Выпуски
RSSv0.62.2
- **`chimera mcp desktop`: Claude can operate the running desktop app.** It is a stdio MCP server that bridges to the app's own local API, and every call runs in the original handler. It is set up in two steps: 1. Turn on **Settings → Claude → "Allow Claude to operate this app"**
v0.62.1
- **The owner's block is read last again.** When `todo_write` was registered, which is every terminal and desktop session, the task-list sentence was appended after the owner's instructions. That broke the precedence study 25 set for every situation: owner > situation > project >
v0.62.0
- **The default agent model is now `openrouter/openai/gpt-6-luna`.** A paired bake-off ran it against `deepseek-v4-flash-0731`, `deepseek-v4.1-flash` and `qwen3.7-flash` on SWE-bench django. The run used the Chimera harness, a network wall and a disk sandbox. - luna resolved **78
v0.61.2
- **`chat`, `assist` and the TUI draw the task list the agent keeps.** `todo_write` is registered on the terminal by default, and the tool itself says it exists so the person watching can see where the agent is. The desktop drew that list; the terminal did not, so the model kept
v0.61.1
- **A reopened turn shows what it was actually charged, beside the estimate.** The receipt's `usd` is priced from the catalogue row of the model id, but one id is served by several routes at different prices. The router's generation record carries `total_cost`, the real charge. T
v0.61.0
- **A typed decision is declared once, and it may only add scrutiny.** Contract v2 (`chimera/decisions/spec.py`): every decision point is a `DecisionSpec` that names its questions, its threshold, its mode (`shadow` by default), what no signal becomes, and the bench that measured
0.60.0 — it continues by itself at the step ceiling, the approval card shows the number that asked, and the sidecar can fetch its browser
- **The Code screen continues a turn that stopped at the step ceiling, by itself, up to three times.** "Continuar sozinho." `max_steps` is the one stop that means *the task was going fine and ran out of room* — the model was working, the ceiling cut it, and the work is incomplete
0.59.0 — hands-free voice that talks while it works, a conversation you can share, and the agent's browser on screen
The nine-item list audited on 2026-09-16 — what the marketing said against what the product did — shipped one pull request per item over two days, each measured before it was built and tested against the running app afterwards. This release is that list, plus what its live tests
0.58.0 — the turn receipt names the route, and a policy review on the Code screen asks you with a card
Two things the live test of 0.57.0 found by using the installed app, fixed the same day.
0.57.0 — what you said at the start survives compaction, and a patch that adds a dangerous call waits for you
When a Code conversation grows past its budget, the app compacts the older part to free room. Until now the dropped span was replaced by a count — *"21 earlier messages were removed"* — which is honest about how much went and says nothing about what still binds. A convention you
0.56.0 — your projects belong to the app, not to one browser window
The Code sidebar groups your conversations by project, and you can add a project before you have ever worked in it. But that list — and the name you give each one — lived in the desktop's own web storage. So it was never really a list of *your* projects: it was a list belonging t
0.55.0 — the approval card asks the safe way, and a turn can stop on its plan
In the desktop app, **"Allow this once" was drawn first and styled as the primary button** — and because the card mounts inside a dialog whose focus trap takes the first focusable element, the caret landed on it. Answering with Enter, without reading, approved. The terminal UI an
0.54.0 — the auditor reads what a slice can do, a runaway is bounded, and the approval question names the command
Two days of measured changes (#425–#447), every one pre-registered, sabotage-verified and published with the predictions that failed.
0.53.0 — the terminal is governed too, and a question finds you
A release about the surfaces a person actually sits at — the terminal, the full-screen app, the desktop chat — which were the ones nothing governed, and about the question that now finds you wherever you are.
0.52.0 — the verify command was the one shell nothing watched
A release about things this project was saying that were not true, and the guards that will not let them come back.
0.51.0 — a tool call now waits for you
The agent used to lose a tool in the middle of a run and not tell you. This release turns that into a question.
0.50.0 — the agent tells you what it is doing, and can reach you
Five things in this release were happening without you being able to see them, reach them, or turn them on.
0.49.3 — six surfaces that said more than they knew
Six defects found by *using* the app rather than reading it — a day of building three real projects with it — plus the model defaults, which had drifted a generation behind.
v0.49.2
The release that makes the updater actually work. Both fixes needed a real machine to find, and neither could have been caught by a test.
v0.49.1
Four fixes, all found by watching 0.49.0 publish and then using it.
v0.49.0
The app speaks your language everywhere now — including the one screen that asks permission to change your machine.
v0.48.1
An audit of the shipped 0.48.0, run against the running app rather than the source. Almost all of it is one defect wearing different clothes: **something existed, had tests, and had never run.**
v0.48.0
Fifty release candidates, and the two things to read before upgrading are both defaults that changed under you.
v0.47.0 — the safety layer said it was protecting you
Not a feature release. Seven fixes came out of one thread: a refused tool call reported itself as a success, and pulling on that reached the kernel, the rules, the audit, and the mode that was supposed to only measure.
v0.46.0 — a skill library you can read, and sixteen claims corrected
A release about **rules written in one place and not applied to the thing next to them**. Nine of those, found by going after the *shape* of the bug rather than the instance — plus the skill card library that grew out of writing the incidents down.
