This repository is for autonomous reliability research, not generic feature work.
Improve the reliability, repeatability, and debuggability of Windows desktop automation on a real machine.
- Reproduce one real failure or run one benchmark.
- Record the observed behavior.
- Patch the narrowest cause.
- Add or update tests when possible.
- Retest locally.
- Retest live if the real MCP path changed.
- Update only the minimum research files needed for handoff.
Do not make strong success claims without a retest.
- reproduced bugs on the real machine
- missing postcondition verification
- repeatable benchmarks for known flaky workflows
- stronger non-UI or semantic routing
- diagnostics that unblock validation
- Do not trust tool success strings alone.
- Prefer
Snapshot, filesystem, clipboard, process, shell, or DOM checks. - Keep blind UI action chains to
1. - If the same workflow fails three times without a system change, stop and classify it.
- Prefer stronger routes when they exist:
- DOM over browser chrome clicking
- protocol URIs over click-navigation
- filesystem/process/shell checks over title guessing
research/failure_taxonomy.md: confirmed failure classes and examplesresearch/test_matrix.md: benchmark definitions and latest statusresearch/patches.md: implemented and planned changesresearch/next_session.md: short handoff onlyresearch/results/*.md: benchmark or repro evidenceresearch/notes.md: short summary, not a second source of truth
Do not repeat the same fact across all of them.
- Prefer a few coherent commits over checkpoint spam.
- Bundle patch, test, and verification when they belong together.
- Avoid repeated
Log ...commits with no new code or benchmark value.