gezel Gezel Help

1.26317 — 13 November 2026

On a model where it was shown to help, a report, document or spec now gets one more read against your request before it is handed in. Moving around the app is steadier: a project you come back to is where you left it, and views appear whole instead of filling in piece by piece. Projects can wear an icon of their own, drawn by a model or uploaded. And Help gains a scorecard for how well each local model does document work.

If you read only one thing: look at Settings → Artificial intelligence → Check work before handing it in.

A second look before work is handed in

A plan's checks confirm that the work exists and has the parts it should. They can't tell whether a report actually answers what you asked. So once a report, document or spec passes its checks, the crew member who wrote it now reads it again against your request and fixes what it can show is wrong. A fix may only remove or correct, or add what the sources say. The checks then run again, and there is one round.

This happens only where it was measured to help. In our test with Qwen 3.8 27B, the reviewed reports were judged better than the same reports before the review, none got worse, and the review added a median of about a minute and a half. On Gemma 4 E4B it made no difference, so models like it skip the review, as do code, data files and slide decks, where a review mostly re-edited work the checks had already verified. Today that means the Q4 build of Qwen 3.8 27B, on plans it runs as one piece of work.

A review can't cost you work that had already passed. If the revised work fails the checks and the task would otherwise pause, the work goes back to how it stood before the review, and the task finishes on that version as long as it still passes. The revision is kept with the review's records.

While the review runs, Activity shows the work as being checked. You can turn it off under Settings → Artificial intelligence → Check work before handing it in. A change there applies to work you start afterwards, never to work already under way.

Views that keep your place

Moving between parts of the app used to reset what you were looking at and draw the next view in stages. Now:

  • Coming back is coming back. Home and the last few places you visited stay ready behind the one you are looking at. Return to a project and it is on the tab, the open file and the scroll position you left, not back on Chat. Settings and the editors still start fresh each time.

  • Views appear whole. The navigation highlights your choice at once. When the next view's information is not ready yet, the current one stays up for a moment, at most about a tenth of a second, so the next one appears complete instead of popping in. A project's tabs no longer blank while the next one loads.

  • Long chats stay smooth. A reply streaming in a chat you are not looking at waits to redraw until you open it, and long conversations redraw less while a reply arrives.

  • The window has less to load at startup. Formatting for search results, for example, now loads when you first use search.

Your own project icons

Two projects of the same type used to look alike in the navigation: two code folders were both braces. A project's Settings → Preferences → Icon now offers every built-in mark to choose from, plus two ways to make one of your own:

  • Draw one… asks a model for a small one-color mark from a description, which starts from the project's name and type and is yours to change. Generate as many as you like; nothing is saved until you choose Set this as the icon.

  • Upload a picture… takes a PNG, JPEG or WebP and scales it down to a small square.

A drawn mark takes the colors of your theme, so it sits beside the built-in marks in light and dark. Automatic goes back to the mark for the project's type. Drawing never happens on its own: it costs a model call, and a small on-device model draws poorly.

In the navigation, project marks are now drawn in ink, like a document's file icon. In the accent color, a project row looked like one more group heading.

A scorecard for document work

The model scorecard asks whether a model can finish a job on its own. The new Docbench article in Help asks how good a model's answer is when DocBlocks hands it one piece of document work: a review, new or rewritten text, a diagram or illustration, slides, study aids, or a cleaned-up dictation. Each model gets a score from 0 to 100, with a range and a band that says which models are too close to tell apart. The first round covers 21 local models, with Qwen 3.8 27B and Gemma 4 12B at the top.

Apps connected to Gezel can now read these scores from its model list, so DocBlocks can suggest the best measured model your computer can run.

Smaller things

  • When a crew member asks another to make a picture, the picture now appears in the chat you are in, beside the reply. Before, it showed only in the other crew member's thread, and the reply could link to a picture that didn't exist.

  • The dark theme no longer turns the chat brown. Your own messages, the message box, Home and hover highlights use a neutral gray at night. The accent color stays on buttons, selections, links and anything that is working.

  • The number of prompt tokens beside a reply's progress bar appears only when you point at the reply's header, unless debug mode is on.

  • A presentation, document, spreadsheet or PDF that the side pane in a chat can't preview now says what it is, instead of showing <Machine File>.

  • In the memory details you open from This PC in the title bar, the list of programs using video memory folds into one line, Gezel against everything else. Open it for the full list.

  • Two video requests now take turns. The second one shows "Waiting for another video to finish…" instead of stopping a long render already running on the other video engine.

  • A plan's reference list leaves out catalog entries that share only one word with your request. "For the board" no longer cites a board game.

  • An app that asks a local model for a structured answer gets one even from a model the engine can't hold to the format, such as Nanbeige 4.2 3B. Before, every such request to that model failed.

  • GPT-OSS 20B now skips its reasoning when an app asks it to. It used to reason anyway and run out of room mid-answer.

  • When a model you added fails its tryout because it wrote its tool call as plain text, the result says so, instead of "fitness check failed". A small model stuck that way no longer fills the reply with junk until it is stopped.

  • Python packages downloaded for on-device engines are now kept inside Gezel's own folder, where storage in Settings and gezel cleanup count them. They used to go to ~/.cache/uv, which neither could see.

  • When an app's own setup command fails while Gezel adds its tools to that app, such as claude mcp add, Settings → Integrations shows what that command said, instead of only "Gezel could not change this app".

