What's new — crenya

1.4.5-r45 7 September 2026 Current

Composer

Multitask is no longer marked deprecated. The badge was inherited, not decided: the mode works, and it is billed. We measured it on the same task as Agent, and it made about five times as many requests and cost about three and a half times as many credits, because it splits the job across subagents that each bill their own turn. That is now stated on the mode itself instead of hidden behind a label that said the wrong thing.
Every mode in the picker now carries a one-line description of what it does and what it will not do. Plan cannot edit; Ask changes nothing; Code Review reads and reports; Multitask costs several times a normal turn.
The session header shows the credits a session has spent. The number is the one the server bills, sent back with every reply, not an estimate made in the editor, so it matches the balance on your account page.

Fixed

When the assistant is refused an action, the message now says who refused it. In Plan mode an edit used to fail with a generic permission error; it now says that Plan does not allow edits and that switching the composer to Agent will. A rule in your own permission settings, a guardrail, and a headless run each name themselves too.

Settings

Background subagents are off by default and now opt-in, with a cap on how many may run at once (three by default, adjustable from one to eight). Before this, nothing limited them. When the cap is reached the assistant is told so, and waits or finishes the work itself, rather than starting one more.

Under the hood

A measurement tool now reports how many turns each request takes per mode, from the session store, so the next round of changes to the modes can be judged by a number rather than a feeling. The baseline it measured: half of all requests finish in four turns or fewer, one in ten needs twenty-eight or more.

1.4.5-r44 4 September 2026

Composer

The composer now has a mode picker with six modes, next to the depth control: Agent for ordinary work, Plan to think a change through without touching files, Debug to chase a fault, Multitask to break a job into pieces, Ask for questions about the code, and Code Review to read a change and report on it.
Five of those six were already built — prompts, permissions and all — and simply had no way to reach them. Plan is the clearest example: it is genuinely unable to modify your code, and until now nobody could select it.
Code Review is the one new mode. The review prompt has driven the /review command for months, but a command reviews once and hands the session back to an agent that edits. As a mode it stays a reviewer: it reads and reports, and does not rewrite what it is judging. Applying its findings is a deliberate switch back to Agent.
A mode says what the assistant may do; the depth control says how much depth to buy. They are separate on purpose, and each is remembered per session.
New sessions start in Agent. They used to start in whichever mode happened to be defined first internally, so adding a mode could have silently changed where everyone began.

Fixed

The depth picker advertised prices that were not real. It read Auto 1x / Pro 1.3x / Max 1.6x / Lite 0.3x. Measured against the running backend, Pro and Max bill identically, and the gap between Max and Lite is far wider than the labels claimed.
Those figures could not be kept honest where they were: pricing changes on the server at any time, while that text ships inside an editor you update every few weeks, so any number written there is wrong the moment the ladder moves and nothing says so. The labels now describe what each setting actually does — Auto routes per task, Pro always uses the frontier tier, Max adds a second pass that re-reads the code it just wrote, Lite stays on the cheapest tier — which stays true whatever the prices are.

1.4.5-r43 28 August 2026

Security

Commands the assistant runs no longer receive your crenya sign-in token. Until now every command started from the chat — and every language server, formatter and background task — was handed the assistant's whole environment, which includes the token your editor uses to sign in to crenya. A command does not need it, but anything that ran could read it: a build script, a package's install hook, or a command a model was steered into suggesting by a file in an untrusted repository.
The fix is at the two places every command is started from, not at each caller, because a caller that scrubs can have its work silently undone one layer down — which is what we found had been happening. A check now runs on every build and fails it if a new way of starting a command skips the scrub.

Under the hood

When a request fails, the reason recorded now says which of five different failures it was. They had all been counted as one, so a single real problem was indistinguishable from four unrelated ones.
A model that cannot be found is now attributed to the side that caused it — ours or the provider's — instead of being logged without the one flag that decides which.
The prompt that nudges a stalled task to finish now records whether it actually achieved anything, not merely that it was sent.

