multivrs

Automated multiverse analysis powered by LLM analyst agents. The CLI scaffolds an OpenClaw-inspired architecture for spawning independent analysts, collecting normalized outputs, and producing machine-readable run artifacts.

Install

npm install
npm run build

Quick start

# Generate starter config + local run directory
npm run dev -- setup

# Validate config
npm run dev -- config validate

# Import Claude/Codex credentials if present
npm run dev -- models auth import --provider all

# Run multiverse harness scaffold (no code execution by default)
npm run dev -- harness run --analysts 3

# Run a single analyst
npm run dev -- agent run --question-id q1

Core commands

• multivrs setup
• multivrs status
• multivrs config validate
• multivrs models auth list|import|select|clear|login
• multivrs agent run
• multivrs harness run

Config

Primary config file is multivrs.yml (see examples/multivrs.yml).

Key sections:

• project: metadata
• dataset: dataset + codebook paths
• questions: analysis prompts/hypotheses
• harness: analyst count, runner, execution mode
• providers: default provider (claude-cli or codex-cli)
• outputs: run artifact root directory

Auth model

Profiles are stored in ~/.config/multivrs/auth-profiles.json.

Credential precedence:

1. Explicit profile (--profile)
2. Active profile for provider
3. Environment fallback (ANTHROPIC_API_KEY, OPENAI_CODEX_API_KEY, OPENAI_API_KEY)

External import support:

• ~/.claude/.credentials.json
• ~/.codex/auth.json

Artifact layout

Each run writes under outputs.rootDir/<run-id>/:

• run.json
• summary.json
• events.jsonl
• analysts/<analyst-id>/analysis-plan.md
• analysts/<analyst-id>/analysis.py|analysis.R
• analysts/<analyst-id>/result.json

Test

npm test