From 2d27b6ae369d59f496500469ee45c6a3d6560676 Mon Sep 17 00:00:00 2001
From: GAP Promoter <bot@gitagent.sh>
Date: Thu, 28 May 2026 00:43:54 +0000
Subject: [PATCH] Add zamalali/DeepGit to the registry
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Adds DeepGit by zamalali — an advanced LangGraph-based GitHub repository
discovery agent using ColBERT v2 retrieval, cross-encoder re-ranking,
hardware-aware filtering, and multi-factor ranking.

Companion PR adding agent.yaml + SOUL.md to the source repo:
https://github.com/zamalali/DeepGit/pull/16
---
 agents/zamalali__deep-git/README.md     | 51 +++++++++++++++++++++++++
 agents/zamalali__deep-git/metadata.json | 15 ++++++++
 2 files changed, 66 insertions(+)
 create mode 100644 agents/zamalali__deep-git/README.md
 create mode 100644 agents/zamalali__deep-git/metadata.json

diff --git a/agents/zamalali__deep-git/README.md b/agents/zamalali__deep-git/README.md
new file mode 100644
index 0000000..891a67d
--- /dev/null
+++ b/agents/zamalali__deep-git/README.md
@@ -0,0 +1,51 @@
+# DeepGit
+
+**DeepGit** is an advanced, LangGraph-based agentic research workflow designed to perform deep discovery across GitHub repositories. It intelligently searches, analyses, and ranks repos based on your natural-language query — surfacing powerful but less-known tools that a standard GitHub search would miss.
+
+## What It Does
+
+Given a query like *"I need a lightweight chain-of-thought reasoning library that runs on CPU"*, DeepGit:
+
+1. **Expands your query** into precise GitHub search tags via an LLM.
+2. **Detects hardware constraints** from your wording (e.g., "GPU-poor", "low RAM") and enforces them throughout.
+3. **Retrieves at scale** using the GitHub API and hybrid dense retrieval (ColBERT v2 token embeddings + BM25 + FAISS).
+4. **Re-ranks** with a cross-encoder (MiniLM-L-6-v2) for passage-level accuracy.
+5. **Filters by dependency compatibility** — inspects `requirements.txt` / `pyproject.toml` to discard repos your hardware can't run.
+6. **Analyses community health** — stars, forks, issue cadence, recent commits, contributor count.
+7. **Scores code quality** — structural signals indicating a well-organised, production-ready project.
+8. **Delivers a ranked table** with links, similarity scores, hardware badges, and health indicators.
+
+## Key Features
+
+- 🔍 **Deep semantic search** — ColBERT v2 multi-dimensional token-level matching, not just cosine similarity
+- ⚛️ **Cross-encoder re-ranking** — passage-level accuracy via `cross-encoder/ms-marco-MiniLM-L-6-v2`
+- 🔩 **Hardware-aware filtering** — only recommends repos that actually run on your machine
+- 📊 **Multi-factor ranking** — combines semantic similarity, community health, and code quality
+- 🚀 **Gradio UI + LangGraph dev dashboard** supported
+- 🐳 **Docker-ready** with a provided Dockerfile
+
+## Example Usage
+
+```bash
+# Set your API keys
+export GROQ_API_KEY=your_key_here
+export GITHUB_API_KEY=your_github_token
+
+# Run the agent
+python app.py
+```
+
+Or launch the full LangGraph dev workflow:
+
+```bash
+langgraph dev
+```
+
+## Model
+
+Uses **Groq** (`deepseek-r1-distill-llama-70b`) by default for query expansion and reasoning. MiniMax (`MiniMax-M2.7`, 204K context) is supported as an alternative.
+
+## Links
+
+- 🔗 [GitHub Repository](https://github.com/zamalali/DeepGit)
+- 🤗 [HuggingFace Space (lite version)](https://huggingface.co/spaces/zamal/DeepGit)
diff --git a/agents/zamalali__deep-git/metadata.json b/agents/zamalali__deep-git/metadata.json
new file mode 100644
index 0000000..cfcfbc4
--- /dev/null
+++ b/agents/zamalali__deep-git/metadata.json
@@ -0,0 +1,15 @@
+{
+  "name": "deep-git",
+  "author": "zamalali",
+  "description": "LangGraph agent that deeply searches, re-ranks, and scores GitHub repos via ColBERT retrieval, cross-encoder reranking, and hardware-aware dependency filtering.",
+  "repository": "https://github.com/zamalali/DeepGit",
+  "path": "",
+  "version": "2.0.0",
+  "category": "research",
+  "tags": ["github", "search", "research", "langgraph", "colbert", "retrieval", "ranking", "open-source-discovery", "hardware-aware"],
+  "license": "Apache-2.0",
+  "model": "groq:deepseek-r1-distill-llama-70b",
+  "adapters": ["claude-code", "system-prompt"],
+  "icon": false,
+  "banner": false
+}