Sudarshan
Sudarshan is a deterministic control plane for AI software builders that keeps the plan, policy, evidence, budget, failures, and verification state outside the conversation. It accepts an idea, PRD, or specification, produces cost estimates, and drives a model through a durable JSON action loop until verification passes. It is for developers who want reproducible, resumable AI-generated builds with enforced budgets and completion gates.
✨ Key features
- Provider-neutral runtime with OpenAI-compatible, Azure, Anthropic, Gemini, local models.
- Explainable token/cost/time estimation before build starts.
- Immutable plan and durable state with atomic checkpoints and append-only events.
- One schema-validated JSON action per turn, executed under policy.
- Automatic verification discovery and completion only after all checks pass.
- Pause, resume, and human input for real-world decisions.
🎯 Use cases
- Build a product from a PRD with enforced budget and verification.
- Resume a build after process kill or provider rate limit without losing state.
- Drive an external agent framework through a one-action JSON contract.
- Prove the loop without an API key using the deterministic command bridge.
- Generate a complete project with passing tests and deployment artifacts.
📦 Installation
🧰 Requirements: Python 3.9+ and an LLM API key or a command bridge; optional capabilities like Git, Node, Docker, SearXNG, OpenClaw are detected but not required.
pip install git+https://github.com/Suraj1235/sudarshan-superharness.git
sudarshan doctor
For development:
git clone https://github.com/Suraj1235/sudarshan-superharness.git
cd sudarshan-superharness
python -m venv .venv
Activate the environment and install:
source .venv/bin/activatepython -m pip install -e .
sudarshan doctor🚀 Usage
export SUDARSHAN_API_KEY="your-key"
sudarshan build \
--prd ./PRD.md \
--workspace ./builds/my-product \
--provider openai-compatible \
--model YOUR_MODEL_ID \
--base-url https://YOUR_PROVIDER/v1 \
--input-price 1.00 \
--output-price 4.00 \
--max-cost 100 \
--verify-command-json '["python","-m","pytest","-q"]' \
--allow-host-commands
⚠️ Good to know
Sudarshan is alpha software; it does not guarantee model reasoning perfection or production-quality output, and generated software still needs human review and threat modeling.
❓ FAQ
What Python version is required?
Python 3.9+ is required for the standalone runtime.
How does Sudarshan handle API keys?
API keys are read from environment variables and are never persisted in run state.
Can I resume a build after a crash?
Yes, use sudarshan resume --workspace ./builds/my-product to restore from durable state.
Does Sudarshan support local models?
Yes, it runs against OpenAI-compatible endpoints, Azure Foundry, Anthropic, Gemini, local models, or any agent framework via a command bridge.
📊 Repository
🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.