synthmerge

AI-powered conflict resolution for Git

synthmerge is a minimalistic command-line tool that leverages AI to automatically resolve conflicts arising from Git commands. Built on the research of the Patchpal project, it provides a pure AI inference layer that seamlessly integrates with your existing Git workflow. While the AI generates code solutions, all code reviews and approvals remain within your favorite code editor.

---

🌟 Core Principles

1. Specialized AI Layer
   Dedicated AI inference system that complements Git without duplicating its core functionality

2. Git Integration
   Leverages Git's diff3 conflict markers as the foundation (requires git config merge.conflictStyle diff3)

3. Editor Agnostic
   Compatible with any development environment (VS Code, Emacs, Vim, etc.)

---

✨ Key Features

• Universal Git Operation Support
  Seamlessly integrates with all Git operations that create conflicts:

  • cherry-pick
  • merge
  • rebase
  • revert
  • stash pop

• Model Flexibility
  No fine-tuning required, any instruct large language model can be used

• Parallel Multi-AI Endpoint Support
  Simultaneously queries multiple AI models to resolve conflicts:

  • Patchpal-backend (fine-tuned specifically for conflict resolution)
  • Self-hosted open-weight open source LLMs with OpenAI-compatible endpoints
  • Gemini (via OpenAI-compatible API)
  • Claude (via Anthropic API)

• Parameter Variants Support
  Each AI endpoint can be configured with multiple parameter variants to run multiple inference strategies:

  • Different reasoning effort levels (high, medium, low)
  • Temperature, top_p, top_k, min_p sampling parameters
  • Context handling options (context: no_diff: no_training: layout: flags)
  • Custom JSON parameters that can be injected into the request payload from the YAML configuration (either at the endpoint level or in each variant)

• Results Deduplication
  Consolidates identical solutions and displays model and/or parameter variant agreement

• Review Using Your Workflow

  • Resolved conflicts appear in your editor with model attribution
  • AI-generated code requires manual review before commit

• Fail-Safe Design

  • When one model fails to resolve a conflict, Git's original conflict remains alongside solutions from other models for that hunk
  • Each AI endpoint can be configured with timeout, delay, and max_delay parameters
  • Custom root certificates can be added to the endpoint configuration
  • Wait time between requests can be specified per endpoint

• Benchmark
  Built-in benchmarking tool (synthmerge_bench) for evaluating model accuracy on conflict resolution tasks

• Context Lines Configuration
  Configurable context lines for code, diff, and patch to control the amount of surrounding information provided to AI models

---

🛠 How It Works

1. Git sets up conflicts

   git config merge.conflictStyle diff3  # Must be set
   git cherry-pick -x <commit>           # Git detects conflicts

2. synthmerge analyzes conflicts

   • Reads Git's diff3 conflict markers
   • Extracts context (3 lines before/after conflict)
   • Generates precise AI prompt

3. AI resolves conflict

   • Sends code + patch to configured endpoint
   • Receives resolved code

4. Git gets updated

   • synthmerge inserts the AI resolution into existing diff3 markers
   • You review in your editor

✅ Works also for git rebase, revert and merge conflict resolutions.

---

🚀 Usage

# Ensure Git is configured for diff3 conflict style
git config merge.conflictStyle diff3

# Attempt cherry-pick (will leave conflicts unresolved)
git cherry-pick -x <commit>

# Resolve conflicts with AI
synthmerge

# Review synthmerge resolved conflicts in each unmerged file ...
git diff --name-only --diff-filter=U

# ... or linearized in a single buffer to edit with ripgrep-edit
rg-edit -E vim -U -e '(?s)^<<<<<<<+ .*?^>>>>>>>+ '
rg-edit -E emacsclient -U -e '(?s)^<<<<<<<+ .*?^>>>>>>>+ '

---

⚙️ Configuration

Create ~/.config/synthmerge.yaml based on synthmerge.yaml:

endpoints:

  - name: "Claude Sonnet 4.5"
    url: "https://api.anthropic.com/v1/messages"
    type: "anthropic"
    x_api_key_file: "~/.keys/anthropic.api-key"
    json:
      max_tokens: 20000
      model: "claude-sonnet-4-5"
      temperature: 0
    headers:
      anthropic-version: "2023-06-01"
    variants:
      - name: "default"
      - name: "no_diff"
        context:
          no_diff: true

  - name: "Vertex Claude Sonnet 4.0"
    url: "https://host/path"
    type: "anthropic"
    api_key_file: "~/.keys/claude.api-key"
    json:
      anthropic_version: "something-YYYY-MM-DD"
      max_tokens: 20000
      temperature: 0
    variants:
      - name: "default"
      - name: "no_diff"
        context:
          no_diff: true
    # Optional root certificate for HTTPS endpoints
    # root_certificate_pem: "~/.ssl/corp-ca.pem"

