This repository is under active development.
A custom Python runtime for durable, bounded AI-agent execution. Runveil is building toward a software-engineering agent that investigates a repository, proposes a patch, waits for human approval, and validates the change in an isolated environment. The project owns the execution loop, persistence, recovery, and authorization logic.
Now: Phase 12, private AWS deployment. The runtime and local workflows are implemented, including sandboxed coding tasks and read-only MCP tools. Terraform and container configuration are implemented through Phase 12B, with local and offline verification. Live AWS acceptance is pending; this is not a production service.
What works today:
- Durable execution with checkpoints, crash recovery, bounded retries, and limits on steps, time, tokens, estimated cost, and tool calls.
- Typed repository read/search tools, read-only MCP integration, and human approval bound to the exact proposed patch before controlled writes.
- Docker-isolated coding workflows across three controlled tasks: inspect, search, run failing tests, propose a repair, approve, apply, and validate.
- Local approval and trace consoles, execution telemetry, and an offline evaluation harness with paired statistical comparisons.
The web showcase at / and /runs presents recorded
public fixture runs. Local operator tools live at /approvals and /traces.
Scripted fixtures verify runtime behavior; they do not establish general coding-agent
or model quality. Phase 10 and 11 closure reviews remain pending.
Runtime, data & tools
Web
Development, testing & observability
Cloud infrastructure (configured; live acceptance pending)
Planned inference experiments
| Layer | In place | Remaining |
|---|---|---|
| Runtime & API | Python 3.12, Pydantic, FastAPI; custom agent loop and workers | Broader agent workflows beyond controlled fixtures |
| Persistence & delivery | PostgreSQL, SQLAlchemy, Alembic; SQS notification adapter | Live AWS delivery and recovery acceptance |
| Web | Next.js, React, TypeScript, Tailwind CSS | Public deployment |
| Models & tools | Scripted and hosted OpenAI-compatible providers, MCP Python SDK, Docker sandbox | Temporary GPU-hosted vLLM and inference benchmarks |
| Observability & evaluation | OpenTelemetry, persisted traces, offline evaluations and paired statistics | Real model-quality evaluation |
| Infrastructure | Terraform for private RDS, SQS, ECS/Fargate, ECR, S3 state, Secrets Manager, CloudWatch logs | Live provisioning, health, recovery, IAM and teardown checks |
| Quality & delivery | GitHub Actions, pytest, Ruff, mypy, Vitest, ESLint, Prettier; image and Terraform checks | Automated production delivery with AWS OIDC |
The AWS configuration currently targets a private read-only API and a fixed scripted worker workflow. Hosting the web app, models, MCP servers, and Docker sandbox workers is outside that deployment slice.
- Review Phase 12B and run bounded private AWS acceptance: migrations, TLS health, queue delivery, worker recovery, permission boundaries, and cleanup.
- Automate production delivery, then benchmark temporary self-hosted inference with vLLM.
- Complete security hardening and publish a live demo with reproducible evidence.
See the roadmap for review gates and the AWS runbook for the immediate next slice.
Start with the development guide for setup and verification. Local development uses Python/uv, Node.js/npm, and Docker Compose.