GeezerVison

Computer Use no longer slows the app it worked in and sends browsers to the MendCode extension, which is on in the full tier; every tool call keeps its own row in the transcript; Claude Code sessions keep your language after compacting; and sub-agents can start bare.

New

• Sub-agents can start bare: with none of your or the project's instructions, memories or skills, only the question, for a second opinion your conventions do not steer. Turn it on with /tier task_bare on and ask for a clean agent.
• A bare sub-agent on Claude Code also runs none of your Claude Code hooks.

Changed

• Computer Use sends Chrome, Brave, Edge, Arc and Chromium to the MendCode browser extension, and tells you how to install it.
• The browser extension is on in the full tier: ask the model to use your browser and it offers to install it. Elsewhere /tier browser_extension on turns it on.
• Claude Code sessions compact at 200K tokens even when compactAtTokens is set higher for long-window models; a per-model compactAt still raises it.
• The whole memory graph groups each category with a dotted line in its color, so memories without links no longer sit apart in a list.

Fixed

• Computer Use no longer leaves an app slowed after it ends: a session that goes quiet ends on its own and turns the app's accessibility mode off again.
• The Computer Use cursor stays where it last went, dimmed, instead of vanishing between steps.
• Every tool call keeps its own row in the transcript: reads, searches and commands no longer fold into Read 5 files or N tools more.
• Waiting for something on screen in Computer Use checks less often the longer it waits, so the app stays responsive.
• A PR review no longer picks up the AGENTS.md of a folder when it reads a file there.
• The code graph now covers a project's own git repository inside a folder that ignores it (a workspace holding the app's repo), so graph searches find that code instead of answering that nothing matches.
• Dream, /btw and memory extraction now wait out a busy or rate-limited provider and try again, instead of failing on the first overload.
• Usage counts now include the tokens Claude Code spends on compacting a conversation.
• Claude Code's own automatic compaction writes its summary in your language, so a long session no longer drifts into English after it.

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}