oscar-agent 0.6.2

OSCAR - GitHub-Specialized AI Coding Assistant (multi-provider: Gemini, OpenAI GPT-5.5/5.4, Claude 4.8/4.6, Groq, Ollama)

OSCAR — GitHub-Specialized AI Coding Assistant

A VS Code extension and CLI powered by the Asterix agentic framework. Specialized for GitHub workflows: branch comparison, PR review, diff analysis, and git automation. Bring your own LLM — Gemini, OpenAI GPT-5, Anthropic Claude 4, Groq, or local Ollama models, switchable with a single environment variable.

What's new in v0.6.2

• Catalog refresh after full online syntax fact-check against platform.claude.com, ai.google.dev, developers.openai.com, and docs.ollama.com.
• Claude Opus 4.8 added (released 2026-05-28); Opus 4.7 retained as legacy.
• Anthropic tool_choice translation now accepts both {"type":"function","function":{"name":"x"}} (legacy) and {"type":"function","name":"x"} (current) OpenAI shapes.

What's new in v0.6.1

• Multi-provider LLM support. Switch between Gemini, OpenAI, Anthropic, Groq, and Ollama via OSCAR_LLM_PROVIDER / OSCAR_LLM_MODEL.
• 26-model catalog including GPT-5.5 / 5.4 family, Claude 4 family (opus/sonnet/haiku), Gemini 3.5/3.1/2.5, and popular open-source models on Ollama.
• New /providers API endpoint and oscar providers CLI command for discovering supported models.
• VS Code status-bar item showing the active provider/model, with a quick-pick command (OSCAR: Select LLM provider and model) to switch.
• Fixed in 0.6.1: OpenAI reasoning models (gpt-5.5, gpt-5.4, o3, o4-mini) now correctly send max_completion_tokens and skip temperature (Asterix's native handler only knew about a few of these — picking o3 previously would crash). Ollama calls no longer send tool_choice (which the OpenAI-compat layer rejects). Gemini default stays on gemini-2.5-flash because the 3.5-flash Vertex AI rollout is still region-gated.

Architecture

┌─────────────────────────────────────────────────────┐
│ VS Code Extension (TypeScript)                       │
│  ├── Sidebar WebviewViewProvider (chat UI)            │
│  ├── Status-bar provider/model indicator              │
│  ├── Quick-pick model switcher                        │
│  └── HTTP/SSE client → FastAPI backend                │
├─────────────────────────────────────────────────────┤
│ FastAPI Server (Python)                              │
│  ├── /chat, /branches, /compare, /review, /history    │
│  ├── /providers (multi-provider model catalog)        │
│  └── SSE streaming for real-time progress             │
├─────────────────────────────────────────────────────┤
│ OSCAR Agent Layer (Python)                           │
│  ├── Asterix Agent (ReAct loop, memory, state)        │
│  ├── asterix_patch.py — multi-provider LLM bridge     │
│  ├── Safety callbacks (on_before_tool_call)           │
│  └── Audit logging (on_after_tool_call)               │
├─────────────────────────────────────────────────────┤
│ Tools (registered via @agent.tool())                 │
│  ├── git_* (status, compare, review, log, diff, ...)  │
│  ├── shell (subprocess with safety checks)            │
│  ├── web_search (Tavily with dual-key fallback)       │
│  └── browser (Playwright: navigate, extract, search)  │
├─────────────────────────────────────────────────────┤
│ LLM providers (pick one)                             │
│  ├── Gemini · 2.5 Pro / Flash / Flash-Lite / 2.0 Flash │
│  ├── OpenAI · GPT-5 · GPT-5 mini · GPT-5 nano · o3 ··· │
│  ├── Anthropic · Claude Opus/Sonnet/Haiku 4.x          │
│  ├── Groq    · Llama 3.3 70B Versatile · 3.1 8B Instant│
│  └── Ollama  · Llama 3.3 / Qwen 2.5 Coder / DeepSeek …  │
└─────────────────────────────────────────────────────┘

Features

• Git-specialized tools — status, branch compare, PR review, log, diff, checkout, commit, push
• Multi-provider LLM — five providers, 21 catalogued models, configurable primary + fallback
• Browser automation — navigate, search, extract content, download files (Playwright)
• Web search — Tavily-based with dual API key fallback
• Shell execution — cross-platform command runner with safe-command allowlist
• Human-in-the-loop safety — auto-approve low risk, confirm medium/high, typed CONFIRM for dangerous ops
• Persistent memory — session context, knowledge base, user preferences via Asterix memory blocks
• Streaming progress — Server-Sent Events through FastAPI for real-time updates
• VS Code sidebar — chat UI with branch comparison widget and status-bar model picker

