Language / 语言: English | 中文

IRVibeOS

IRVibeOS is an LLVM IR-native VibeOS/runtime prototype. The 1.0 hosted release focuses on one useful loop:

intent -> LLVM IR -> verify -> save as module -> run

The long-term goal is an operating system surface where software is created from intent and stored as verifiable LLVM IR. The bigger north star is a complete bare-metal OS with a coding-agent interface comparable in workflow quality to Claude Code or Codex. The current usable target is hosted mode on an existing OS.

Current Status

Usable in 1.0 hosted mode

• Verify all LLVM IR files with tools/verify.ps1.
• Build all tracked .ll files with tools/build.ps1.
• Run hosted modules from modules/<name>/main.ll.
• Generate a module from an intent with host/hosted_vibe.py.
• Run the hosted shell with lli src_ir\irvibeos.ll.

Experimental / roadmap

• Bare-metal EXEC payload loading.
• Tier1 ESP32 seed.
• Tier2 UEFI seed.
• On-device vibe_engine.ll API calls and response parsing.
• Dependency solving beyond the simple deps.txt module convention.

Quick Start

Requirements:

• LLVM tools in PATH: llvm-as, llc, lli.
• Python 3 for host tooling.
• Optional: requests for AI providers.

Verify and build:

.\tools\verify.ps1
.\tools\build.ps1 -Clean

Run the hosted shell:

lli src_ir\irvibeos.ll apps
"hello" | lli src_ir\irvibeos.ll run
lli src_ir\irvibeos.ll verify

Create a module without network access:

python host\hosted_vibe.py --name demo --intent "print a hello message" --provider template --run

Create a module with an OpenAI-compatible API:

python host\hosted_vibe.py `
  --name ai_demo `
  --intent "print three short lines about LLVM IR" `
  --provider openai-compatible `
  --api-base http://localhost:11434/v1 `
  --api-key dummy `
  --model llama3 `
  --run

For OpenAI or Claude, pass --provider openai or --provider claude with --api-key and --model, or set OPENAI_API_KEY / ANTHROPIC_API_KEY plus IRVIBEOS_MODEL.

Architecture

IRVibeOS/
  src_ir/
    irvibeos.ll          hosted IR shell: apps/run/deps/vibe/verify
    vibe_engine.ll       experimental on-device vibe loop

  host/
    hosted_vibe.py       hosted 1.0 intent -> IR -> module tool
    ai_host.py           experimental TALK host for seeds

  modules/
    <name>/main.ll       runnable hosted module
    <name>/deps.txt      simple dependency metadata

  seed/
    tier0_mcu/seed.ll    TALK seed core
    tier3_hosted/seed.ll hosted byte I/O adapter
    tier1_connected/     planned ESP32 seed
    tier2_pc/            planned UEFI seed

  knowledge/             AI reference material, not runtime source
  tools/                 verification and build scripts

Module Contract

A hosted module is a directory under modules/:

modules/example/
  main.ll
  deps.txt

main.ll must define:

define i32 @main()

tools/verify.ps1 checks all .ll files with llvm-as and verifies that every module directory has main.ll and deps.txt.

Source Rule

System/device source is LLVM IR (.ll or .bc). Other languages are allowed only for:

• host/: development tools that run on the PC.
• tools/: repository verification/build tools.
• docs/ and knowledge/: documentation and AI reference material.

Roadmap

1. Hosted 1.0: intent-to-module workflow, verification, build, module shell.
2. Hosted 1.x: safer module names, richer metadata, dependency checks, more examples.
3. Seed runtime: define an executable payload format for TALK EXEC.
4. Device-driven vibe: implement vibe_engine.ll JSON/API parsing and platform externals.
5. Hardware tiers: implement ESP32 and UEFI seeds.
6. Bare-metal agent OS: QEMU boot, kernel, storage, network, IR runtime, agent tool API, permissions, and rollback.

See docs/BAREMETAL_AGENT_OS.md for the full target.

References

• LLVM Language Reference: https://llvm.org/docs/LangRef.html
• Repository: https://github.com/Pulsareon/IRVibeOS.git