transcripts it was built from, because it stores the text again to search it.

Principles

1. Files are the truth. Everything is derived from ~/.claude/.
2. One writer per file, always. Every branch is its own session file.
3. The graph is for humans; each path is linear for the model.
4. Search is the front door; the graph is the confirmation.
5. A window you glance at, not a place you live.

Non-goals

braids does not talk to any model, proxy your session, or replace your terminal. It does not sync anything anywhere. It is one window next to the ones you already have open.

Uninstall

braids hooks --remove                     # if you installed them
rm -rf ~/.braids                          # the index, the bin, the sidecars
rm "$(go env GOPATH)/bin/braids"          # the binary

Your conversations are untouched by all three. They were never braids' to begin with.

Security

The threat model, the permissions braids sets and how to report something are in SECURITY.md. The one worth repeating here: BRAIDS_SPAWN runs through a shell, so every value braids substitutes into it is shell-quoted first, because a conversation called x; rm -rf ~ # has to be a name rather than a command. Your template supplies the shell syntax; the values never do, so do not put quotes around a placeholder yourself.

Other harnesses

Claude Code is the only source today. The seam is deliberate: a Source port with optional capability interfaces, so a harness that cannot branch still gets the map and search. See SPEC.md.

Contributing

make ci     # fmt, vet, lint, test, race, cover: the same gate CI runs

CONTRIBUTING.md has the layout, the decisions that look like missing features, and what a good change looks like. SPEC.md is the design doc: nearly every decision in it traces to something measured on a real machine rather than assumed, and the ones that went wrong went wrong because they were assumed.

License

MIT