Installation

Prerequisites

• Python >= 3.10
• (Optional) Google Cloud CLI — only needed for the Gemini/Vertex AI default
• (Optional) Ollama — only needed if you want local models

Python Backend

git clone https://github.com/adityasarade/OSCAR.git
cd OSCAR
pip install -e .
playwright install chromium

Choose your LLM provider

OSCAR ships with five providers ready to use. Copy .env.example to .env and set the provider you want:

cp .env.example .env

Provider	Set in .env	Credentials
Gemini (default)	OSCAR_LLM_PROVIDER=gemini	gcloud auth application-default login (or GEMINI_API_KEY)
OpenAI	OSCAR_LLM_PROVIDER=openai	OPENAI_API_KEY
Anthropic Claude	OSCAR_LLM_PROVIDER=anthropic	ANTHROPIC_API_KEY
Groq	OSCAR_LLM_PROVIDER=groq	GROQ_API_KEY
Ollama (local)	OSCAR_LLM_PROVIDER=ollama	none — just have Ollama running locally

The full model catalog, with recommendations:

Provider	Models
Gemini	gemini-2.5-flash ★ · gemini-3.5-flash† · gemini-3.1-pro-preview† · gemini-3.1-flash-lite† · gemini-2.5-pro · gemini-2.5-flash-lite
OpenAI	gpt-5.4-mini ★ · gpt-5.5 · gpt-5.5-2026-04-23 · gpt-5.4 · gpt-5.4-nano · o4-mini · o3 · gpt-4.1 · gpt-4o
Anthropic	claude-sonnet-4-6 ★ · claude-opus-4-8 · claude-haiku-4-5-20251001 · claude-opus-4-7
Groq	llama-3.3-70b-versatile ★ · llama-3.1-8b-instant
Ollama	llama3.3 ★ · llama3.1 · qwen2.5-coder · deepseek-coder-v2 · mistral-nemo

(★ = recommended default for the provider; † = Gemini 3.x is on the Developer API now; Vertex AI rollout is still region-gated as of 2026-05-27)

Run oscar providers for the live, color-coded version of this table with API-key status indicators.

VS Code Extension

cd vscode-oscar
npm install
npm run compile
# Press F5 in VS Code to launch the Extension Development Host

The extension shows the active LLM in the status bar. Click it (or run OSCAR: Select LLM provider and model from the Command Palette) to pick a different one — the chosen env vars are copied to your clipboard, ready to paste into .env.

Usage

CLI

oscar                  # Start interactive session
oscar --debug          # Debug mode
oscar --dry-run        # Dry run (no destructive ops)
oscar --config-check   # Verify configuration

Inside the prompt:

OSCAR> providers           # List all supported providers and models
OSCAR> config              # Show current provider/model + tool count
OSCAR> test                # Round-trip the current LLM
OSCAR> git status
OSCAR> compare main and feature-branch
OSCAR> review feature-branch against main
OSCAR> search for Python async best practices
OSCAR> navigate to https://docs.python.org

API

oscar-server                                          # Start FastAPI on port 8420
curl http://127.0.0.1:8420/providers | jq            # Inspect the live model catalog
curl -X POST http://127.0.0.1:8420/chat \
     -H 'Content-Type: application/json' \
     -d '{"message":"show me git status"}'

Project Structure

src/oscar/
├── api/server.py               # FastAPI endpoints incl. /providers
├── cli/main.py                 # CLI with `providers`, `config`, `test`
├── config/
│   ├── providers.py            # 5-provider / 21-model catalog (single source of truth)
│   ├── prompts.py              # System prompt
│   └── settings.py             # Env-driven settings, API-key resolution
├── core/
│   ├── agent.py                # Asterix agent orchestrator
│   ├── asterix_patch.py        # Multi-provider LLM bridge (Gemini/OpenAI/Claude/Groq/Ollama)
│   ├── safety.py               # Human-in-the-loop safety callbacks
│   └── repo_context.py         # Active-repo handling
└── tools/
    ├── git_tool.py             # Git operations (9 functions)
    ├── shell.py                # Shell command execution
    ├── web_search.py           # Tavily web search
    └── browser.py              # Playwright browser automation

Built With

• Asterix — Agentic framework (ReAct loop, memory, tool management)
• Gemini · OpenAI · Anthropic · Groq · Ollama — pluggable LLM providers
• google-genai · openai · anthropic · ollama — official Python SDKs
• FastAPI — HTTP backend with SSE streaming
• Playwright — Headless browser automation
• Tavily — Web search API
• Rich — Terminal UI and formatting
• Click — CLI framework
• TypeScript — VS Code extension

License

MIT

Author

Built by Aditya Sarade — Final year AI & Data Science, AISSMS IOIT, Pune.