1.26279 — 6 October 2026

Four days after 1.26275. This release is about seeing and trusting what your gezels do. Gezels are the named AI craftspeople on your crew. There is now one place that shows everything running, waiting, or waiting on you. Gezels that write for you cite where each fact came from, and refuse to put an unsourced fact into your document. On a Mac, your chat no longer waits behind background work.

If you read only one thing: the Activity key at the top of the window replaces Updates. Open it to see what every gezel is doing, and answer what needs you without leaving the screen you are on.

Activity: everything going on, in one place

Before, what your crew was doing lived in three places. An Updates button appeared only when a gezel had a question. A queue meter appeared only while a model was busy. And a task waiting on you showed up nowhere until you opened it. Now one Activity key always sits at the top of the window. It summarises the moment, for example "2 need you · 1 working" or "All quiet", and shows an amber dot when something needs you.

Opening it shows What's going on, across all your projects, in four sections:

Needs you. Questions from gezels, which you answer in place: pick a choice or type a reply, then Submit. Text you have started is kept while you look elsewhere or switch projects, until you close gezel. A task step assigned to you also appears here. Type what should happen next and choose Continue task, or pause or cancel the task. You no longer have to open the task to move it on.
Working. Each gezel using a model right now, with Stop.
Next. Work waiting for a turn or scheduled for the Night Shift, the work gezels do overnight. You can move it up, move it down, or cancel it. When automatic work is paused by your AI engagement setting, the row says so and offers Activity settings.
Ready. Finished work and Night Shift results to review.

Each row says in plain words what is happening, such as "Waiting for a free model slot" or "Still active. No new progress reported in the last two minutes." A task now shows as working only when something is actually running, not because it was once started. Paused tasks stay out of the list. If gezel cannot refresh the panel, it says so and keeps the last update. It never shows "All quiet" when it does not know.

The ? beside a project in the side bar opens Activity for just that project, with Show all projects to widen it. On Home, the "waiting on you" and "ready for you" chips now say Open Activity.

Gezels that cite their sources

Local models are fluent with names, dates and family relationships, and often wrong about them. Asked for George Washington's family tree, a writer gezel produced several confident mistakes. Now a gezel that writes facts for people must cite them.

It looks up what it does not have. Each source it reads gets a number, and its reply cites them, like this: [1].
Unsourced facts are refused before they reach your document. A gezel writing into an open Word, PowerPoint or LibreOffice document is turned back when a sentence states a name, date, number or quote that none of its sources shows. It is told which lookup to make next. The same check covers prose files it saves, such as .md and .txt. A saved file that cites sources gets a Sources list at the end. Code and data files are never checked. After two refusals the gezel may write anyway, and its reply warns you how many statements could not be matched to a source.
In chat, nothing is blocked; it is shown. Each [1] is a link to its source. Under the reply, a collapsed row such as "3 sources · 1 statement not found in any source" lists the sources and the sentences that did not match. The row is a quiet note, not a red error.

This applies by role. Writers, researchers, reviewers, editors, journalists, historians and similar gezels work this way; novelists and poets do not. Any gezel writing into an open document works this way too. To change it, open the gezel's page and use Sources: to choose By role, Always cite or Off. In Word and LibreOffice, a pane with no gezel chosen yet now starts with one that cites its sources.

Measured on Gemma 4 31B, with 20 questions about the Washington family each way: with sources on, 3 of 147 statements were wrong, against 7 of 137 with them off, and 2 replies of 20 contained an error, against 6. Every reply looked something up. The cost was about 15 seconds more per answer at the median, and a gezel leaves out what it cannot find. That is one model, one topic and one run, so treat it as a direction, not a rate.

The check is literal, so it has limits. It catches invented facts, but not a wrong fact copied faithfully from a wrong source, or two true names joined by a false relationship. With GitHub Copilot and the Claude and Codex command-line tools, gezel asks the model to name its sources but cannot check them.

Separately, a link a gezel writes to a reference catalog passage (it begins knowledge://) is now clickable in any reply, and opens that passage in the Knowledge area.