  - name: "Patchpal AI"
    type: "patchpal"
    url: "http://patchpal.usersys.redhat.com:9080/v1"

  - name: "Gemini 2.5 Flash"
    url: "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions"
    type: "openai"
    api_key_file: "~/.keys/gemini.api-key"
    json:
      model: "gemini-2.5-flash"
      # "none" (only available with Flash) works better with default layout
      reasoning_effort: "none"
    variants:
      - name: "default"
      - name: "no_diff"
        context:
          no_diff: true

  - name: "Gemini 2.5 Pro"
    url: "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions"
    type: "openai"
    api_key_file: "~/.keys/gemini.api-key"
    json:
      model: "gemini-2.5-pro"
      reasoning_effort: "low"
    context:
      # reasoning_effort != none needs the prompt at the top of system_message
      layout:
        system_message:
          - prompt
          - training
          - diff
        user_message: []

  - name: "Gemini 3 Pro preview"
    url: "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions"
    type: "openai"
    api_key_file: "~/.keys/gemini.api-key"
    json:
      model: "gemini-3-pro-preview"
      reasoning_effort: "low"
    context:
      layout:
        system_message:
          - prompt
          - training
          - diff
        user_message: []

  - name: "llama.cpp vulkan minimal" # requires --no-jinja
    url: "http://localhost:8811/v1/chat/completions"
    type: "openai"

  - name: "llama.cpp vulkan" # requires --no-jinja
    url: "http://localhost:8811/v1/chat/completions"
    #timeout: 600000
    #retries: 10
    #delay: 1000
    #max_delay: 600000
    #wait: 1000
    type: "openai"
    json:
      #temperature: 0.7
      #top_p: 0.8
      #top_k: 20
      #min_p: 0

      # n_probs: 1 provides the probability of the lowest probability
      # token in the resolved conflict
      n_probs: 1

      # n_probs: 2 same as n_probs: 1 but it also provides two more
      # beams with the perplexity search algorithm of synthmerge
      # applied to the logprobs, which is a client side only
      # approximated beam search
      #n_probs: 2
    variants:
      # one query for each entry in the variants list
      - name: "default"
      - name: "no_diff"
        context:
          no_diff: true
      #- name: "min_p"
      #  json:
      #    temperature: 0.3
      #    top_p: 1.0
      #    top_k: 0
      #    min_p: 0.9

  - name: "llama.cpp vulkan no_chat" # requires --no-jinja
    url: "http://localhost:8811/v1/completions"
    type: "openai"
    no_chat: true
    context:
      no_training: true

---

🌐 Supported AI Endpoints

Endpoint Type	Example Configuration	Notes
Patchpal-backend	type: "patchpal"	Fine-tuned for patch resolution
OpenAI protocol	type: "openai"	Self-hosted LLMs (e.g., llama.cpp) and Gemini
Anthropic protocol	type: "anthropic"	Claude models

✅ Gemini supports a compatible OpenAI endpoint
✅ Models work with stock weights – the prompt engineering simulates Patchpal's fine-tuned behavior.

---

⚙️ Context Layout Configuration

The context: layout: configuration allows fine-grained control over how information is structured in a LLM request.

• Prompt placement: All models tested so far (including Gemini 2.5 Flash with reasoning_effort: none) perform best when the most important directives are closest to the generation
• Gemini thinking models exception: Gemini models with reasoning_effort != none require the prompt explaining the challenge at hand to be at the top of the system message
• Layout flexibility: The layout configuration enables each model to select the optimal information structure

Available layout elements:

• prompt: The high-level prompt explaining the challenge
• training: The synthetic training examples
• diff: The full git diff showing all other changes of the commit

Context control flags:

• no_diff: Disable diff inclusion in context
• no_training: Disable training examples in context

Configuration examples:

# Set layout at endpoint level
context:
  layout:
    system_message:
      - prompt
    user_message:
      - training
      - diff

# Override layout in a variant
variants:
  - name: "no_diff"
    context:
      no_diff: true

The layout can be configured either at the endpoint level or in individual variants, but not both simultaneously in the same endpoint.

---

🛠 Installation

Fedora

A Fedora Copr package is available:

1. Install Synthmerge:

   sudo dnf copr enable vittyvk/synthmerge
   sudo dnf install synthmerge

2. Configuration:

   cp -a /usr/share/synthmerge/synthmerge.yaml ~/.config/
   $EDITOR ~/.config/synthmerge.yaml

From source code

1. Install Synthmerge:

   git clone https://gitlab.com/aarcange/synthmerge.git
   cd synthmerge
   cargo build --release
   sudo cp target/release/synthmerge /usr/local/bin/

2. Configuration:

   cp synthmerge.yaml ~/.config/
   $EDITOR ~/.config/synthmerge.yaml

---

🎥 Demo

---

📊 Benchmark Statistics

The following statistics were generated using the synthmerge_bench tool on a C language dataset to evaluate model performance on conflict resolution tasks. These results may vary depending on prompt, context, and other variables.

