kronumos 0.1.0

"You write the features. Kronumos heals the bugs."

📌 What is Kronumos?

Kronumos is a specialized autonomous software engineering agent, fine-tuned specifically for end-to-end bug remediation and automated self-healing — not generic chatbot coding.

Generic AI coding assistants try to do everything: generating unverified applications, hallucinating missing functions, and bloating context with massive raw runtime logs.

Kronumos is scoped narrower and deeper. It handles a single, closed-loop engineering workflow:

1. Diagnose runtime failures & test crashes from raw error logs.
2. Sub-Cortex Token Surgery: Excise framework noise and redact sensitive credentials (JWT, AWS, DB keys) using Tokenectomy Rust engine (91.3% token reduction).
3. Analyze Blast Radius: Map caller dependency graphs before applying edits.
4. Synthesize & Verify Atomic Patch: Apply search-and-replace AST patches with zero dirty diffs.
5. Git Delivery: Automatically branch, commit, open Pull Requests, and close incident tracking issues.

📄 Academic Paper & Preprint

Read the formal preprint paper on Research Square (Springer Nature): DOI: 10.21203/rs.3.rs-11205335/v1 or Zenodo: DOI: 10.5281/zenodo.22929676, read the Technical Report, or view the publication-ready LaTeX source in paper/main.tex:

"Kronumos 2 Kairos: Cost-Bounded Automated Program Repair via Dual-Brain Cybernetic Sub-Cortex on SWE-bench Verified"
Author: Muhammad Naufal Daffa (ORCID: 0009-0000-7909-4916)
Research Square (Springer Nature): 10.21203/rs.3.rs-11205335/v1
Zenodo Archive: 10.5281/zenodo.22929676

🗺️ Product & Engineering Roadmap

Read the complete engineering roadmap and evaluation methodology in kronumos_product.roadmap.md, or explore the raw 16-hour sprint build log and incident postmortem in JOURNEY.md.

📊 Empirical Benchmark Results (SWE-bench Verified)

Read the official full 500-instance benchmark report in BENCHMARK_500_REPORT.md.

• Evaluated Tasks: 500 / 500 tasks (475 synthesized, 25 structural safe refusals)
• Officially Resolved Tasks (Pass@1): 12 Resolved (15.0% candidate precision on Docker-evaluated tasks, 2.4% full-benchmark lower bound [1.4%, 4.1%])
• Verified Ecosystems: Solved production issues in Django (7), Scikit-Learn (2), Pytest (1), PyData Xarray (1), and SymPy (1)
• GNU Patch Validity: 100.0% clean application rate (0 hunk errors on evaluated instances)
• Average Turns to Remediation: 2.34 turns
• Average Token Consumption: 3,361.0 tokens / task (91.3% token bloat reduction vs raw context)
• Average Remediation Latency: 43.6 seconds / task
• Marginal Inference Cost: $0.00 (Self-Hosted / Cloudflare Edge)

🛠️ Tool Schema (Agent Sub-Cortex)

Kronumos is fine-tuned to emit structured JSON tool calls:

Tool	Purpose
get_error_context	Excise framework noise, redact credentials, and extract exact offending code snippets from raw logs
apply_code_patch	Apply an atomic search-and-replace AST patch, verified before commit
inspect_docker	Diagnose container crashes (e.g. exit code 137 OOMKilled) via logs and resource stats
probe_database	Triage connection pool starvation and lock deadlocks
sentinel_analyze_blast_radius	Map caller dependency graph for a symbol/file before applying a patch
create_fix_branch	Create a new isolated git branch for the fix
commit_fix	Commit the verified patch to the fix branch
open_pull_request	Open a PR from the fix branch to the target branch
create_incident_issue	Open a tracking issue for a diagnosed incident
link_issue_to_fix_pr	Link an existing incident issue to its resolving PR
close_incident_issue	Close the incident issue once verified and merged

💻 Interactive CLI Agent (Transparent Terminal REPL)

Kronumos includes a native agentic CLI interface featuring an ambient, transparent-terminal aesthetic, live streaming tokens, and an autonomous repair loop:

    ██╗  ██╗██████╗  ██████╗ ███╗   ██╗██╗   ██╗███╗   ███╗ ██████╗ ███████╗
    ██║ ██╔╝██╔══██╗██╔═══██╗████╗  ██║██║   ██║████╗ ████║██╔═══██╗██╔════╝
    █████╔╝ ██████╔╝██║   ██║██╔██╗ ██║██║   ██║██╔████╔██║██║   ██║███████╗
    ██╔═██╗ ██╔══██╗██║   ██║██║╚██╗██║██║   ██║██║╚██╔╝██║██║   ██║╚════██║
    ██║  ██╗██║  ██║╚██████╔╝██║ ╚████║╚██████╔╝██║ ╚═╝ ██║╚██████╔╝███████║
    ╚═╝  ╚═╝╚═╝  ╚═╝ ╚═════╝ ╚═╝  ╚═══╝ ╚═════╝ ╚═╝     ╚═╝ ╚═════╝ ╚══════╝

                     · K R O N U M O S   K A I R O S ·
               Autonomous Code Remediation & SRE Agent • v1.0

Interactive Agent Commands:
  Any text     chat with Kronumos or explain code/errors
  /fix         autonomous diagnostics & repair loop
  /diff        inspect git diff in workspace
  /test        run test suite on physical hardware
  /doctor      run diagnostic environment & gateway health check
  /stats       display FinOps token savings & telemetry
  /undo        revert uncommitted patches (zero dirty diff)
  /clear       clear conversation memory buffer
  /help        display help and shortcuts
  /exit        exit Kronumos cleanly

