https://github.com/gametimesf/prbuddy
AI-powered PR review companion — authors train agents with context, reviewers ask questions with voice or text
https://github.com/gametimesf/prbuddy
Last synced: 19 days ago
JSON representation
AI-powered PR review companion — authors train agents with context, reviewers ask questions with voice or text
- Host: GitHub
- URL: https://github.com/gametimesf/prbuddy
- Owner: gametimesf
- Created: 2026-04-07T15:41:08.000Z (2 months ago)
- Default Branch: main
- Last Pushed: 2026-04-08T07:18:26.000Z (2 months ago)
- Last Synced: 2026-04-08T08:30:02.997Z (2 months ago)
- Language: Python
- Size: 251 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# PR Buddy
An AI-powered review companion for pull requests. Authors train agents with context about their PR, and reviewers can then ask questions — getting answers grounded in author knowledge, code diffs, and team context.
## How It Works
```mermaid
sequenceDiagram
participant Author
participant PRBuddy as PR Buddy
participant Weaviate as Knowledge Base
(Weaviate)
participant GitHub
participant Unblocked
Note over PRBuddy,Weaviate: Session Creation
PRBuddy->>GitHub: Fetch diff, description, comments
GitHub-->>PRBuddy: PR content
PRBuddy->>Weaviate: Auto-index diff chunks,
description, comments
Note over Author,Weaviate: Author Training
Author->>PRBuddy: "Edgar likes this approach"
PRBuddy->>Weaviate: Index author_explanation
(preserves entity names)
PRBuddy-->>Author: Got it - Edgar approves
the explicit access approach.
opt Research (team context, history, Slack)
Author->>PRBuddy: "What did the team discuss?"
PRBuddy->>Unblocked: Search team context,
Slack, historical decisions
Unblocked-->>PRBuddy: Relevant discussions
PRBuddy->>Weaviate: Index findings
PRBuddy-->>Author: Team discussed via Envoy...
end
```
```mermaid
sequenceDiagram
participant Reviewer
participant PRBuddy as PR Buddy
participant Weaviate as Knowledge Base
(Weaviate)
participant Unblocked
Note over Reviewer,Weaviate: Reviewer Q&A (auto-injection)
Reviewer->>PRBuddy: "Is Edgar ok with this?"
PRBuddy->>Weaviate: 3 parallel queries:
hybrid + BM25 + vector
Weaviate-->>PRBuddy: Matching docs
(author explanations, diff, etc.)
Note over PRBuddy: Deduplicate, rank,
inject as system context
PRBuddy-->>Reviewer: Edgar reviewed and approves.
He likes the explicit single-service access.
opt Research (when KB lacks context)
Reviewer->>PRBuddy: "How long has this been requested?"
PRBuddy->>Unblocked: Search Slack, Jira,
historical PRs
Unblocked-->>PRBuddy: Timeline + discussions
PRBuddy->>Weaviate: Index findings
PRBuddy-->>Reviewer: Requested for ~6 months,
tracked in INFRA-432.
end
```
## Architecture
```mermaid
graph TB
subgraph Frontend
WEB[Web UI
Voice + Text]
EXT[Chrome Extension
Side Panel + Diff Selection]
end
subgraph Server["FastAPI Server"]
WS[WebSocket Handler]
TEXT[Text Session]
PIPE[Pipeline Session
STT + Agent + TTS]
CI[Context Injection
Multi-query RAG]
end
subgraph Agents["Agent System (YAML-configured)"]
AT[AuthorTraining]
RQ[ReviewerQA]
RE[Research]
AT <-->|handoff| RE
RQ <-->|handoff| RE
end
subgraph Storage
WV[(Weaviate
Vector DB)]
end
subgraph External
GH[GitHub API]
UB[Unblocked MCP
Team Knowledge]
end
WEB --> WS
EXT --> WS
WS --> TEXT
WS --> PIPE
TEXT --> CI
CI --> WV
TEXT --> Agents
PIPE --> Agents
Agents --> WV
RE --> GH
RE --> UB
```
## Key Features
- **Auto-indexed PR context** — Diff, description, and comments are fetched from GitHub and indexed into Weaviate on session creation. Agents start with full context immediately.
- **Auto-context injection** — Before each reviewer question, 3 parallel RAG queries (hybrid, BM25, vector) surface relevant author knowledge automatically. The agent doesn't need to decide to search.
- **Entity-aware indexing** — Author explanations preserve people, team, and system names for reliable keyword matching (e.g., "Edgar" stays "Edgar", not "a stakeholder").
- **Voice + Text modes** — Text chat or voice (Whisper STT + Polly/OpenAI TTS).
- **Diff selection** — Chrome extension lets reviewers highlight code and ask questions with that context.
- **Research via Unblocked MCP** — Agents can search Slack, Jira, historical PRs, and team discussions.
## Quick Start
```bash
# Setup
make setup
# Edit .env and add your OPENAI_API_KEY
# Start Weaviate + dev server
make dev
```
Then open http://localhost:8000 in your browser.
### Chrome Extension
Load the extension for the side panel experience on GitHub PRs:
1. Go to `chrome://extensions`
2. Enable **Developer mode**
3. Click **Load unpacked** and select the `prbuddy-extension/` directory
### Voice Mode Requirements
For voice modes (Pipeline), you need:
- **ffmpeg** — `brew install ffmpeg` (macOS) or `sudo apt install ffmpeg` (Linux)
- **OPENAI_API_KEY** — For Whisper STT
- **AWS credentials** or **ELEVENLABS_API_KEY** — For TTS
## Session Modes
| Mode | Description | Latency |
|------|-------------|---------|
| **Text** | Chat-based Q&A | Instant |
| **Pipeline** | Whisper STT + Agent + TTS | ~2-3s |
## Configuration
Agent behaviors are configured via YAML files in `config/agents/`:
| Directory | Agents | Purpose |
|-----------|--------|---------|
| `common/` | Research | Context gathering from GitHub + Unblocked |
| `author/` | AuthorTraining, AuthorEngagement | Author interview and knowledge capture |
| `reviewer/` | ReviewerQA | Reviewer Q&A with citations |
## Development
```bash
make setup # Install dependencies
make dev # Start Weaviate + server (hot reload)
make test # Run unit tests (81 tests)
make eval # Run evals (requires Weaviate + OPENAI_API_KEY)
make clean # Clean up generated files
```
## Evals
The `evals/` directory contains end-to-end scenarios that verify author-to-reviewer knowledge sharing:
| Scenario | Tests |
|----------|-------|
| `edgar_approval` | Author says "Edgar likes this" — reviewer can find it with 5 phrasings |
| `technical_decision` | Author explains tradeoff — reviewer finds the reasoning |
| `indirect_reference` | Author provides context — reviewer finds it with different wording |
```bash
make eval # Requires Weaviate running + OPENAI_API_KEY
```
## License
MIT