Skip to content

SamDev1303/Thanos-Skill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

┌──────────────────────────────────────────────────────────────────────────┐
│                                                                          │
│  ████████╗██╗  ██╗ █████╗ ███╗   ██╗ ██████╗ ███████╗         │
│     ██╔══╝██║  ██║██╔══██╗████╗  ██║██╔═══██╗██╔════╝         │
│     ██║   ███████║███████║██╔██╗ ██║██║   ██║███████╗         │
│     ██║   ██╔══██║██╔══██║██║╚██╗██║██║   ██║╚════██║         │
│     ██║   ██║  ██║██║  ██║██║ ╚████║╚██████╔╝███████║         │
│     ╚═╝   ╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═══╝ ╚═════╝ ╚══════╝         │
│                                                                          │
│  T H E   I N F I N I T Y   A G E N T   H A R N E S S                    │
│                                                                          │
│  🟠 SOUL  🔴 REALITY  💜 POWER  🔵 TIME  🔷 SPACE  🟡 MIND               │
│                                                                          │
│  Autonomous. Self-healing. Ruthlessly correct.                           │
│  Works with: Claude • Codex • Gemini • Aider • any CLI agent          │
└──────────────────────────────────────────────────────────────────────────┘

CI License: MIT Works with Claude Works with Codex Works with Gemini


"I am inevitable."

Thanos is an autonomous agent harness that loops relentlessly until your goal is achieved with machine-verifiable proof. It combines four battle-tested patterns — /goals, GSD phases, the Ralph Loop, and Hermes self-healing — into a single CLI you can drop into any project.


⚡ One-Line Install

curl -fsSL https://raw.githubusercontent.com/SamDev1303/Thanos-Skill/main/install.sh | bash

Or with npm:

npx thanos-skill

🚀 Quick Start

# Open ANY project folder and run:
cd your-project
thanos

# Thanos will:
# 1. Scan EVERY file in your project
# 2. Ask you the REAL goal (not the surface ask)
# 3. Launch your installed CLI (Claude / Codex / Gemini)
# 4. Loop until machine-verifiable proof of completion
# 5. Self-heal and learn from every failure

Specific CLIs

thanos claude                        # Claude Code
thanos codex "fix the auth module"   # Codex with inline goal
thanos gemini                        # Gemini CLI
thanos aider                         # Aider

🌌 The Six Infinity Stones

Thanos tracks all state in six persistent Markdown files inside .thanos/. These survive crashes, context resets, and restarts.

.thanos/
├── SOUL.md     🟠  Active goal + stop condition + assumptions
├── REALITY.md  🔴  Every file in your project, read into agent context
├── POWER.md    💜  Test/build/lint exit codes + pass rates
├── TIME.md     🔵  Loop history + Hermes lessons (grows automatically)
├── SPACE.md    🔷  Phase queue + task backlog + blockers
└── MIND.md     🟡  Critic scores (0–100) + visual proof log
Stone Role Written by
🟠 SOUL The goal. The stop condition. The truth. Discuss phase (you + agent)
🔴 REALITY Full code scan — every file, every line Auto-scanned at launch
💜 POWER Machine proof — exit codes, test output Sub agent after each action
🔵 TIME Loop history + injected anti-rules Hermes after each loop
🔷 SPACE Phase queue, what’s done, what’s next Main agent during planning
🟡 MIND Critic scores. The hardest reviewer. Critic agent (smallest model)

🔄 The Thanos Loop

┌───────────────────────────────────────────────────────────────────┐
│                    THANOS LOOP                              │
└───────────────────────────────────────────────────────────────────┘

