Roll back to last green
Tests passed, then a later step broke them. Return to the last checkpoint annotated tests=pass, without a WIP commit.
Early preview · v0.1.0-rc.1 · being tested in Cloop and Automaze Works
High-frequency source control for AI agents.
ghf saves your worktree every time an agent finishes an edit. Restore, diff, apply or fork any of those states, on any machine. Git never knows it was there.
$ git hf trace log --annotation tests=pass
#7 ck_01m5054dg9rkmmeasv0rcxd8x0 2026-10-15T16:08:12Z 2 change(s) claude-code:s1 [tests=pass]
$ git hf trace restore ck_01m5054dg9rk --in-place
restored ck_01m5054dg9rkmmeasv0rcxd8x0 in place (2 written, 1 deleted); recorded as ck_01m5054dmjth7gspksqcfddsh0Git
holds accepted history: commits, branches, pull requests.
ghf
holds recoverable working history, with its own retention, permissions and sync.
Nothing else changes
No commits, trailers, notes or refs. Plain Git users never need to know ghf exists.
What it does for an agent fleet
Hand off between machines
Another worker continues from the exact file state, uncommitted changes included.
Fork an attempt
Two models try different approaches from the same checkpoint, each in its own worktree and trace.
Show a human what was tried
When a run escalates, list the trace and diff any two points.
Survive a crash
A checkpoint is acknowledged only once it is durable locally. Sync catches up when the network returns.
How it works
git hf after every edit. It writes checkpoints to its own store under .git/ghf and syncs them to ghf-server, from which any other machine can restore. Git itself is only read.Every checkpoint is complete
A checkpoint is the full worktree state, stored as content-addressed objects (BLAKE3, zstd). Each one restores on its own, and unchanged directories are shared.
Restore never eats work
Restore goes into a new worktree by default. In-place restore and apply refuse if anything uncaptured would be lost, and every restore is itself a checkpoint.
One writer per worktree
Parallel agents use separate Git worktrees, each with its own trace. A second writer gets E_WRITER_LOCKED.
Its own lifecycle
Traces close, finalize against the commit that landed, and are pruned 30 days later. Pins and holds keep what matters. Pruning never touches Git.
Built for agents to call
Every command takes --json and returns one versioned envelope with stable error and exit codes. Nothing prompts in automation, and mutating commands accept an idempotency key.
{"schema":"ghf/v1","command":"trace apply","ok":false,
"error":{"code":"E_APPLY_CONFLICT","message":"1 conflicting path(s); the working tree is unchanged"},
"data":{"applied":false,"conflicts":[{"path":"src/lib.rs","reason":"content changed on both sides"}]}}Quick start
Run a server
ghf-server --data-dir /var/lib/ghf token create --name fleet --repos 'acme/*' ghf-server --data-dir /var/lib/ghf serve --listen 0.0.0.0:7878 \ --tls-cert cert.pem --tls-key key.pemEnable a repository
git hf init --server https://ghf.example.com --repository acme/project git add .ghf && git commit -m "Enable ghf"Set up each machine
git hf trust echo "$GHF_TOKEN" | git hf login git hf hooks install claude-code # or: cursor git hf skill install
That's it: every agent edit is now a checkpoint. The user guide covers the rest.
Fast enough to run on every edit
| 2,000-file repository, release build | p50 | p95 |
|---|---|---|
| Checkpoint after editing files | 25 ms | 28 ms |
| Checkpoint after creating a file | 42 ms | 45–51 ms |
| Restore on another machine | 0.6 s | |
Early preview. ghf is currently being tested in Cloop and Automaze Works. Anything can change before v1 is officially released, including commands, the JSON schema, the server API and the storage format.