Your chat goes first on a Mac, and answers come back whole

On a Mac, a message you type goes ahead of background work. When gezels were working through a task, your question could sit behind them, labelled "Processing prompt", until it timed out without starting. In one case a 300-token question waited behind a 55,000-token task step for nearly seven minutes, then gave up. Now the model on your Mac (run through MLX, Apple's engine for its own chips) serves what you type, your answers to questions, and a gezel consulting another one for you ahead of task steps and other background work. Background work pauses between steps and resumes exactly where it left off. While you wait, the status reads "Waiting for another chat to finish" or, on the background reply, "Paused while a chat goes first". The one exception is a Mac already short of memory, where the background work keeps its place.

Long conversations survive quitting, restarting and switching models. The model keeps a copy of what it has already read, so the next message only reads what is new. That copy was often lost:

When gezel stopped the model. The model had three seconds to save, and a 98,000-token conversation was lost that way. Gezel now asks it to save first, and waits up to 30 seconds.
After a restart, once a gezel had used a tool. Last release fixed this only up to the first time a gezel used a tool. On a Mac, gezel now saves the exact conversation the model saw, and replays it after a restart. A resumed 69,000-token task step used to read 55,000 tokens again, about four minutes on a 27B model.
When you changed a setting. Changing a setting no longer throws the copy away.
When you stopped a reply. A stopped or cancelled reply keeps what was read, so Retry starts from there.

Answers after a lookup are whole again. Since 1.26274, some local models, Gemma among them, were cut short after using a tool. A reply longer than about 700 characters was reduced to its last paragraph, and the rest folded away as "thinking". This affected Macs and other computers alike. Replies after a search now keep their full answer.

Search loops stop sooner. A model that kept rewording a search could go round indefinitely, because each wording looked new. Gezel now compares the sources a search returns, not its words. On the third search that finds nothing new, the model is told to stop rephrasing; on the fourth, the turn ends.

When a reply on a Mac does run out of time, the message now says why in plain words. For example: "This message was waiting for the AI on your Mac to finish another chat, and it never got its turn. Nothing is wrong with the AI — send the message again."

Reference catalogs find the right page

Knowledge catalogs are offline reference libraries, such as shelves of Wikipedia, that gezels search alongside your own files (How knowledge works in Gezel). Searching them got noticeably better.

The first search after starting gezel no longer comes back empty. That search used to load part of the catalog itself, run out of time, and return nothing at all. The catalog now loads in the background as soon as gezel starts.
Questions phrased the way people talk find the right pages. Words like "can you tell me how to" used to steer the search: on an Azure documentation catalog, "Can you tell me how to integrate Azure search with blob storage?" put a page about "eID-Me" first. Those words are now ignored. A word found in most of a catalog, like "azure" in that one, no longer slows the search: that search took 550 ms and now takes 40 ms.
A question that names an article finds that article. "Who were the members of ABBA?" did not show the ABBA article in its first six results. Now it is first. Looking up "What is …?" found the right article 92% of the time in the music catalog, up from 86%, and 91% in film and television, up from 84%.
Searching one catalog in the Knowledge area shows up to 20 results. It stopped at six, even when hundreds of titles matched.
Three new Wikipedia shelves: Film & Television (272,280 articles), Music (317,534) and Sports & Games (371,095), from the April 2026 English Wikipedia. There are now 20.
Installing is steadier. An install no longer hangs when the search model's download stalls; after five minutes it finishes, and keyword search works until the model arrives. Progress now reads "Verifying the catalog…" and "Checking the file…". A + beside the catalog picker, Add knowledge catalogs, opens Settings → Knowledge.

A gezel can ask before editing a read-only project

Last release made Allow built-in tools and background work to modify the workspace, in Project → Settings, hold. A gezel that needed to change files in such a project could only tell you it couldn't. Now it can ask.

