AIOS Long-Running Harness

Overview

Use this harness to keep long tasks stable under UI drift, model variability, and partial failures. It maps Anthropic's long-running-agent harness ideas into this repository's file-based workflow.

Harness Loop

1. Preflight: lock objective, stop conditions, budgets, and required artifacts.
2. Plan: split into idempotent steps with explicit success/failure evidence.
3. Execute: run one step at a time with tool output capture.
4. Verify: assert completion from page evidence, not assumptions.
5. Checkpoint: persist current state, artifacts, and next action.
6. Recover: on failure, classify and retry only with a changed hypothesis.
7. Complete: run final verification and write summary doc.

Pairing with Superpowers Skills

• Plan step should be produced through superpowers:writing-plans (or superpowers:brainstorming first when scope is unclear).
• For 2+ independent domains, use superpowers:dispatching-parallel-agents; for coupled domains, run sequentially.
• If the runtime has no true subagent tool, emulate dispatch with explicit per-domain task queues and only parallelize safe independent reads/checks.
• Always finish with superpowers:verification-before-completion before claiming run success.