  ┌────────────────────────────────────────────────────────────┐
  │ 1. READ ALL 6 STONES   (SOUL•REALITY•POWER•TIME•SPACE•MIND) │
  └────────────────────────────────────────────────────────────┘
                              │
  ┌────────────────────────────────────────────────────────────┐
  │ 2. DISCUSS  🎯 Deep Q&A session — find the REAL goal         │
  │            One loop is never enough. Keep asking.          │
  └────────────────────────────────────────────────────────────┘
                              │
  ┌────────────────────────────────────────────────────────────┐
  │ 3. PLAN     📝 Write acceptance criteria BEFORE any code     │
  └────────────────────────────────────────────────────────────┘
                              │
  ┌────────────────────────────────────────────────────────────┐
  │ 4. EXECUTE  ⚔️ Sub agent builds. ONE action at a time.       │
  └────────────────────────────────────────────────────────────┘
                              │
  ┌────────────────────────────────────────────────────────────┐
  │ 5. VERIFY   💜 Run tests + lint + build → write POWER.md      │
  └────────────────────────────────────────────────────────────┘
                              │
  ┌────────────────────────────────────────────────────────────┐
  │ 6. CRITIQUE 🟡 Critic agent runs COLD. No loop history.       │
  │            Haiku/Flash-Lite/4o-mini. 6 scores. All ≥ 95.  │
  │            Visual proof required for any UI change.        │
  └────────────────────────────────────────────────────────────┘
                              │
  ┌────────────────────────────────────────────────────────────┐
  │ 7. LEARN    🧬 Hermes injects anti-rules into TIME.md        │
  │            Lessons accumulate. Quality only goes up.       │
  └────────────────────────────────────────────────────────────┘
                              │
              ┌───────────────────────────┐
              │  All scores ≥ 95?           │
              └───────────────────────────┘
              YES ↓                 NO ↩
         💥 SNAP ✔            Back to EXECUTE
         Done. It is           with issues list
         inevitable.           from critic

🧐 The Three Agents

Main Agent — The Architect

The largest, smartest model available. Reads all six stones, interprets critic verdicts, decides loop/snap/pause. Never writes production code directly. This is Thanos himself.

Platform Recommended Model
Claude claude-opus-4 or claude-sonnet-4-5-20251101
OpenAI o3 or o4-mini
Gemini gemini-2.5-pro

Sub Agent — The Builder

A capable mid-tier model. Writes code, runs shell commands, edits files, updates the Power Stone after every action. Executes ONE scoped task per cycle.

Platform Recommended Model
Claude claude-sonnet-4-5
OpenAI gpt-4.1
Gemini gemini-2.5-flash

Critic Agent — The Ruthless Reviewer

The smallest, fastest, cheapest model. Runs cold with zero loop history — sees only the current repository state. Its job is to find every flaw. A score below 95 means the loop continues. This is the most important agent in the system.

Platform Recommended Model
Claude claude-haiku-3-5
OpenAI gpt-4o-mini
Gemini gemini-2.0-flash-lite

Why the cheapest model as critic? It runs on EVERY loop. Speed and cost matter. And a cheaper model with no context bias gives you the most honest, unforgiving review. You want it to be hard to impress.


🧬 Hermes — Self-Healing & Self-Learning

Hermes is the self-improvement protocol embedded in Thanos. After every critic check that results in a LOOP AGAIN verdict, Hermes:

  1. Classifies the failure: logic / style / perf / security / visual
  2. Checks TIME.md — has this pattern appeared before?
  3. Injects an ANTI-RULE into THANOS.md in this format:
    [ANTI-RULE Loop N]: NEVER {action} because {consequence} — detected by critic in loop N
    
  4. Pattern matches — before the next execution cycle, the main agent reads all accumulated anti-rules and avoids repeat mistakes
  5. Escalates if the same issue appears 3+ consecutive loops: pauses and asks the human for help

The ANTI-RULES section of THANOS.md grows automatically as the system runs. Each project builds its own muscle memory.


📋 GSD Phase Commands

These commands can be used inside any running agent session:

Command Action
THANOS:discuss Force deep Q&A goal clarification
THANOS:plan Write phase plan to SPACE.md
THANOS:execute Run one sub-agent execution cycle
THANOS:verify Run verifier command, update POWER.md
THANOS:critique Spawn critic agent, score, write MIND.md
THANOS:learn Run Hermes — inject lessons from this loop
THANOS:snap Force-evaluate snap condition
THANOS:status Print all six stone summaries
THANOS:reset Clear all stones, start fresh
THANOS:pause Write pause state, exit cleanly
THANOS:resume Read pause state, continue from last checkpoint

💥 CLI Reference

thanos                              # auto-detect CLI, run discuss, launch
thanos claude                       # Claude Code
thanos claude "build a REST API"    # Claude with inline goal (skips discuss)
thanos codex                        # OpenAI Codex CLI
thanos codex "fix the auth module" 
thanos gemini                       # Gemini CLI
thanos aider                        # Aider
thanos --discuss                    # Run Q&A goal clarification only
thanos --status                     # Print all six stones
thanos --reset                      # Wipe .thanos/ and start fresh
thanos --resume                     # Resume from last checkpoint
thanos --test                       # Run E2E test suite
thanos --version                    # v2.0.0
thanos --help                       # Full usage

📊 Critic Scoring Rubric

All six scores must reach ≥ 95 before Thanos snaps. This is not negotiable.

