Skip to content

v0.38.0

Latest

Choose a tag to compare

@github-actions github-actions released this 14 Aug 07:38
b5b5de1

Features

  • Reliable Computer Use loop — Desktop control now follows an enforced Observe → Act → Verify cycle. Every writable action consumes a short-lived, single-use observation state bound to the target app and process, then requires a fresh observation before another action.
  • Guided permissions and recovery — Abu requests only the macOS permissions required by the current task, explains when a relaunch is needed, and can return to the original conversation without reusing an old authorization or accessibility session.
  • Clear model capability modes — Models are shown as full, structured, unsupported, or unknown. Tool-capable models without image input, including supported DeepSeek configurations, can use structured accessibility data without being presented as visually capable.

Reliability and Safety

  • Consequential actions are classified again by the Electron host and require a one-attempt confirmation; an ambiguous native result stops the task instead of retrying a possible side effect.
  • Stale observations, target-process changes, helper restarts, renderer reloads, and duplicate writes fail closed. Windows desktop control also serializes writes across approval and native-input paths.
  • Repeated no-change observations allow one bounded recovery attempt, then stop with an explicit result rather than looping indefinitely.

Diagnostics

  • The capability page and task status now show the active target, capability mode, permission state, and Observe/Act/Verify phase.
  • Local runtime traces correlate the conversation, loop, tool call, observation state, helper generation, and verification result through an allowlist that excludes prompts, screenshots, accessibility labels, user input, and tool-result bodies.

Full Changelog: v0.37.4...v0.38.0