  • Model downloads from Hugging Face hold up on an unsteady connection. A dropped connection now picks up where it stopped instead of throwing away hundreds of megabytes already received, the progress bar no longer jumps back, and a few drops in a row no longer end the download.

  • The Use this model link on Hugging Face now fills the model in under Add a model… and waits for you to choose Check model. Any web page can send that link, so Gezel doesn't contact Hugging Face until you ask.

  • Adding an MLX model now also checks its video settings for program code, and turns the model away when a settings file is too large to check, instead of letting it through unchecked.

  • The tryout after a model download now waits until no other model has worked for three minutes. It had unloaded a model a task was still using, between two of the task's turns.

  • An on-device model that gets stuck repeating itself now has only that reply stopped, and the engine keeps serving your other chats and tasks. A second time on the same engine within half an hour restarts it.

  • When a crew member opens several files at once, an Office document among them is now converted to text, as it is when opened alone. A review had reported a finished deck as one that would not open.

  • A plan's citation check no longer counts a report that cites itself as a source.

  • After making a picture, a crew member is now told how to animate it when an installed video model can. A permission check had kept that hint from ever appearing.

  • When a video render fails on a paired server, the message no longer shows file paths from that computer. It gives a request number to look up in that server's log.

  • An app update keeps a schedule time you changed. It used to reset the time to the one in the new version.

  • An app in an archived or read-only project, or one shared by every account on this computer, now offers its update for you to apply. Its update line used to say "An update is being applied." and never finish.

  • In a narrow window, a project's Output page now moves aside on the Settings tab, as it already did on Tasks.

  • Apps connected to Gezel, such as DocBlocks, now list the recommended image models first, instead of whichever model's id sorts first. After a chat model downloads, they keep showing its name, such as "Qwen 3.5 (0.8B, Q4)", instead of its id.

  • The llama.cpp engine that runs GGUF models on your computer is updated to version 1.0.0. In our checks, replies and speed were unchanged, and three more model designs can now be added from Hugging Face: Kimi K3.1, Ripple and Xing 4.0.

  • A web page open in your browser can no longer send requests to the llama.cpp engine, even a page served from your own computer. The engine already required a key that Gezel sets when it starts it. Nothing changes in how you use Gezel.

  • Catalog content updated (gilde 0.3.10).

In documents

  • Exporting a document to PowerPoint puts its footnotes in the speaker notes of the slides that cite them, instead of leaving them out or printing them on the slide.

  • In a slideshow, a slide with a video stays up until the video has played, trims and cuts included, instead of moving on partway through.

  • A table, diagram or scene at the end of a document now fills the space left in the window instead of stopping at a fixed height. Short tables and diagrams you sized yourself keep their size.

  • A document that ends in a diagram, table or code block always has a place after it to keep typing, without an extra blank line in the file.

  • Checklists on slides show check boxes instead of [ ] and [x], and an empty checklist item is no longer dropped when the document is saved.

  • A data file shown in a document shows its name with spaces, such as sales 2.csv, instead of sales%202.csv.

  • A slide deck a crew member makes now draws a short list as a picture of how its items relate: a numbered list as steps, Label: detail items as cards, dated items as a timeline. Only a list that is a slide's whole body is drawn, and the shapes stay editable in PowerPoint.

  • Document editor updated (Squisq 2.13.16), and the document maker crew members use (DocBlocks 2.7.16).

In the terminal

  • gezel task wait <ref> and gezel do … --wait now print where a finished task's result is: the main file, what else it made, and its folder. gezel task wait <ref> --out <dir> copies it into that folder.

  • gezel run sends text you pipe or redirect to it after the prompt: gezel run "Summarize these notes for the board" < notes.txt, or git diff | gezel run "Write a commit message for this change".

  • gezel run no longer prints the service's warnings above the reply. They go to the service log, and only errors reach the terminal.

  • gezel model list marks the model chats use and the models another Gezel on this computer shares. gezel model pull downloads nothing when a shared copy is already here.

  • gezel agent show and gezel agent output-limit take a crew member's name as well as an id.

  • When you let gezel do change a folder's files, with --allow-edits or by answering yes, and the task then fails to start, the folder goes back to how it was.

  • gezel knowledge export-parquet downloads DuckDB when it is missing and you add --download. Without it, the message names the command that gets DuckDB, instead of pointing at the desktop app.

  • Download progress written to a file or a log shows the size once and then 25, 50 and 75%, instead of hundreds of updates on one line.

New in Help

For developers

  • Each downloadable entry in /v1/models carries its docbench score and a docbench_recommendation that ranks models for document work on this machine. The phone's catalog carries the same.

  • Installed catalog models in /v1/models carry the catalog's display name, and /v1/images/models lists the curated image models first, best first.

  • A @bendyline/gezel-capacitor release no longer waits for a native build when native/runtime has moved past the pinned release. It ships the pinned release as built, and still refuses a plugin that calls runtime API that release lacks.

Things to see and do after update

  • If you run Qwen 3.8 27B, start a plan that writes a report and watch Activity show it being checked. To skip that step, use Settings → Artificial intelligence → Check work before handing it in.

  • Open a project's Settings → Preferences → Icon and give it a mark of its own.

Watch this article as a slideshow · Read it as markdown, for AI agents