REBUSS.Pure – Pull Request Analysis MCP Server

REBUSS.Pure is a lightweight MCP (Model Context Protocol) server designed to enable intelligent pull request analysis by AI agents such as GitHub Copilot.

The server provides structured access to pull request data from Azure DevOps while minimizing context usage and avoiding the need to clone repositories locally.

Its primary goal is to support efficient AI-assisted code reviews by exposing small, well-defined tools that allow agents to analyze pull requests incrementally.

---

Why this project exists

Modern AI coding assistants often struggle with large pull requests because:

• loading an entire PR diff can overflow the model context window
• cloning repositories locally is slow and unnecessary
• AI agents need structured access to repository data instead of large text blobs

REBUSS.Pure solves this problem by exposing small composable MCP tools that allow an AI agent to retrieve only the information it needs.

This enables scalable pull request analysis even for large repositories.

---

Architecture

The system follows a layered architecture:

GitHub Copilot
│
│ prompt file  (stdio / JSON-RPC 2.0)
▼
MCP Server (REBUSS.Pure)
│
│ Azure DevOps REST API
▼
Azure DevOps Repository

Instead of sending a large diff to the AI model, the MCP server exposes tools that allow the agent to retrieve data incrementally.

---

Available MCP Tools

get_pr_metadata(prNumber)

Returns high-level information about the pull request.

Used by the AI agent to determine the review strategy and understand the scope of the changes.

Example fields:

• PR title
• author
• base branch and commit SHA
• head branch and commit SHA
• number of commits
• number of changed files
• additions and deletions
• truncated PR description

This call is always the first step in the review workflow.

---

get_pr_files(prNumber)

Returns a structured list of files changed in the pull request.

Each file entry includes:

• file path
• change status (add, edit, delete, rename)
• additions and deletions
• total number of changes
• file extension
• binary, generated, and test file flags
• review priority (high / medium / low)

The response also includes an aggregated summary by file category (source, test, config, docs, binary, generated).

This allows the AI agent to review the pull request file-by-file instead of loading the entire diff.

---

get_pr_diff(prNumber)

Returns a structured JSON object with per-file hunks for all changed files in the pull request.

The response includes for each file:

• file path and change type
• skip reason (if the diff was skipped)
• additions and deletions count
• a list of hunks, each with location metadata and ordered lines
• each line includes an operation (+, -, or ) and the line text

Files where a diff is not meaningful (deletions, renames, binary files, generated files, and full-file rewrites) are automatically skipped — see Diff Skip Behavior below. Skipped files have a skipReason field and an empty hunks list.

This tool is useful for a quick overview of all changes at once. For large pull requests, prefer get_file_diff to retrieve changes file-by-file.

---

get_file_diff(prNumber, path)

Returns a structured JSON object with the diff for a specific file in the pull request.

The response includes:

• file path and change type
• skip reason (if the diff was skipped)
• additions and deletions count
• a list of hunks, each with location metadata and ordered lines
• each line includes an operation (+, -, or ) and the line text

If the requested file falls into a skip category (deleted, renamed, binary, generated, or full-file rewrite) the response includes a skipReason field and an empty hunks list — see Diff Skip Behavior below.

The agent uses this to analyze code changes with minimal context cost.

This is the default analysis method for reviewing code changes.

---

get_file_content_at_ref(path, ref)

Returns the full content of a file for a specific revision.

The ref parameter can be:

• a commit SHA
• a branch name
• a tag

This is typically used with:

• head.sha – the new version of the file
• base.sha – the previous version of the file

Full file retrieval is used only when the diff alone is not sufficient to understand the change.

---

Diff Skip Behavior

The diff provider automatically skips diff generation for files where a diff would be unnecessary, misleading, or wasteful.

When a diff is skipped, the file entry includes a skipReason field that explains why, and the hunks list is empty.

Skip categories

Category	skipReason value	Description
File deletions	file deleted	The file was removed entirely. No content is fetched.
File renames	file renamed	A pure rename. Fetching content at the new path against the base commit would produce a misleading full-file diff.
Binary files	binary file	Detected by file extension (.dll, .png, .zip, .pdf, .woff2, etc.). Diffs are meaningless for binary content.
Generated files	generated file	Detected by path patterns (/obj/, /bin/, node_modules/, .g.cs, .designer.cs, lock files, etc.). Generated output changes are noise.
Full-file rewrites	full file rewrite	Both file versions exist and have at least 10 lines, but the diff contains zero unchanged context lines — every line was replaced. This typically indicates a formatting rewrite, re-encoding, or tooling regeneration.

