GitHub - AllyMarthaJ/git-reabsorb: Branch-based commit organisational tool. With some AI.

git-scramble git-reabsorb

Ever make a mess of your commits while working on a feature branch? This tool helps you reorganize them before merge.

It works by soft-resetting your branch to a base commit, then recommitting your changes using a strategy you choose—whether that's preserving the original structure, grouping by file, squashing everything, or letting an LLM figure out a sensible organization.

Installation

Or build from source:

Quick Start

# See what it would do (dry run)
git-reabsorb -n

# Reorganize commits on current branch
git-reabsorb

# Let an LLM organize your commits intelligently
git-reabsorb -s llm

# Specify a base branch explicitly
git-reabsorb --base main

Strategies

Strategy	Flag	What it does
`preserve`	`-s preserve`	Keep original commit structure (default)
`by-file`	`-s by-file`	One commit per file
`squash`	`-s squash`	Everything in one commit
`llm`	`-s llm`	AI-powered reorganization
`hierarchical`	`-s hierarchical`	Multi-phase LLM for large changes

Useful tips

Plan and Apply Separately

# Generate and save a plan (doesn't modify your branch)
git-reabsorb --save-plan

# Review the plan, then apply when ready
git-reabsorb apply

Undo

# Reset to pre-reabsorb state
git-reabsorb reset

Assess Commit Quality

# Score your commits against quality criteria
git-reabsorb assess

# Compare before/after
git-reabsorb assess --save before.json
git-reabsorb
git-reabsorb assess --save after.json
git-reabsorb compare before.json after.json

LLM Configuration

For the llm and hierarchical strategies, configure your provider:

# Use Claude (default, requires claude CLI)
git-reabsorb -s llm

# Use a local model via OpenCode
git-reabsorb -s llm --llm-provider opencode --opencode-backend lmstudio

# Or via environment
export GIT_REABSORB_LLM_PROVIDER=claude
export GIT_REABSORB_LLM_MODEL=claude-sonnet-4-20250514

As of writing, we default to claude when no provider is specified for the best performance.

License

Licensed under either of Apache License 2.0 or MIT at your option.