v0.45.1 — the gateway boots again
**If you run Chimera headless in Docker, v0.45.0 does not start. Upgrade to this.**
v0.45.0 — the agent nobody is watching
v0.44.0 made the desktop app an editor. This one is about the other half: the agent that runs on a schedule, with money, while you sleep.
v0.44.0 — the app is an editor now
The desktop app stopped being a console for agents and became a place to write code. Seven phases of work land together: an editor with tabs, diagnostics that agree with CI, inline completion from a local model, search across the workspace, retrieval over the repository, a comman
v0.43.0 — uma chave para qualquer provedor, e as telas que faltavam
The release where a key for any provider LiteLLM reaches starts working, three things that existed only in `.env` got a control, and two features that were built but unreachable finally have a surface: attaching a file, and the registry the board dispatches to. ### Added - **Any
v0.42.0 — as configurações passam a valer, e o agente passa a ser seu
Você podia trocar o modelo padrão nas Configurações, ver a tela confirmar, reler e ver o novo — e a resposta seguinte vinha do antigo. Não era um bug de uma configuração: eram sete. Esta release é sobre a tela parar de mentir, sobre duas cercas de segurança que só existiam no `.e
v0.41.0 — anexar um PDF e ditar funcionam de fábrica
A v0.40.0 trouxe anexar e ditar. Nenhum dos dois funcionava quando você instalava o app — o leitor de documentos e o modelo de fala eram extras opcionais, e num aplicativo de desktop "opcional" quer dizer "ausente". Agora vêm no pacote.
v0.40.1 — anexar um arquivo de texto não precisa de conversor
Correção de uma coisa só.
v0.40.0 — uma conversa só, e o app diz se o modelo realmente enxerga
A v0.39.0 tirou coisas da tela. Esta devolve três — anexar, ditar, aprender — e fecha uma porta que estava aberta desde sempre.
v0.39.0 — a tela para de perguntar, e passa a dizer o que fez
Esta release é, na maior parte, uma subtração. A tela Código abria com **onze decisões** antes de você poder escrever uma palavra: pasta, árvore de arquivos, executor de comandos, dois seletores de segurança de três valores cada, tabela de perfis, comando de verificação, número d
v0.38.0 — o loop pode ser medido, e avisa quando parou de chegar a algum lugar
Esta release quase não tem funcionalidade nova. Ela mexe no *harness* — tudo que existe em volta do modelo — depois de um estudo que cruzou quatro fontes independentes sobre engenharia de harness com uma auditoria do próprio Chimera.
v0.37.0 — o app de desktop vira um espaço de trabalho do agente
O app de desktop deixou de ser um menu de funcionalidades e virou um espaço de trabalho.
v0.36.2 — a replicação segurou, e a decomposição contradisse nossa própria previsão
O resultado do SWE-bench **replicou fora da amostra** — e uma quarta rodada decompôs o ganho por componente, contradizendo a previsão que nós mesmos tínhamos registrado.
v0.36.1 — a citação apontava para o arquivo errado
Um patch de honestidade. A v0.36.0 exibia os números certos com a **procedência errada** — e num projeto cuja tese é medição rastreável, isso é a falha que menos podemos deixar passar.
v0.36.0 — o furo de graduação, o número que subiu ao ser corrigido, e o primeiro placar externo
Esta versão é sobre uma coisa: **descobrimos que nosso próprio benchmark tinha um furo, consertamos, e o número subiu.** Depois fomos buscar um número que não é nosso.
v0.35.0 — the desktop app wakes up, remembers, and can reach you
The desktop app stops being something you have to open and ask. It now fires schedules on its own, can keep what you tell it to remember, and can reach you on Discord — all configured from the UI, no terminal.
v0.34.1 — installable again on Windows/macOS and from a clone
Patch release. Since 2026-07-12, `pip install chimera-agent` failed on **Windows and macOS**, and building from a clone failed **everywhere**. Both are fixed here. No feature changes — everything else on `main` waits for the next scheduled minor.
v0.34.0 — stop a batch, and tests that bite
**Cancellation is cooperative, not an instant kill:** a task finishes its current attempt and then halts, because an in-flight model call can't be interrupted. The UI says exactly that.
v0.33.0 — Intel Macs + a real frontend test suite
`macos-latest` is arm64-only, so until now an **Intel Mac literally couldn't install** the native app — the aarch64 `.dmg` won't run — and got no updater entry either. This release builds on `macos-15-intel` too, publishing an **x86_64 `.dmg`** plus a signed `darwin-x86_64` updat
v0.32.2 — updater manifest Linux fix
A patch fixing the updater manifest: the Linux entry in `latest.json` now correctly points at the `.AppImage` (v0.32.1 pointed at the `.deb`, which the Linux auto-updater can't apply). Windows and macOS were already correct. This republishes a correct `latest.json` for all three
v0.32.1 — updater manifest published
A patch that completes the in-place auto-updater from v0.32.0: the macOS build now emits the signed updater artifact (built with `app,dmg`), so this release publishes the `latest.json` manifest across **all three** platforms. The updater is now live end-to-end.
v0.32.0 — native in-place auto-update
The native desktop app now **updates itself** — signed and consent-prompted.
v0.31.0 — Stop a run · browser-verify · update signal
Three honest-by-construction additions to the desktop app.
v0.30.0 — Agent Manager: parallel isolated runs
The **Agent Manager** — run several coding tasks at once, each isolated, with a live board and an honest conflict report. This is the parallel, agent-first surface, built on Chimera's existing git-worktree isolation.