You see a Permission request card in the chat and in Activity. It says exactly what would change: create, edit, rename and delete files in that project's folder, for all gezels, until you turn it off again, and where to turn it off. Choose Allow project file edits and continue and the gezel carries on, or Keep current permissions and it is told not to ask again.

Only you can grant it, and only from the desktop app. A gezel cannot answer its own card, and a typed reply never grants anything. A meester or project lead can no longer switch the setting on by itself either. If the project's folder moved after the gezel asked, the request is void and must be made again.

Security

On Windows, the background service that loads models can no longer change gezel's program files. That service runs with few rights because it opens model files from the internet. Its folder sat inside one it could change, so it could have rewritten code that every account on the computer runs with that person's own rights. It can now only read that code, and every install, including a repair, unpacks a fresh copy instead of trusting the one it finds. Nothing you do changes.
Private links stay out of diagnostics. Preview links and connected-app tokens are removed from the service log and from the freeze reports added last release.
A shortcut that leads outside a folder is refused with a clear message. Changing a file through a shortcut that points outside your library or a project's artifacts used to show "Something went wrong inside Gezel". It now says the path goes through a shortcut to somewhere gezel won't change files.

Smaller things

Web previews run TypeScript without a build. A page in the workspace that loads .ts, .tsx or .jsx files used to show "Preview cannot run unbuilt TypeScript/JSX". The preview now converts them as they load and runs the page. A missing file, an npm package or a syntax error shows on the page with its file and line, instead of a blank frame.
Apple Intelligence on a Mac. On a Mac with Apple Intelligence turned on, Apple Intelligence is a choice under Settings → Artificial Intelligence, with a readiness line such as "Ready · 8,192 tokens of context" and Check again. Apple manages the model, so there is nothing to download, and it works offline. It is small with a short memory: good for chat, notes and short tasks.
A new gezel role, the Conversationalist, talks things through with you, looks things up, and brings in another gezel or a craftbook when that helps. A craftbook is one of gezel's saved recipes for a piece of work. A new Chat project now comes with one. See The Conversationalist.
The task page. Steps are now stops along a route, with the current one centred and marked with its gezel. Scroll or drag to move along it. In a task's Chat tab the message box stays at the bottom of the window. The latest file sits above the tabs, with Show in folder and Save a copy… as icons.
Documents open as documents everywhere. Opening a Word, PowerPoint or PDF file from a task's Open or from search now gives the same preview as the file list, instead of garbled text. A PowerPoint opens as a slideshow. Other binary files say there is no text preview.
Craftbook steps that update their own task can finish again. Since 1.26275, a step whose own check edited its task conflicted with itself and could never advance.
Try again is not undone by an old check. A check still running when you chose Try again could pause the task again moments later. Its result is now discarded.
A stray numbered folder no longer blocks new tasks. A folder a gezel created with a huge number in the task area made every new task in that project fail. Such folders are now ignored.
Restoring a backup that adds gezels or projects is quick again. Last release's careful copy took minutes for a large workspace, and a copy that died midway blocked every retry. New items are now moved into place in one step, still without overwriting anything newer.
In Word, a diagram that cannot be placed leaves the document as it was. Before, the text went in without the diagram, and a placeholder could be left behind.
Double-clicking an answer sends it once.
No more "Loading…" flashes. Across about 25 screens, gezel now keeps the frame still while it loads instead of flashing the word.
Gezel rests more when idle. The freeze detector added last release made an idle gezel wake about 59 times a second, against 13 before. It now checks four times a second.
Codex CLI sessions take less disk. Each one downloaded about 57 MB of plugins and catalogs it never used. Your own Codex setup is untouched.
Catalog content (the gilde) moved from 0.1.80 to 0.1.83. Qwen3.8 27B uses a slightly smaller file on llama.cpp (16.5 GB, was 17.1 GB) and, on a Mac, gains a 239 MB draft model for speculative decoding. The PowerPoint deck craftbook now ends when the file is saved, in five steps instead of seven.
llama.cpp moves to v0.6.0. Its settings page (This PC (llama), or This Mac (llama), under Artificial Intelligence) gains Draft sampling, an experimental choice for how a draft model's guesses are picked; the default is unchanged.