Category Weight What the critic checks
Logic Correctness 20% Does it actually do what was asked?
Code Quality 20% DRY, patterns, no anti-patterns, readability
Test Coverage 20% Edge cases covered, tests are meaningful
Visual/UI Proof 20% Screenshot/DOM snapshot required for UI changes
Security 10% OWASP Top 10, no secrets, no injection
Performance 10% No N+1, no blocking I/O, no memory leaks

The critic runs cold — no previous loop context. It reads only the current repository state. This is the Ralph Loop’s most important rule: never let the agent grade its own work from memory.


🔮 The Four Inspirations

/goals (OpenAI Codex)

Codex’s /goal command attaches a persistent objective to the session. It injects goals/continuation.md at the end of every turn to decide whether to loop again, and enforces a token budget via goals/budget_limit.md. Thanos extends this with six-stone state, multi-model orchestration, and Hermes self-learning.

GSD — Get Shit Done

GSD solves context rot by breaking work into scoped phases with their own specs, assumption lists, and acceptance criteria. The flow is always: Discuss → Assumptions → Plan → Execute → Verify → Critique. Thanos uses GSD as the backbone of every loop.

The Ralph Loop

The Ralph Loop wipes conversation context between iterations so the AI cannot grade its own work from memory. Instead it uses the filesystem, git, and objective machine signals (test exit codes, coverage %, build output) as the only truth. The critic agent in Thanos IS the Ralph Loop — it runs cold every single time.

Hermes (Self-Learning Agent)

Hermes is an agent architecture built for self-improvement. It finds bad patterns, writes them as anti-rules, and reads them back before the next execution cycle. Thanos’ TIME stone and the ANTI-RULES section of THANOS.md implement the Hermes protocol. Quality only ever goes up.


💾 What Gets Installed

~/.local/bin/thanos                    ← global CLI command
~/.agents/skills/thanos/THANOS.md      ← Codex skill loader path
~/.claude/skills/thanos/THANOS.md      ← Claude Code skill path  
~/.codex/skills/thanos/THANOS.md       ← Codex instructions path

📁 Repository Structure

Thanos-Skill/
├── thanos.sh          ← Main CLI harness (this is what you run)
├── THANOS.md          ← Agent skill file (embedded into agent context)
├── GAUNTLET.md        ← Deep architecture + philosophy reference
├── SKILL.md           ← Agent loader frontmatter
├── install.sh         ← One-line global install
├── package.json       ← npx thanos-skill support
├── templates/         ← Stone file templates
├── tests/             ← E2E test suite
└── docs/              ← Architecture diagrams

🏁 The Snap Condition

Thanos snaps when ALL seven boxes are checked:

  • [ ] SOUL.md stop condition verified: every criterion met
  • [ ] POWER.md: all verifier commands exit 0
  • [ ] MIND.md: critic score ≥ 95 on ALL six categories
  • [ ] MIND.md: visual proof exists for any UI changes
  • [ ] TIME.md: this loop’s lesson written
  • [ ] SPACE.md: task queue empty
  • [ ] REALITY.md: file state matches what was planned

Only when all seven boxes are checked: 💥 SNAP. It is done. It is inevitable.


🔭 The Vision

Today’s AI coding tools are reactive. You ask, they answer. You find the bug, they fix it. You notice the security hole, they patch it.

Thanos is proactive. It reads every file in your project before you say a word. It asks the real question, not the surface question. It loops until the machine says it’s done — not until the human thinks it looks done.

The long-term vision:

  • Any folder, any language, any agentthanos works everywhere
  • Self-improving skill files — ANTI-RULES accumulate across projects, shared as community knowledge
  • Multi-repo orchestration — Thanos coordinates across services, not just files
  • Visual regression built in — screenshot diffing as a first-class MIND stone metric
  • Cost dashboard — track tokens spent per loop, per stone, per project
  • One command to rule them all:
curl -fsSL https://raw.githubusercontent.com/SamDev1303/Thanos-Skill/main/install.sh | bash
# then:
thanos

🤝 Contributing

Pull requests welcome. The only requirement: your PR must pass thanos --test before you submit it. The critic is ruthless. Impress it.


📄 License

MIT — use it, fork it, make it yours.


"You could not live with your own failure. Where did that bring you? Back to me."

— Thanos, on every loop that leads back to the EXECUTE phase

About

⚡ Thanos Skill — Autonomous AI agent skill for any CLI. Six Infinity Files track state, progress, goals and rules. Works with Claude Code, Codex CLI, Gemini CLI, OpenCode. Snap and it's done.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages