Dream-RSI SDK · Alpha

Bring your agent. Record its search. Evolve executable exploration policies.

An independent, unofficial Python SDK inspired by Dream-RSI research. Maintained by TheAstrayDev, who is not a Google or Google DeepMind employee. This is a personal research initiative, not a commercial Google development, official SDK or endorsed product.

Python 3.11+ · zero core third-party dependencies · Apache-2.0 · version 0.2.0a4.

Install

python -m pip install --upgrade dreamrsi==0.2.0a4

The package includes the Python API and the dreamrsi command. Publishing policy packages to GitHub additionally requires GitHub CLI (gh).

New in 0.2.0a4

The replay optimizer can synthesize a shorter PrefixPolicy without model requests: preserve the incumbent's original decisions and complete batches until every training world's full raw quality has been reached, then stop. Nested source policies, checkpoints and portable bundles are supported. Independent holdout validation is still required; recorded quality does not guarantee generalization. The original replay objective remains unchanged.

Disable proposals with DreamRSIConfig(optimizer_prefix_search=False). Developer eligibility now accounts for root-branch ordering when bounding recorded probe cost. Prefix bundles need 0.2.0a4 or newer; existing trees and older policy bundles remain readable. Start a new campaign for checkpoints created with the older runtime configuration. See the release notes and prefix guide.

Third-party policy and replay packages

Version 0.2.0a2 adds portable JSON bundles and GitHub-hosted package sharing. Community developers can distribute policy versions and recorded discovery trees; recipients import them into local SQLite memory with their own agent, quality metric, task-family contract and acceptance rules. These replay datasets contain observed search transitions, not model weights. Saved policies can avoid repeated development when they meet the recipient's quality floor; new tasks still run the recipient's agent.

dreamrsi list
dreamrsi install OWNER/REPOSITORY
dreamrsi save my-policy-pack --all
dreamrsi publish .dreamrsi/packages/exports/my-policy-pack.dreamrsi.json

Use a real package reference for OWNER/REPOSITORY. Export requires previously saved memory; publication requires GitHub CLI and creates a public repository. See the complete walkthrough for the diagram, executable integration example, and compatibility requirements.

Basic agent integration

from dreamrsi import Budget, DreamRSI

rsi = DreamRSI(
    agent=lambda task: task.upper(),
    evaluator=lambda answer: float(len(answer)),
    budget=Budget(model_calls=4),
)
result = rsi.run_sync("hello")
print(result.best)  # HELLO

This example demonstrates integration, not quality improvement. Stateful adapters support generate/evaluate/refine tasks. Replay uses recorded outcomes without new discovery calls. LLMPolicyDeveloper writes and iteratively rewrites executable source from measured feedback. Version 0.2.0a1 adds opt-in persistent champion reuse and more conservative replay cost accounting. It can replay already recorded runs without a new training call, delays holdout collection until a replay-improving candidate exists, and stops repeated policy revisions. The default promotion gate requires paired raw-quality and probe evidence; applications needing score-only decisions can explicitly use ReplayOnlyGate. An end-to-end cost advantage for an LLM discovery agent has not been proven. The SDK includes configurable Docker-free policy interpreters, optional process execution, held-out validation, SQLite recovery and reported token/USD accounting.

Two measured experiments

In the Bonsai Q2 experiment, a real local model wrote executable policy code while a deterministic agent solved the tasks. Across 64 fresh fixture tasks, counted operations fell from 256 to 248 with raw quality 0.9 throughout. This is a logical-operation proxy, not a measured token or dollar saving for an LLM agent.

In the GPT-6 Luna xhigh experiment, the real model both solved tasks and developed policies. Deployment fell from 24 to six model requests across six held-out tasks, but 94 preparation requests made the full path 100 versus 24. Mean reported score was slightly lower. This demonstrates learning and reuse, not an all-in economic win.

• Full documentation and roadmap
• Bonsai Q2 experiment
• GPT-6 Luna experiment
• Integration and recovery
• Sandbox configuration
• Issues
• Original Dream-RSI paper

Alpha APIs may change. Generated source runs in a bounded Python-syntax subset, not arbitrary Python. External state isolation and remote cancellation require adapter support. Usage caps depend on accurate provider reports and ceilings. Research-scale performance and broad generalization remain unverified.