Skipped file format

Skipped files appear in the structured JSON output with:

{
  "path": "path/to/file",
  "changeType": "delete",
  "skipReason": "file deleted",
  "additions": 0,
  "deletions": 0,
  "hunks": []
}

This allows consumers to see the file in the changed-files list while keeping the diff output minimal.

When diffs are still computed

Diff generation proceeds normally for:

• file additions (add)
• file edits (edit) to non-binary, non-generated source files
• any file that does not match a skip category

---

AI-Driven Review Workflow

The intended workflow for AI agents is:

1. Retrieve pull request metadata
2. Retrieve the list of changed files
3. Review files one by one (or the full diff for small PRs)
4. Recognize and skip files whose diffs were automatically omitted
5. Retrieve full file content only when necessary
6. Produce a structured review report

This incremental approach prevents context overload and allows AI tools to analyze large pull requests efficiently.

---

Example Workflow

get_pr_metadata(PR)

↓ determine review strategy

get_pr_files(PR)

↓ iterate files

get_pr_diff(PR)              ← full diff (small PRs)
or
get_file_diff(PR, path)      ← per-file diff (all PR sizes)

↓ if more context required

get_file_content_at_ref(path, head.sha)

---

Design Goals

REBUSS.Pure was designed with the following goals:

• minimize AI context usage
• avoid cloning repositories locally
• provide deterministic tool outputs
• skip diff generation for files where diffs are not meaningful
• enable scalable pull request analysis
• integrate seamlessly with GitHub Copilot

---

Configuration

REBUSS.Pure requires Azure DevOps connection settings.

Required settings under the AzureDevOps section:

Setting	Description
OrganizationName	Azure DevOps organization name
ProjectName	Azure DevOps project name
RepositoryName	Git repository name within the project
PersonalAccessToken	PAT with read access to code and pull requests

The server will fail to start if any required field is missing.

---

Storing Secrets Locally

Never put real secrets in appsettings.json. That file is committed to the repository and must only contain empty defaults or non-sensitive values.

Option 1 — appsettings.Local.json (recommended)

Create a file named appsettings.Local.json in the REBUSS.Pure project directory:

{
  "AzureDevOps": {
    "OrganizationName": "your-org",
    "ProjectName": "your-project",
    "RepositoryName": "your-repo",
    "PersonalAccessToken": "your-pat"
  }
}

This file is already excluded from Git via .gitignore (appsettings.*.json) and will never be committed.

The application loads it automatically and its values override anything in appsettings.json.

Option 2 — Environment Variables

Set environment variables using the AzureDevOps__ prefix (double underscore):

AzureDevOps__OrganizationName=your-org
AzureDevOps__ProjectName=your-project
AzureDevOps__RepositoryName=your-repo
AzureDevOps__PersonalAccessToken=your-pat

Environment variables take the highest priority and override both JSON files.

Configuration priority (lowest → highest)

1. appsettings.json — committed, contains defaults (no secrets)
2. appsettings.Local.json — not committed, contains your personal secrets
3. Environment variables — highest priority, useful for CI/CD or container deployments

---

Integration with GitHub Copilot

This repository includes a prompt file:

.github/prompts/review-pr.prompt.md

The prompt instructs GitHub Copilot how to perform a structured code review using the MCP tools provided by REBUSS.Pure.

The agent follows a controlled workflow:

1. metadata analysis
2. file list analysis
3. diff-based inspection
4. optional full file retrieval

---

Example Usage

In GitHub Copilot Chat you can trigger the review workflow:

PullRequest 123 #review-pr

Copilot will then:

1. load the review prompt file
2. call the MCP tools
3. analyze the pull request incrementally
4. generate a structured code review report

---

Project Status

This project is an experimental foundation for AI-assisted repository intelligence tools.

Future improvements may include:

• repository architecture analysis
• semantic code indexing
• automated PR risk scoring
• deeper integration with AI planning agents
• automated code quality metrics

---

License

MIT License

---

Author

Michał Korbecki

Application Architect & Software Engineer
Creator of the REBUSS developer tooling ecosystem

• https://github.com/rebuss
• https://REBUSS.Pureo