In the terminal and editor

If you use the gezel command (The gezel command line) or the VS Code extension:

Running gezel from your home folder works. Gezel never makes a project of your home folder, the temp folder or the top of a drive. The command used to stop with an error there. It now uses the Default project, says so, and tells you how to pick another. VS Code does the same when the open folder is your home folder.
When gezel is not running, the message says what to do: "Gezel is not running for this user. Start it from the Gezel app, or run gezel start."
gezel run no longer prints a memory warning on every run. The memory step it interrupted now finishes on a later turn.
Catalogs can find articles near a place. gezel knowledge nearby <catalog> --latitude … --longitude … --radius-meters … lists articles near a point, and gezel knowledge search takes the same three options. This needs a catalog that records places; none of the Wikipedia shelves do yet. See Working with knowledge catalogs.
gezel knowledge build is more forgiving. --skip-images builds without images and keeps their descriptions, and an unusable image is skipped with a warning instead of stopping the build. A page listed in several sections of a table of contents now appears in each of them.
Catalogs built with this version use format 0.7, which older gezel releases cannot open. Gezel still reads catalogs in 0.5 and 0.6.

For developers

Activity. GET /api/activity returns each item's section, title, detail and links, plus questions and queues; client.getActivityStatus(). The same resolver is exported from @bendyline/gezel as resolveActivity().
Permission requests. ask_user_question takes permissionRequest: "workspace-write". The service builds the card; only the first-party app can answer it. A session token that tries to change managedWorkspaceWritePolicy or allowGezelWrites through PUT /api/projects/:id gets 403.
Factual writing. factualWriting: true | false in gezel.md (also POST /api/gezels/:id/settings), GezelSummary.writesFactually, and ChatMessage.grounding (evidence, counts, and up to 20 unmatched statements). Craftbook citationsResolve gates now check knowledge:// citations against installed catalogs.
Human steps with instructions. A task assignee can be { kind: "user", instructions: "…" }; the instructions lead the Activity row and the review card.
Knowledge for connected apps. /v1/knowledge/state, /update and /retrieve, behind a new grantable knowledge scope and the Connected Apps switch, return reranked, cited passages and never touch projects or files. Retrieval requires the relevance model (409 reranker_required otherwise). app.knowledge.* in @bendyline/gezel-app-sdk.
Embedding. createDesktopEmbedding (@bendyline/gezel-app-sdk/host) and createMobileEmbedding (@bendyline/gezel-capacitor, still a private preview) own the AI lifecycle for an app that embeds gezel. A ModelManager lists, inspects, prepares and watches models, and never downloads without allowDownload: true. Packaging checks ship as @bendyline/gezel-service/packaging and @bendyline/gezel-capacitor/packaging.
Client. connectToLocalGezel() reads the running service's port, token and certificate and pins the certificate; client.streamSessionEvents(). The README quick start, which failed on the self-signed certificate, is rewritten and now runs in the package check. gezel-mcp started from an outside MCP client's config finds the running service by itself.
Search tools. search takes catalogs (knowledge only) and spatial; a new knowledge_nearby tool and POST /api/knowledge/nearby. read_document on a passage citation opens that passage.
gezk 0.7. Typed document locations, radius queries and shared table-of-contents placements, with JSON Schemas under packages/gezk/schemas/0.7/.
Diagnostics. GEZEL_MEMORY_DIAGNOSTICS=1 logs memory use once a minute. A write through a shortcut that escapes its root is 403 with code symlink-escape.
Dependencies: @bendyline/gilde 0.1.83; native engines 0.1.48, with llama.cpp v0.6.0.

Things to see and do after update

Open Activity at the top of the window and answer whatever is in Needs you. If a gezel writes for you, open its page and check Sources: is set the way you want. If you use reference catalogs, the new Film & Television, Music and Sports & Games shelves are in Settings → Knowledge.