aios-long-running-harness

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.

Progress self-report and replan cadence

Stuckness is self-reported by you in structured observations, never judged by a second model call and never guessed from message keywords. After every Execute/Verify round, emit:

• progress_made: true|false — true only when the round produced new evidence, a new artifact, or a state change versus the previous round. Repeating the same tool call with the same result is false.
• blocked_reason: <empty|one line> — empty when progress was made; otherwise the concrete blocker (selector gone, auth lost, tool error, waiting on input).

Rules:

• The harness counts consecutive progress_made: false rounds (threshold 3 unless the run preflight set another). At threshold, stop repeating: run Recover with a changed hypothesis or escalate to the user. Never run the identical action a fourth time hoping for a different result.
• An identical tool call issued twice in a row is flagged immediately as no-progress, regardless of the self-report.
• Every N steps (planning_interval, default 5, set in preflight), re-emit the frontier: ready work vs. blocked work with reason and evidence, and replan the remaining steps from evidence. A replan is a checkpointed event, not a silent course change.