Atlas Core

Standalone loop engine for routed, evaluated, retryable AI-assisted work.

Atlas Core is a bounded task loop. It observes context, chooses a route, plans the next action, executes, evaluates the result, retries when that helps, and stops when the answer is good enough.

observe → route → plan → execute → evaluate → retry / replan → finalize

It is intentionally separate from the MQ stack. MQ, GitHub, Obsidian, ChatGPT Skills, and local memory plug in as adapters rather than becoming the core.

Quick start

git clone https://github.com/MCamner/atlas-core.git
cd atlas-core
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
python -m unittest discover -s tests

Example output

A real run against this repository, abbreviated:

$ atlas run "granska atlas-core och hitta nästa bästa förbättring" --repo-path .

# Prompt Improvement

## Goal
granska atlas-core och hitta nästa bästa förbättring

## Recommendation
Gör prompten till ett tunt gränssnitt ovanpå en loop/state-machine.

## Next step
Skriv routes som data, inte som långa promptstycken.

---
Atlas route: prompt_improvement
Iterations: 1/2
Quality score: 0.9
Status: passed

Routes

Routes are data, not prompt prose. Seven ship in routes/routes.json:

• repo_review — observe, summarize, find gaps, prioritize
• architecture_decision — requirements, design, risk, options
• root_cause — symptoms vs causes, causal chain, actions
• decision_tradeoff — context, options, trade-offs, recommendation
• learning — plain explanation, common confusion, example
• prompt_improvement — diagnose, failure modes, rewrite, test cases
• general — understand, answer, next step