⚡ kronumos ❯ 

Installation

# 1-Line Cryptographic Installer (Linux x86_64, with automated SHA256 verification):
curl -fsSL https://raw.githubusercontent.com/Tokenectomy-Labs/Kronomus/main/cli/install.sh | bash

# Or download pre-compiled standalone release binaries directly:
# https://github.com/Tokenectomy-Labs/Kronomus/releases/latest

Usage Modes

# 1. Autonomous TDD self-healing loop (runs test suite, patches, verifies on hardware):
kronumos --fix
kronumos --fix -t "pytest tests/test_auth.py"  # Custom test command override
kronumos --fix --auto-rollback                  # Auto-revert broken patches on failure (0 dirty diff)
kronumos --fix --branch fix/auth --commit      # Auto Git branch & commit on pass
kronumos --fix --pr                            # End-to-end Git delivery: auto-branch, commit, & open Pull Request
kronumos --fix --json                          # Machine-to-machine JSON output for CI/CD

# 2. Diagnostic environment & edge connectivity check:
kronumos --doctor
kronumos --doctor --json                       # Structured diagnostic output for CI/CD

# 3. One-shot terminal command execution (headless):
kronumos "Explain the blast radius of refactoring auth module"

# 4. Unix piping (clean quiet mode for CI/CD or log diagnosis):
cat error.log | kronumos -q
pytest 2>&1 | kronumos -q "Diagnose and fix assertions"

# 5. Interactive ambient terminal REPL:
kronumos

# 6. Multi-Backend inference (Cloudflare Edge, Local Ollama, OpenAI/Groq):
kronumos --backend cloudflare                 # Hosted edge gateway (zero-config, free tier)
kronumos --backend ollama --ollama-model kronumos/Kronumos-2-kairos
kronumos --backend openai --openai-key $GROQ_API_KEY --openai-model llama-3.3-70b-versatile

GitHub Actions CI/CD Integration

Automate bug remediation directly inside your repository pipelines with the official GitHub Action:

name: "Autonomous Remediation"
on: [pull_request, workflow_dispatch]

jobs:
  self-heal:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Auto-Remediate Test Failures
        uses: Tokenectomy-Labs/Kronomus@main
        with:
          test-cmd: "pytest tests/"
          auto-rollback: "true"
          commit: "true"

🧪 Live Bug Arenas & Reproducible Benchmarks

Explore standalone, authentic bug arenas designed to test autonomous repair agents on physical hardware without mocks or simulations:

• Linux Kernel Circular Buffer (kfifo) Boundary Defect: Authentic off-by-one boundary defect (> vs >=) causing ring buffer silent byte corruption. Resolved by Kronumos in 18 seconds (round 4).
• CVE Memory Lifecycle Defect (AddressSanitizer Heap-Use-After-Free): Authentic systems security bug under GCC -fsanitize=address -fsanitize=undefined modeled after network daemon CVE patterns (Redis/libcurl). Naive LLMs trigger LeakSanitizer aborts; Kronumos synthesizes a surgical memory-safe patch in round 1.

☁️ Zero-Cost Cloudflare Edge Gateway ($0 Serverless)

Deploy Kronumos on Cloudflare's global edge network in under 60 seconds with $0 server infrastructure cost using Cloudflare Workers AI:

# 1. Navigate to the gateway directory:
cd gateway

# 2. Deploy directly to the edge:
npm run deploy

• Free Tier Allocation: 10,000 Neurons per day (~500+ debugging turns/day for free).
• Global Edge Acceleration: Runs on Cloudflare edge GPUs across 300+ datacenters worldwide.
• Architecture Details: See gateway/README.md.

🏠 Offline Local Inference via Ollama

Run Kronumos 100% locally and offline with GGUF quantization:

ollama run kronumos/Kronumos-2-kairos

🐍 Python SDK (Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Kronumos/Kronumos-Kairos-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

🏢 Organization, Inquiries & Community

• Author & Research Lead: Muhammad Naufal Daffa (ORCID: 0009-0000-7909-4916)
• Hugging Face Organization: huggingface.co/Kronumos
• Titan Flagship (671B MoE): Kronumos Aion
• Edge Workstation (7B / 14B): Kronumos Kairos v2 | GGUF
• Official Inquiries: kronumos@proton.me
• Research Publication & Substack: https://kronumos.substack.com
• Ollama Hub Registry: https://ollama.com/kronumos/Kronumos-2-kairos
• License: Dual License (GNU Affero General Public License v3 / Commercial Enterprise) — see LICENSE
• Terms of Service: TERMS.md
• Privacy & Zero-Retention Policy: PRIVACY.md
• Security Policy: SECURITY.md

📖 Citation

If you use Kronumos in your research or benchmarks, please cite our latest preprint:

@article{daffa2026kronumos2,
  author    = {Muhammad Naufal Daffa},
  title     = {Kronumos 2 Kairos: Cost-Bounded Automated Program Repair via Dual-Brain Cybernetic Sub-Cortex on SWE-bench Verified},
  journal   = {Research Square},
  year      = {2026},
  doi       = {10.21203/rs.3.rs-11205335/v1},
  url       = {https://doi.org/10.21203/rs.3.rs-11205335/v1}
}