1.4.5-r38 20 August 2026

Security

The library that draws diagrams in chat has been updated. The old version could be sent a diagram that made it loop forever, alter shared settings, or apply styling outside its own box — and diagrams come from the model's reply, so the input is not always yours. Ten published advisories are closed by this update.

Fixed

On macOS the notarization ticket is now attached to the app before the disk image is built, so both carry it. Without it a first launch had to reach Apple over the network to be approved, which failed for anyone installing offline.

Under the hood

Every build now records the exact source revision it was made from, and each published download is checked on a machine of its own architecture before release. Nothing changes for you — it means we can prove which code is in the copy you are running instead of inferring it.

1.4.5-r37 16 August 2026

Security

The bundled agent shipped a compression library carrying a critical published advisory. It has been updated.
The session database — which holds your sign-in credentials — was written so that any other account on the same machine could read it. It is now readable only by you.

Fixed

A single common word in your message could load an entire skill's instructions into the conversation, crowding out your own context. Matching is much stricter now.
crenya worked out your project's stack and then ignored its own answer when choosing which skills to apply. It uses it now.
"Next.js" matched nothing, because the punctuation split the word before it was compared. Short two-word requests were being skipped entirely as well.
A safety rule meant to stop curl … | bash was also blocking ordinary work — piping a downloaded response into python3 -c or -m json.tool to read it. Of the three commands it had ever blocked, two were harmless. It now checks whether the interpreter would actually execute what it receives.
When a tool call was missing an argument, the assistant was told its search text "did not match uniquely" and to retry with a shorter one — advice about text it had never sent. It now reports what actually went wrong.

Windows

Windows builds are now checked, before release, that every bundled native component was compiled for the machine it will actually run on. A build that fails is not published — previously such a build would start normally and then have no terminal and no file watching, with nothing on screen to explain why.

Windows x64 only. macOS and Linux stay on 1.4.5-r36 and are not offered this update. Windows on ARM is not published yet — a corrected build is being verified on ARM hardware, and until it passes we offer no ARM build rather than one whose terminal does not work.

1.4.5-r36 11 August 2026

Updates

Every release was being held back from every user for five days. The update client waits a set number of hours before installing a new build; the value meant to shorten that had been written in a place nothing reads. A fix now reaches you in hours instead of the better part of a week.
On Windows, updating a per-user install fetched the system-wide installer, which then asked for an administrator. Each install type now gets its own.

Fixed

The Windows installer showed a black square where the crenya logo should be, on every page. The image was correct; it was stored in a format the installer cannot draw.
"Release Notes", "Documentation" and "Request Feature" in the About dialog opened Microsoft's VS Code pages. They open ours now — this page, in fact.

macOS (Apple Silicon & Intel), Windows and Linux — x64 and arm64. macOS: drag crenya into your Applications folder rather than running it from the disk image — macOS runs apps launched from a disk image in a read-only location, and updates cannot install there.

1.4.5-r35 11 August 2026

Security

The agent could open .env files without asking. It now asks first — a blanket "don't be naggy" read permission had been overriding the rule that protects secrets.
A project you open but don't trust could pre-approve shell commands through its own agent and command files. Those are now stripped, and crenya tells you when it has done it instead of failing silently.

Fixed

Your editor theme was dropped instead of carried across when a project's config was upgraded.
The setup tip pointed at a themes folder crenya no longer installs to.
A codebase search that failed reported "empty codebase", which sent you looking for an indexing problem that wasn't there. A search failure now says it failed.
A one-off failure to load the model list was remembered as "no models available" for the rest of the session.
The feedback prompt could never appear once a conversation got long.
The settings screen and the agent disagreed about which config folder was in use, so an edit in one was invisible to the other.

Faster

A panel that only displayed information was costing an extra model call on every single turn. Removed — same output, one less round trip.

Look

The app icon's backdrop is now white, on every platform and in the installers.

macOS (Apple Silicon & Intel), Windows and Linux — x64 and arm64.

Your codebase, understood — and reviewed.

Ask in plain English

Knows your stack

One-command code review

Connected in seconds

The whole toolkit

Release notes