Accuracy checks if the AI resolved conflict is an exact match including all spaces, tabs, and newlines.

Accuracy (aligned) checks equality of whitespace patterns up until the first non-whitespace character, ignoring differences in lines without non-whitespace characters and whitespace variations after the first non-whitespace character (i.e. Python equivalence).

Accuracy (stripped) compresses all whitespaces and newlines into a single space (i.e. C/C++/Rust/JavaScript equivalence).

This measurement used only new test data never exposed to the model during the fine tuning process.

Claude Sonnet 4.5 and Gemini 3 Pro preview not done yet.

Model: Claude Sonnet 4.0 (default)
  Accuracy: 66.70% (753/1129)
  Accuracy (aligned): 70.42% (795/1129)
  Accuracy (stripped): 73.34% (828/1129)
  Error Rate: 0.00% (0/1129)
  Average tokens: 5730.47
  Average duration: 7.03 s

Model: Claude Sonnet 4.0 (no_diff)
  Accuracy: 65.19% (736/1129)
  Accuracy (aligned): 68.29% (771/1129)
  Accuracy (stripped): 71.48% (807/1129)
  Error Rate: 0.00% (0/1129)
  Average tokens: 1184.14
  Average duration: 6.34 s

# only the Patchpal Beam 0 is comparable to the non Patchpal models
Model: Patchpal AI
  Accuracy: 64.57% (729/1129)
  Accuracy (aligned): 68.47% (773/1129) # might be duplicate with other beams
  Accuracy (stripped): 71.12% (803/1129) # might be duplicate with other beams
  Error Rate: 0.44% (5/1129)

Model: Gemini 2.5 Pro (high) # reasoning_effort: high
  Accuracy: 55.18% (623/1129)
  Accuracy (aligned): 60.67% (685/1129)
  Accuracy (stripped): 63.42% (716/1129)
  Error Rate: 0.00% (0/1129)

Model: Gemini 2.5 Flash (none no_diff) # reasoning_effort: none
  Accuracy: 53.06% (599/1129)
  Accuracy (aligned): 63.24% (714/1129)
  Accuracy (stripped): 66.25% (748/1129)
  Error Rate: 3.28% (37/1129)
  Average tokens: 1036.06
  Average duration: 1.18 s

# context: layout: system_message: [ prompt ] user_message: [ training, diff ]
Model: Gemini 2.5 Pro (low userctx) # reasoning_effort: low
  Accuracy: 52.44% (592/1129)
  Accuracy (aligned): 56.95% (643/1129)
  Accuracy (stripped): 59.70% (674/1129)
  Error Rate: 5.49% (62/1129)
  Average tokens: 6014.82
  Average duration: 9.68 s

# context: layout: system_message: [ prompt ] user_message: [ training, diff ]
Model: Gemini 2.5 Pro (low no_diff) # reasoning_effort: low
  Accuracy: 51.99% (587/1129)
  Accuracy (aligned): 55.36% (625/1129)
  Accuracy (stripped): 58.02% (655/1129)
  Error Rate: 2.92% (33/1129)
  Average tokens: 1931.27
  Average duration: 9.11 s

# temperature: 0.7 top_k: 20 top_p: 0.8 min_p: 0
# llama.cpp vulkan Q6_K
Model: Qwen3-Coder-30B-A3B-Instruct (default)
  Accuracy: 49.69% (561/1129)
  Accuracy (aligned): 54.21% (612/1129)
  Accuracy (stripped): 56.78% (641/1129)
  Error Rate: 0.09% (1/1129)
  Average tokens: 4252.31
  Average duration: 9.18 s
  Average prob: 33.1% (+- 35.4)
  Average prob (incorrect): 16.3% (+- 40.7)
  Average prob (stripped): 56.7% (+- 27.4)
  Average prob (aligned): 58.0% (+- 27.2)
  Average prob (correct): 61.6% (+- 25.9)

Model: Gemini 2.5 Flash (none default) # reasoning_effort: none
  Accuracy: 49.60% (560/1129)
  Accuracy (aligned): 60.41% (682/1129)
  Accuracy (stripped): 63.42% (716/1129)
  Error Rate: 6.20% (70/1129)
  Average tokens: 5069.04
  Average duration: 1.15 s

# context: layout: system_message: [ prompt ] user_message: [ training, diff ]
Model: Gemini 2.5 Flash (low no_diff userctx) # reasoning_effort low
  Accuracy: 48.72% (550/1129)
  Accuracy (aligned): 58.19% (657/1129)
  Accuracy (stripped): 62.00% (700/1129)
  Error Rate: 2.66% (30/1129)
  Average tokens: 1916.70
  Average duration: 4.62 s

