Skip to content

v0.2.0-beta

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 19 Aug 16:02
· 3 commits to master since this release

v0.2.0-beta

Features

  • New subagent roster — the agent can now dispatch work to six specialist subagents: code-reviewer, commit-writer, pr-creator, security-reviewer, simplifier, and web-research. Agents declare their dispatch capabilities in a catalog the model can see, so parallel investigation and review requests route to the right specialist.
  • Background dispatch — subagents can run in the background while you keep chatting. Dispatches persist across app restarts, can be resumed with follow-up instructions, and can be aborted mid-run. Ships behind an experimental toggle in Settings.
  • Read-only git_repo tool — inspect any remote or local repository without cloning it: browse files at a ref, read file contents, view logs, diffs, and blame. Reference research no longer pollutes your workspace.
  • Notification sounds — a soft chime when a turn finishes, a brighter ping when a permission prompt is waiting, and a low tone when a turn ends in an error or limit. Toggleable in Settings → General, independent of OS notifications.
  • Skill catalog & live todo plan — the skill list is now rendered into the model's tool catalog, and your todo plan is included in the system prompt so the agent tracks progress against it.
  • Chat polish — new pixel-grid loading animation, redesigned chat message blocks, and streaming render tuning for smoother output during long turns.

Fixes

  • Protocol compatibility — thinking-effort and max-token parameters are now handled correctly for non-native endpoints (Anthropic-protocol thinking allowlist, OpenAI effort clamping), fixing errors on third-party providers.
  • Malformed follow-up questions — GLM-family models occasionally emit invalid ask_followup arguments; these are now repaired on the fly instead of failing the tool call.
  • RAG model setup — workspace setup now requires an explicit embedding-model download instead of proceeding with a missing index, dev builds lazy-download the model exactly like production, and the vendored model no longer bloats dev launches.
  • UI — chat no longer jumps while streaming when scrolled up, duplicated sidebar buttons removed, and the File Viewer and Commit sheets clamp to the viewport width instead of overflowing.
  • Session titles and compaction summaries are more accurate after prompt refinements.

Full changelog: v0.1.4-beta...v0.2.0-beta