Generate images and short clips locally, and keep engine plugins current

The agent can now generate and edit images on your own GPU, produce short video clips, and install or update Unity, Godot and Unreal plugins against their newest released versions.

Featured

Local image generation on your own GPU

The agent generates images locally with stable-diffusion.cpp, running on your own hardware. Model downloads happen only after you approve them, the model storage folder can be moved, and finished images land in your project and are imported into Unity, Godot and Unreal. You can drive it from the Desktop image-models panel or the CLI /images command. (#759)

More image models, image editing and local video

Adds Qwen-Image and the SDXL family, image editing from reference images, and LoRA support. A new Video tool, off by default, generates short clips locally with LTX-2.5 or Wan 2.2. (#776)

Engine plugins install from the newest release

Installing or checking a Unity, Godot or Unreal plugin now targets the newest released plugin version instead of one fixed when the app was built. The built-in version is used only when the newest one cannot be looked up, for example offline, and the app says so. It never downgrades a plugin that is already newer, and the update prompt names exactly the version the Update button installs. (#777)

New

Agent questions and plan mode

The agent can ask you a question with choices, and plan mode lets it read and plan while every change is refused until it presents a plan. Plan mode is a per-session switch in the composer, and the plan review card offers Approve and run or Keep planning with feedback. (#749)

Fresh-start suggestion for a long session

When a long session is about to take on an unrelated new task, the composer offers to start fresh, which saves your limit. The offer appears only when the agent is idle and the context is large, and you can keep working in the current session instead. (#757)

Improved

Chat keeps your place while it works

Chat stops auto-following the moment you scroll up, and new messages and streaming text keep your reading position until you reach the bottom or use the scroll-to-bottom control. Tool and hook work folds into quiet disclosures while final answers stay visible, and one animated activity status replaces repeated status bubbles. (#744)

New installs start on the built-in claude agent

A fresh install now begins on the built-in claude agent instead of no agent, so the first session works without picking a persona first. Existing setups keep the agent they already selected. (#752)

Cleaner alerts with smoother animation

Alerts and the mid-chat model-switch confirmation gained a text-only ghost answer style, hover tooltips on every button including close, and smooth enter and exit animations that respect reduced-motion settings. (#756)

Permission and fresh-start prompts in the composer drawer

Tool permission requests and the fresh-start suggestion now appear in the same composer drawer the agent's questions use, instead of a transcript card and a popup. The session shows a waiting-on-you mark while a prompt is open and returns to running as soon as you answer. (#768)

Fixed

Accurate context and cache readouts

The composer now reports the model's real prompt and cache usage instead of early zeros, shows an unknown state rather than 0% when usage is missing, and counts cache warmth down from the reported retention. Very large streaming replies render as plain text so the window stays responsive, while completed messages keep their formatting. (#745)

Very large saved replies stay responsive

Reopening a saved session with an extremely long reply no longer freezes the window. Such replies render as plain text with every character and line break intact, while normal replies keep Markdown, highlighting and clickable file paths. (#746)

No more early compaction on large-context models

Models the bundled CLI does not know now receive the context window the server declares, so a million-token model no longer compacts around 167k. When the server lists no window, the app leaves the CLI's own conservative default in place rather than guessing. (#748)

Themed composer shortcut hint

The idle composer hint now reads "Ask anything, @tag files/folders, $use skills, or / for commands", with the @, $ and / triggers in the accent colour and the rest in muted text that follows the current theme. (#750)

An unusable built-in helper agent is blocked

The SDK's statusline-setup subagent is no longer offered or runnable, because it only writes to a personal settings file the app never reads. It disappears from the agent picker, and any saved selection of it falls back safely. (#747)

Question drawer placement and editable surveys

Agent questions now slide out from behind the input card without moving the transcript, and multiple questions become an editable survey: click any question to revisit it, restore earlier answers, skip ahead, and submit only once every question has an answer. (#751)

Test runs no longer touch real app settings

Test runs are guarded against writing to the real app settings directory, so an unrelated open copy of the app can no longer break the build. This changes test tooling only. (#753)

Quieter errors from panes whose session is not loaded

Panes that have not loaded their session no longer report session-state read failures, removing bursts of spurious error reports when you switch between the app and your engine editor. (#754)

Diagnostics for vague agent failures

Agent turn failures that used to arrive as a bare code now write one detailed local log line, with secret masking, so a subprocess exit, a dropped connection and an SDK error can be told apart. Agent behaviour is unchanged. (#761)

Failure reports carry enough context to diagnose

Failed agent turns now send masked error text, HTTP status, server request id, session id and model id with the report, so a failure can be matched to a specific request without guessing from timestamps. Secrets and home directory paths are scrubbed before anything is sent. (#767)

A file path in code renders as a single clickable chip

When a whole inline code span is a file path, it now renders as one clickable file chip instead of a path wrapped in a code box. Clicking, keyboard activation and hover all target the chip, while code that mixes text and paths looks the same as before. (#769)

Focused session row has the strongest contrast

On dark themes the focused session row now has the most text-versus-background contrast of any row, and open sessions are only a subtle step from closed ones. Light themes are unchanged. (#770)

Under the hood

• docs: migrate agent instructions to AGENTS.md (#771)
• docs: rename docs/claude/ to docs/agents/ (#774)