# temperature: 0.7 top_k: 20 top_p: 0.8 min_p: 0
# llama.cpp vulkan Q6_K
Model: Qwen3-Coder-30B-A3B-Instruct (no_diff)
  Accuracy: 46.94% (530/1129)
  Accuracy (aligned): 51.02% (576/1129)
  Accuracy (stripped): 53.76% (607/1129)
  Error Rate: 0.00% (0/1129)
  Average tokens: 904.89
  Average duration: 4.37 s
  Average prob: 37.1% (+- 35.1)
  Average prob (incorrect): 24.0% (+- 39.1)
  Average prob (stripped): 53.8% (+- 29.1)
  Average prob (aligned): 57.3% (+- 27.9)
  Average prob (correct): 62.6% (+- 26.5)

# context: layout: system_message: [ prompt ] user_message: [ training, diff ]
Model: Gemini 2.5 Flash (low default userctx) # reasoning_effort: low
  Accuracy: 42.52% (480/1129)
  Accuracy (aligned): 52.70% (595/1129)
  Accuracy (stripped): 55.98% (632/1129)
  Error Rate: 13.82% (156/1129)
  Average tokens: 5942.75
  Average duration: 4.22 s

# if Beam 0 is wrong, Beam 1 is right 10.54% of the time
Model: Patchpal AI #1
  Accuracy: 10.54% (119/1129)
  Accuracy (aligned): 21.17% (239/1129) # might be duplicate with other beams
  Accuracy (stripped): 30.03% (339/1129) # might be duplicate with other beams
  Error Rate: 0.53% (6/1129)

Model: Gemini 2.5 Flash (low default) # reasoning_effort: low
  Accuracy: 7.97% (90/1129)
  Accuracy (aligned): 9.57% (108/1129)
  Accuracy (stripped): 10.27% (116/1129)
  Error Rate: 85.56% (966/1129) # default layout fails with Gemini thinking mode
  Average tokens: 3719.80
  Average duration: 0.51 s

# this is comparable to Patchpal AI #1
Model: Qwen3-Coder-30B-A3B-Instruct (no_diff#1) # perplexity beam #1
  Accuracy: 7.71% (87/1129)
  Accuracy (aligned): 11.87% (134/1129) # might be duplicate with other beams
  Accuracy (stripped): 16.56% (187/1129) # might be duplicate with other beams
  Error Rate: 0.18% (2/1129)
  Average tokens: 910.68
  Average duration: 1.17 s # kvcached

# if Beam 0 and Beam 1 are wrong, Beam 2 is right 3.37% of the time
Model: Patchpal AI #2
  Accuracy: 3.37% (38/1129)
  Accuracy (aligned): 16.21% (183/1129) # might be duplicate with other beams
  Accuracy (stripped): 23.83% (269/1129) # might be duplicate with other beams
  Error Rate: 0.44% (5/1129)

# this is comparable to Patchpal AI #2
Model: Qwen3-Coder-30B-A3B-Instruct (default#2) # perplexity beam #2
  Accuracy: 1.95% (22/1129)
  Accuracy (aligned): 6.91% (78/1129) # might be duplicate with other beams
  Accuracy (stripped): 11.87% (134/1129) # might be duplicate with other beams
  Error Rate: 0.09% (1/1129)
  Average tokens: 913.69
  Average duration: 1.18 s # kvcached

---

📊 Benchmark Aggregate Accuracy

Aggregate accuracy represents the combined performance when multiple models/variants/beams are used in parallel: a conflict is considered successfully resolved if at least one model/variant/beam produces a correct solution.

Configuration	Accuracy	Accuracy (aligned)	Accuracy (stripped)
Qwen3-Coder-30B (default)	49.69%	54.21%	56.78%
Qwen3-Coder-30B (no_diff)	46.94%	51.02%	53.76%
Aggregate: Qwen3-Coder-30B (default + no_diff)	55.80%	60.50%	63.33%
(Perplexity) beams added to Qwen3-Coder-30B	63.24%	69.18%	71.83%
Claude Sonnet 4.0 (default)	66.70%	70.42%	73.34%
Qwen3-Coder-30B + Claude Sonnet 4.0	75.02%	78.39%	80.96%
Gemini 2.5 Flash (none)	49.60%	60.41%	63.42%
Gemini 2.5 Pro (low)	52.44%	56.95%	59.70%
Qwen3-Coder-30B + beams + Claude Sonnet 4.0 + Gemini 2.5 Flash + Gemini 2.5 Pro	79.98%	82.82%	84.68%
Patchpal AI (Beam 0)	64.57%	68.47%	71.12%
Aggregate: Patchpal AI (3 beams)	78.39%	81.05%	82.46%
✅ All models + all variants + all beams	84.85%	87.51%	88.66%

---

License