पवन Pawan
Self-healing CLI coding agent. Named after Power Star Pawan Kalyan. Built in Rust. Powered by NVIDIA NIM.
9,400 LOC · 107 tests · 18 commands · 17 tools · Zero subscription
Self-Healing
Automatically fixes compilation errors, clippy warnings, and failing tests. Runs cargo check → clippy → test and applies targeted fixes.
18 Subcommands
heal, commit, review, explain, test, watch, run, init, doctor, and more — covering the full development loop.
17 Built-in Tools
File ops, search, bash, git (status/diff/add/commit/log/blame/branch/checkout/stash), sub-agents. Full filesystem + repo access.
AI Code Review
Reviews your changes with severity levels (critical/major/minor) and actionable fixes. Streaming output via SSE.
AI Commit
Stages files, generates commit message from diff, confirms and commits. Follows conventional commit format.
Headless Execution
pawan run for scripted tasks, CI integration, and autonomous sprint execution. No TTY required.
Quick Start
Self-Heal in Action
pawan heal runs cargo check → clippy → test → AI fix → repeat until green. Zero human intervention.
Supported NVIDIA NIM Models
Set model = in pawan.toml to switch instantly. Your API key works across all.
| Model | Status | Notes |
|---|---|---|
Devstral 2 123B mistralai/devstral-2-123b-instruct-2512 | recommended | Best for coding tasks, tool use |
Qwen3-Coder 32B qwen/qwq-32b | works | Strong reasoning + code |
Nemotron Super 49B nvidia/llama-3.3-nemotron-super-49b-v1 | works | NVIDIA's flagship |
StepFun Flash stepfun-ai/step-3.5-flash | works | Fast inference, lower cost |
DeepSeek R1 32B deepseek-ai/deepseek-r1-distill-qwen-32b | caution | Context drift on long tasks |
All models accessed via integrate.api.nvidia.com — one API key, 15+ models.
pawan vs Everything Else
Why we replaced Claude Code, Cursor, and OpenCode with pawan for dirmacs development.
| Feature | pawan | Claude Code / Cursor |
|---|---|---|
| Subscription cost | Free (NVIDIA NIM API) | $20–$40/month |
| Self-heal loop | Native (pawan heal) | Manual or basic |
| Runtime | Rust binary, ~4MB | 500MB+ Node/Python |
| Tool calling | 17 built-in tools | Limited or plugin-based |
| CI integration | pawan run (headless) | Not designed for CI |
| Model choice | 15+ NVIDIA NIM models | Locked to vendor |
Architecture
4-crate Rust workspace. pawan-core has zero internal dependencies — works standalone with any OpenAI-compatible API.
Configuration
pawan.toml in your project root. Switch models with one line.