Bounded, verified work
Run simple single-file repository taskpacks or fine-grained operator-pinned controller pipelines inside native OS sandboxes with lease-fenced state and independent gate receipts.
Architecture
- Intake: Validate scope, author or compile TaskContract, and lock trusted configuration.
- State: Reserve authority in the fenced SQLite StateStore before physical dispatch.
- Execute: Run isolated Pi adapter with optional System-One decision advisory.
- Verify & Gate: Bind independent check evidence to candidate digests and enforce authenticated gates.
- Publish & Compact: Reconcile external effect receipts and compact reviewed memory into CAS stores.
Documentation Sections
Getting Started
Architecture foundations, two operating modes, runtime baseline, and installation.
Operations
Operate Loopwright across GitHub Actions, GitHub Apps, and local CLI environments.
Walkthroughs
Choose between zero-boilerplate repository taskpacks and fine-grained controller governance.
Configuration
Complete Pretty YAML schemas and field dictionaries for both operating modes.
Bootstrap or Migrate with One Prompt
Paste this prompt into any coding agent inside your repository to initialize .loopwright/taskpack.yaml or migrate an existing automatic synchronized workflow interactively via llms.txt:
Read https://docs.gpustack.ai/loopwright/llms.txt and fetch the Markdown guides for Declarative Taskpack, Taskpack, and GitHub Action. Inspect this repository (including existing automatic synchronized workflows, scripts, version tables, and directory layout), confirm my upstream sources, writable/protected scope, verification checks, and PR settings with me, then generate .loopwright/taskpack.yaml and .github/workflows/loopwright.yml (apiVersion: loopwright.dev/v1alpha1) without hardcoding secrets.Or install the laptop CLI globally: npm install -g --ignore-scripts @gpustack/loopwright — see Installation and Managed Controller.