Agent Complex mode for complex tasks: A new agent mode for multi-step engineering work. It plans the task first, then carries out file and code operations, tracks progress against the plan, and reviews and wraps up when needed. It suits work that has to keep moving, such as refactoring, migrations, batch implementation and test fixes. Each chat remembers the mode you picked, so you can move between the lightweight Agent, Chat questions and Complex automation.
Global chat search: The box at the top is now a search entry point: click it or press Ctrl/Cmd + F. It searches chat titles, summaries and message text across all projects (global included) and leaves out internal detail such as system prompts, tool calls and tool results. Results are grouped by project, matches are highlighted, you can filter by project and time, and each group loads more on demand.
Subagents that run on their own, with a live window: Research subtasks the Agent hands out get their own identity, their own prompt and a restricted tool surface. The run appears in its own tab on the right, where reasoning, tool calls, reports and the final conclusion arrive as they happen, instead of the main chat showing only a condensed summary.
Upgraded built-in browser automation: Browser tools now drive the embedded browser you can see in the editor area, so the Agent works on the page in front of you. Opening, navigating, clicking, typing, reading text, running scripts and taking screenshots all behave consistently.
A better structured changelog center: Notes now read as a list of titles and dates on the left with the note itself on the right, which makes switching between versions clearer.
Core capabilities
Native page capture and DOM automation: Browser screenshots now capture the WebView page contents directly instead of cropping a desktop screenshot, so you get the real page even when the window is covered. Reading text and running scripts now return results over a native channel instead of navigating the page to send them back.
Smarter browser tool fallback: If a visible browser tab is already open in the editor area, the Agent can bind to that page for screenshots and text reads even when it did not open the browser itself this turn.
Live workspace refresh: After the Agent creates, changes, deletes, moves or copies a file, the file tree, open editors and Git status refresh right away instead of waiting for the stream to end or the window to regain focus.
Generated content saved as it arrives: During long tasks, content is saved to the chat as it is produced, so switching pages, refreshing the window or hitting an error no longer loses work.
Chat workspace kept alive: Switching top-level pages no longer unloads the chat area, which avoids blank screens when streaming Markdown rendering is cut off, and the state is still there when you come back.
Context compaction you can follow: Automatic compaction is now an observable step with clear states for started, finished and failed, instead of a silent background action. The trigger threshold is a share of each model's context window, which suits models with different context lengths.
Better judgement of task boundaries in Agent Complex: Building on the new Complex mode, this release sharpens how follow-up turns are judged. Complex no longer sends every follow-up through the planner: explanations, reviews and follow-ups go straight to an answer, while tasks that really need orchestration (refactoring, migrations, batch implementation, tests) still enter Complex planning.
Easier to review what happened: Subagent runs, LLM retries, context compaction and failed tool calls now show clearer notices in the interface and leave audit records for developers, so problems no longer hide in the background.
Large images and multimodal input: A single image can now be up to 5 MB instead of 1 MB, Pillow ships with the installed build, and image compression works on a clean machine too.
Stability and reliability
LLM timeout guard: Timeout budgets across the stack extend to 15 minutes, with classified errors, retry backoff and visible error frames, so long reasoning and long tool calls interrupt silently less often.
Claude tool-use truncation guard: Fixed Claude output that was truncated while generating large tool arguments being treated as a success. Incomplete tool calls, max_tokens cut-offs and upstream streams that end without a terminal event are recognized, and chunked writes can be corrected.
Compaction loop fixed: Fixed compaction repeating on every turn once it had fired. Older history is now replaced correctly, which avoids wasted LLM calls and a context that keeps growing.
Slow second turn in Complex fixed: Fixed a question asked after the first Complex run repeatedly hitting three timed-out retries before it answered. Planner input now trims large artifacts, and a failing helper step no longer inherits the long backoff of the main answer.
More reliable subtask results: Subagents must report through a structured report_result. Progress events, success and failure states and error types are recorded for review and troubleshooting.
Clearer browser tool errors: Requests made without an open browser, against an unavailable target or on an unsupported platform now return actionable messages instead of a bare 400 Bad Request.
More reliable screenshot saving: Screenshots land reliably in the current chat's output, so failed captures, lost results and images you cannot look at again are rarer.
Steadier local running: System proxies, network conditions and process cleanup interfere less with chats and tool calls, which cuts down "the model is not responding" and "the tool stopped" surprises.
Interface and interaction fixes
Context menu and refresh protection: The chat area and the main sidebar no longer show the WebView's default context menu, and F5 is caught so an accidental refresh no longer cuts the desktop app's current state.
Updated message bubbles: Your messages no longer carry a redundant time and button row at the bottom. When a turn has a plan, your message pins to the top like a task card and stays put while you scroll.
Clearer plan and to-do cards: Running PlanCard and TodoListCard views pin, collapse and show progress more sensibly, so the current goal and how far along it is are easier to follow in long tasks.
Accurate mode selection when editing: Editing a message and sending it again reuses the thread's mode choice instead of drifting because of local editing state.
Menus that fit the viewport: The model and mode pickers open upward or downward depending on the space available and are no longer clipped near the top of the window.
Fixed browser tab titles and page state going out of sync.
Fixed cases where a subtask or tool call looked finished but had not actually completed.
Fixed compaction summary requests being counted in the chat's own records.