Skip to main content
LoopGuard is an alpha project. The package metadata and published PyPI release are 0.2.1, with Python 3.11+ required and classifiers for Python 3.11 and 3.12. These docs describe the repository implementation. Inspection of the published 0.2.1 wheel confirmed that Gemini embeddings/provider selection and controlled ReAct demo scenarios are not bundled in that release. The checkout also adds recommended_action and stop_reason to serialized decisions, along with clearer stop messages. Use a source installation for those features.

​
Available today

  • LangGraph/LangChain tool callbacks, with observe-only reporting or stop exceptions.
  • A wrapper for the supported live graph-update shapes.
  • Exact, semantic, progress, failure, cycle, handoff, and budget detectors.
  • JSON saved-trace analysis and CLI pass/fail checks.
  • A local FastAPI/WebSocket demo with a Next.js UI.
  • Offline tests and a small detector evaluation harness.
REPLAN and PAUSE are recommendations. Automatic replanning, persisted pause/resume, JavaScript SDK support, and a hosted LoopGuard runtime are not implemented here.

​
Directions to explore

The existing roadmap focuses on tuning thresholds with real traces, improving healthy retry/polling examples, and adding GitHub check annotations for trace failures. LangGraph.js/LangChain.js support remains an idea for later. These are possible next steps, not shipped features or delivery commitments.

​
Find your way around the source

​
Contribute

Open an issue with the smallest trace that reproduces a problem, the settings you used, and the decision you expected. For a code change, run the offline suite and include validation relevant to the behavior you changed. The contributing guide explains CI checks and the documentation review expected for developer-facing changes.