claw-branch

claw-branch is the branch orchestration engine for ClawDB-style workflows.

It provides isolated, SQLite-backed branch execution for agent and application workloads where you need to fork a trusted baseline, run experiments safely, inspect exact diffs, and merge with explicit conflict policy.

Why this crate exists

Traditional branch workflows in application databases are hard to make deterministic, inspectable, and safe for concurrent agents. claw-branch solves this with:

• File-level branch isolation (one SQLite DB file per branch).
• Explicit lineage and merge-base tracking via DAG.
• Three-way merge and selective commit primitives.
• Sandboxed simulation with recommendation output.
• Built-in metrics and divergence reporting.
• Snapshot verification and garbage-collection tooling.

Installation

[dependencies]
claw-branch = "0.1.1"
tokio = { version = "1", features = ["macros", "rt-multi-thread"] }

Optional guarded mode (integrates with claw-guard policy checks):

[dependencies]
claw-branch = { version = "0.1.1", features = ["guarded"] }

Feature flags

Architecture overview

┌─────────────────────────────────────────────────────────────────┐
│                         BranchEngine                           │
│   ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌─────────────┐    │
│   │ Lifecycle│  │  Commit  │  │  Merge   │  │   Sandbox   │    │
│   │ fork/    │  │ cherry-  │  │ 3-way +  │  │ run/eval    │    │
│   │ archive  │  │ pick/all │  │ resolver │  │ recommendation│   │
│   └────┬─────┘  └────┬─────┘  └────┬─────┘  └──────┬──────┘    │
│        │              │              │                │           │
│   ┌────▼──────────────▼──────────────▼────────────────▼──────┐   │
│   │                  BranchStore (registry DB)                │   │
│   └───────────────────────────────────────────────────────────┘   │
│                                                                   │
│   ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌─────────────┐      │
│   │   DAG    │  │   Diff   │  │ Metrics  │  │  Snapshot   │      │
│   │ lineage  │  │ entity + │  │ tracker/ │  │ copy/verify │      │
│   │ + LCA    │  │ field    │  │ reporter │  │ + gc        │      │
│   └──────────┘  └──────────┘  └──────────┘  └─────────────┘      │
└─────────────────────────────────────────────────────────────────┘

Each branch is a separate SQLite file.
The trunk is the canonical root branch and is never discarded.

Quick start

use claw_branch::prelude::*;
use std::path::PathBuf;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let config = BranchConfig::builder()
        .workspace_id(uuid::Uuid::new_v4())
        .branches_dir(PathBuf::from("/tmp/demo/branches"))
        .build()?;

    let engine = BranchEngine::new(config, std::path::Path::new("/data/source.db")).await?;
    let feature = engine.fork_trunk("feature/summariser").await?;

    let trunk = engine.trunk().await?;
    let diff = engine.diff(trunk.id, feature.id).await?;
    println!(
        "changed: +{} ~{} -{}",
        diff.stats.added, diff.stats.modified, diff.stats.removed
    );

    let report = engine
        .simulate(
            feature.id,
            SimulationScenario {
                name: "trial-run".into(),
                description: "evaluate summarisation strategy".into(),
                max_ops: Some(100),
                timeout_secs: Some(30),
                seed_data: None,
            },
            |pool| async move {
                let _ = pool;
                Ok(serde_json::json!({"status": "done", "summaries_generated": 42}))
            },
        )
        .await?;

    match report.recommendation {
        Recommendation::Commit => engine.commit_to_trunk(feature.id).await?,
        Recommendation::Discard => engine.discard(feature.id).await?,
        Recommendation::NeedsReview(notes) => {
            println!("manual review required: {notes:?}");
        }
    }

    Ok(())
}

Core concepts

Public module map

BranchConfig reference

Merge strategies

Simulation lifecycle

1. Fork a temporary simulation branch from a parent.
2. Run agent logic against the simulation SQLite pool.
3. Compute diff and metrics.
4. Produce Recommendation::Commit, Recommendation::Discard, or Recommendation::NeedsReview.
5. Apply workflow policy (promote/discard/review).

Error model

Most operations return Result<T, claw_branch::error::BranchError> and are designed to fail explicitly on:

• Invalid branch names and lifecycle transitions.
• Snapshot or sidecar integrity violations.
• DAG cycle or invalid ancestry operations.
• SQLite IO and migration failures.
• Merge conflicts requiring manual strategy.

Safety and invariants

• No unwrap/expect in library source (clippy::unwrap_used denied).
• Public API docs required (missing_docs denied).
• SQLite isolation by branch file to avoid cross-branch mutable overlap.
• DAG cycle prevention before edge insertion.
• Atomic merge and commit transactions.
• Snapshot integrity checks via BLAKE3 hashes.

Performance targets (design goals)

These targets are validated by benchmark suites in benches/branch_bench.rs.

Compatibility

• Rust edition: 2021
• MSRV: 1.75 (see Cargo.toml)
• Runtime: Tokio
• Storage backend: SQLite via SQLx

Development and CI commands

cargo build --all-features
cargo test --all-features
cargo clippy --all-targets --all-features -- -D warnings
cargo fmt --all -- --check
cargo deny check
cargo audit
cargo bench --no-run
cargo bench -- --test

Release process

1. Update Cargo.toml version and CHANGELOG.md.
2. Ensure CI passes on main.
3. Create and push a version tag (vX.Y.Z).
4. publish.yml verifies the tag/version match, publishes to crates.io, and creates a GitHub release.