Honest status (13 Oct 2026): Ye ek CLI-first multi-agent orchestration prototype hai — Sakana AI ke Fugu / Marlin / TRINITY ideas se inspired. Architecture ke building blocks real hain (real CLI spawning, real AB-MCTS, real safety tripwires), par is environment mein koi LLM CLI installed nahi hai — isliye LLM steps documented mock fallback par chalte hain. Jo kaam karta hai wo neeche clearly marked hai; jo limited hai wo Current limitations mein.
⚠️ License note:LICENSEfile mein MIT hai — par license choice user confirmation pending hai. Client/production use se pehle confirm karo.
By Prosenjit Paul (aka Prosen) · Zero npm dependencies · Pure Node.js stdlib
- Ek CLI-first multi-agent orchestration prototype — ek user goal andar jata hai, planner steps banata hai, conductor unhe CLI workers (Claude Code, GitHub CLI, Copilot CLI) par chalata hai, verifier quality check karta hai.
- Conductor ab real
adapter.run()se CLIs spawn karta hai (30s timeout, argv-only execution); CLI binary missing ho to documented mock fallback (🧪) milta hai — output mein hamesha visible rehta hai ki mock tha ya real. - Search/selection real hai: AB-MCTS (UCB1, adaptive wider/deeper branching, backprop), TRINITY turn-loop (worker → verifier → revise), safety tripwires har adapter call par — par scoring deterministic rubric par chalti hai; LLM-as-judge optional hai aur abhi unconfigured.
Har command neeche 13 Oct 2026 ko khud chalakar verify kiya gaya hai. Folder names mein spaces + emoji hain — path quotes mein lo.
| Tool | Zaroori? | Note |
|---|---|---|
| Node.js v18+ | ✅ Yes | Sirf yehi required hai (tested: v24.20.0) |
gh CLI |
Missing → uske steps mock fallback par | |
claude CLI |
Missing → mock fallback par (is env mein missing) | |
github-copilot-cli |
Missing → mock fallback par (is env mein missing) | |
| Python 3.10+ | Sirf mini_fugu_conductor.py POC ke liye |
# Clone the repository
git clone https://github.com/PsProsen-Dev/RTXCoreOrchestrate.git
cd RTXCoreOrchestrate
# Engine directory (note the quotes — spaces + emoji in folder names)
cd "Antigravity Ultimate (RTX⚡)"/RTX-Engine
# System health check — kaunsi CLI real chalegi, kaunsi mock par
node scripts/health-check.jsExpected output (is env mein):
⚠️ [claude_code ] claude NOT FOUND → mock mode
✅ [gh_cli ] gh gh version 2.102.0
⚠️ [copilot_cli ] github-copilot-cli NOT FOUND → mock mode
✅ [node ] node v24.20.0
✅ [git ] git git version 2.43.0
✅ [RTX-HealthCheck] All REQUIRED binaries present. Engine is go! 🚀
🧪 [RTX-HealthCheck] Mock mode CLIs: [claude_code, copilot_cli]
# Standard pipeline: planner → conductor → verifier → session
node core/rtx-conductor.js "Build a FastAPI backend with JWT authentication"
# ULTRA mode: parallel candidate racing + verifier best-selection + zyada turn budget
node core/rtx-conductor.js --ultra "Create a hello world REST API"
# AB-MCTS standalone: real tree search (koi API key nahi chahiye — rubric scoring)
node core/rtx-mcts.js --depth 2 --width 3 --budget 12 "Best architecture for a real-time trading system"
# Marlin: 3-layer research report apne notes/sources se
RTX_MARLIN_SOURCES=/path/to/your/notes node core/marlin-conductor.js "Your research question"
# Repo root se: eval harness (target ABSOLUTE path hona chahiye — relative path fail hota hai)
cd ../.. # wapas repo root par
node eval/rtx-eval.js --target "$PWD" --out /tmp/eval-results.json
# Smoke tests (repo root se; zero-dep, plain node:assert)
node test/smoke-mcts.js
node test/smoke-router.js
node test/smoke-safety.js
node test/smoke-verifier.js
node test/smoke-planner.js| Entry point | Command | Kya karta hai (verified) |
|---|---|---|
| Main pipeline | node core/rtx-conductor.js [--ultra] "<goal>" |
Plan → per-step TRINITY turn-loop (worker→verifier→revise, budget 2 standard / 4 ultra) → real adapter.run() spawn ya documented mock → credential check → sanitize → session log → COMPLETED ✅ |
| ULTRA racing | (upar wala, --ultra flag) |
requires_mcts steps par real AB-MCTS search (budget 8, worker pool = step.tool + fallback chain); hard non-MCTS steps par parallel candidate racing (Promise.all across router fallback chain, verifier winner chunta hai) |
| Research reports | node core/marlin-conductor.js "<goal>" |
5 stages: evidence gather (RTX_MARLIN_SOURCES dir / RTX_MARLIN_NOTES file / optional queryAdapter) → hypotheses → real ContradictionTracker → real AB-MCTS → 3-layer report (Verified Facts / Working Assumptions / Inferences), har line source-linked |
| Standalone search | node core/rtx-mcts.js --depth N --width M --budget K "<query>" |
Real AB-MCTS: tree nodes, UCB1 (c=1.414) selection, adaptive GO WIDER / GO DEEPER, backprop, topK() best-first. Defaults: --depth 1 --width 3 --budget 12. Bina LLM ke chalta hai — deterministic rubric scoring |
| Health diagnostics | node scripts/health-check.js |
5 CLIs scan karta hai (claude, gh, github-copilot-cli, node, git), health-report.json likhta hai, missing CLIs ko mock mode declare karta hai |
| Eval harness | node eval/rtx-eval.js --target <ABSOLUTE path> [--golden eval/golden-tasks.json] [--out results.json] |
Golden tasks ko actual engine modules se chala kar behaviour score karta hai. Honesty rule: jo measure nahi hua wo "unevaluated" report hota hai, kabhi 0 nahi. --target relative path (.) par fail hota hai (exit 1) — absolute path use karo |
| Smoke tests | node test/smoke-{mcts,router,safety,verifier,planner}.js |
Zero-dep assertions har core module par — sab exit 0 (verified 13 Oct 2026) |
| Python POC | python3 "Antigravity 2.0 (RTX⚡)"/mini_fugu_conductor.py "<task>" |
Lightweight Fugu-style orchestrator POC — sirf echo-based mocks, real execution nahi. argv-based execution (shell=True injection fix landed) |
Run karne ki jagah: engine commands Antigravity Ultimate (RTX⚡)/RTX-Engine/ ke andar se; eval/ aur test/ repo root se.
config/cli-registry.json mein sirf 3 CLIs hain — router inhi par route karta hai, inke bahar kuch nahi:
| ID | Binary | Type |
|---|---|---|
claude_code |
claude |
LLM coder |
gh_cli |
gh |
VCS (repo/PR/deploy) |
copilot_cli |
gh copilot |
LLM coder |
Note: GPT / Gemini registry mein nahin hain. README ke purane "Claude, GPT, Gemini" claim ko hata diya gaya hai.
...[truncated 8553 chars]