Status
What Gin does today, what is being built, and what has been measured. This page reflects the Gin repository's main branch on 7 November 2026.
Gin is pre-release. The version is 0.1.0, there are no published binaries or compatibility promises yet, and you build it from source.
On main today
These ship in the source on main and are covered by CI on Linux.
| Area | What works |
|---|---|
| Workspaces | Private workspaces with journals, conditional saves, savepoints and local checkpoints |
| Publication | Publishing exact checkpoints on channels, whole or not at all |
| Dependent work | Forking from published parents, preview and advance (restack), multi-parent composition |
| Conflicts | Durable candidates, conflict paging, side reads and explicit resolutions |
| Concurrency | Many workspaces saving into one native store; process-kill recovery tests. The concurrent store is not yet release-qualified |
| Durability | Git-parity defaults: no extra device flushes. full durability is opt-in (details) |
| Adapters | Embedded Rust API, gin CLI, gin mcp, gin serve (gRPC with TLS, local IPC) |
| Evidence | CI uploads, execution receipts, baseline classification, advisory reuse plans. The test-result parts are planned to move to Toki |
| Comments | Append-only annotations on exact checkpoints |
| Git | Bundle import endpoint, history on request, checkpoint export to a Gin-owned repository (details) |
| Views | gin view: materialised checkouts of exact snapshots for build systems |
| Maintenance | gin verify, gin repair, gin retention |
| LFS | The gin-lfs library only. No gin command uses it yet |
Being built
This work is in progress and is not on main. Names and behaviour may change before it lands.
| Work | What it adds |
|---|---|
| Git sync redesign | Import and export as one plain tree diff between Git's head and Gin's head (design) |
| Tama provider | Gin as the source backend for Tama tasks |
| Agent-first CLI and MCP | Lifecycle commands for agents, a gin guide command that teaches an agent how to use Gin, and a read-only git stand-in that answers common Git commands inside Gin checkouts |
| DAG qualification | The remaining measurements for agents on a task DAG, such as publication latency and kill safety under both durability modes |
| Git LFS | LFS content in workspaces, checkouts and Git sync |
| Submodules | Git submodules in Gin trees |
| Windows | Native Windows stores and checkouts |
Test history belongs to Toki. Gin's CI test ingestion is planned to move there; Gin will then keep at most a link to Toki's history for a checkpoint.
Platforms
| Platform | Status on main |
|---|---|
| Linux | Primary. CI builds, lints and tests every change on Linux. |
| macOS | Native stores work (Unix). Not tested in CI; CI only cross-compiles an Apple Silicon binary. Benchmarks ran on an M2 Mac mini. |
| Windows | CI cross-compiles and links a Windows binary. Native stores, journals and local IPC are refused. Use WSL, or use the CLI as a client of a remote gin serve (not covered by CI). |
See Getting started for the practical consequences.
Measurements
Gin quotes only numbers recorded in its repository. The headline comparison is docs/benchmarks/dag-agents-2026-11-07/README.md, with raw JSON beside it and the method in docs/dag-benchmark.md.
Workload. A 32-task, 51-edge, dependency-closed prefix of a real project's task DAG, including a five-parent join. Scripted agents edit 5 to 50 real Rust files per task with a fixed amount of CPU work. Children start from their parents' published contracts while the parents continue. All strategies must produce the same final tree, checked against an independent oracle.
Strategies. Gin's embedded native store; Git with physical worktrees and merge updates; Git with worktrees and rebase updates; and Tama's current flow, where a task starts only after its parents land.
| Host | Gin | Best Git | Tama today | Gin vs Git | Gin vs Tama | Restack p50 vs git rebase |
|---|---|---|---|---|---|---|
| Apple M2 Mac mini | 12.65 s | 78.93 s (rebase) | 59.90 s | 6.24x | 4.74x | 13.38x |
| Linux container, 12 CPUs | 20.58 s | 74.18 s (merge) | 77.93 s | 3.61x | 3.79x | 15.05x |
All eight final trees matched, with zero failures or refusals. The 3x target against Git and the 5x restack target pass on both hosts. The 5x target against Tama fails on both and remains open.
What these numbers do not show:
- They are one warm run per strategy, not production tail latencies.
- No language models and no builds ran; agents were scripts.
- Gin was driven through its Rust API, not the CLI, MCP or daemon.
- The Tama provider, its checkout projection and the Git sync redesign are not merged and were not measured.
- Default versus
fulldurability and a complete multi-process kill matrix still need qualification.
Other recorded measurements, such as the concurrent store's throughput under both durability modes, are in docs/concurrency-round-2.md.