Changelog

• At Manual and Assisted autonomy, a run that Jev judges complete now ends with "Done, awaiting your review". In 0.2.0 a fixed threshold left some finished runs marked "Not done", 3 of the 12 recorded sessions where Jev had said done.
• When the LLM answers a retry without changing anything, Noul uses Jev's last verdict and stops asking Jev again about the same state.
• Jev now reads the LLM's final answer in the done check, so a task that asks Noul to run, read or report something is judged on the answer even when no files changed.
• About, Finder's Get Info and noul --version show Noul's own version, 0.2.1, with the VS Code version next to it.

Download Noul again to get 0.2.0. It does not update itself yet.

• Every reply ends with one row that shows what Jev decided and what the LLM wrote, with a call count for each. Expand it for the calls, wait time and tokens of each job.
• When an Edit or Agent request is missing a detail, Noul asks up to three short questions before it starts. Jev preselects the answer it judges likely. noul.chat.clarify turns this off.
• Approvals have an Edit button, so you can change a command or a tool's arguments before it runs. Jev checks the edited version again. Reject with Reason tells the LLM why, so it can try something else.
• The done check sees new compile and lint errors in the files the agent edited and asks the LLM to fix them. Errors that were there before the task are left out.
• A .noulignore file keeps files away from the agent, Tab, Cmd+K and the index, and Noul can also read .cursorignore. Rules in .noul/permissions.json deny, ask about or allow actions. Allow rules apply only after you trust them.
• /review asks the LLM for findings on your uncommitted changes, or on a branch with /review main. Jev decides which findings are real and how severe they are. In Agent or Edit mode you pick the ones Noul fixes.
• Larger Agent and Edit requests start with a short plan, shown as a to-do list above the chat input. Jev ticks each to-do off from the diff and the tests. /plan lets you edit or reject a plan before anything changes, and /implement runs it.
• The agent can read public web pages with web_fetch. Jev screens every page before the LLM reads it, and a page you did not name asks for approval first.
• Answers appear in the chat as the LLM writes them. Text written before you press Stop stays in the reply.

Jev call and token counts are higher than in earlier builds because every Jev call and the prompt cache's tokens are now counted.

Noul opens with its own setup on first launch.

• The mark draws itself in three strokes, then the wordmark fades in. Click or press any key to skip.
• Pick Light, Dark or Other. Other opens the full theme picker.
• Connect Jev and a model. Paste a TypeSafe key for Jev and an Anthropic or DeepSeek key for code. Each field links to where you get the key. Skip it and Noul asks the first time you open the chat.

Keys are stored in your system keychain and can be changed later from the Command Palette.

Noul for macOS on Apple silicon ships as a disk image. Open it and drag Noul into Applications.

• Every build is signed with a Developer ID and notarized by Apple. The first time you open it, macOS asks you to confirm, as it does for any app downloaded from the internet.
• Auto-update arrives in a later release. Until then, download Noul again to get a new version.

Every decision Noul makes in a turn appears as a chip in the chat, drawn from the decision record rather than written by the model.

• Intake shows the intent and its confidence. Gather lists the selected files with their relevance scores.
• Each tool call shows its risk score and whether it was applied automatically, approved by you or refused.
• The verifier's verdict, the test delta and the regression risk close the turn.
• The chips stay pinned under the final answer, so the reasoning is still there after the reply streams in.
• A Decisions view in the side bar keeps the complete log for the session, and /log prints it in the chat.

• Edits the agent makes open as reviewable diffs in the editor. Accept or reject each file before it lands.
• Tab completions appear only when Jev judges they are likely to be accepted, so ghost text stops interrupting you.
• A running turn can be cancelled at any stage, including mid tool call.

Bring an Anthropic or DeepSeek key and Noul's chat and completions run on your account. The same models are available to any extension that uses the language model API.

• Questions that only need to read code get an answer without the edit and verify steps.
• Commands to set the TypeSafe, DeepSeek and Anthropic keys, stored in your keychain.

Noul ships as a full desktop editor with a Jev-native agent built in as the default chat.

• Jev understands each request, picks the files that matter, weighs every action and judges the result.
• Four autonomy levels, Manual to Background.
• Text planted in files or tool output cannot steer the agent, and secrets are kept out of the model's view.
• Verification runs your tests against a baseline taken before the task, so only new failures count.
• A command-line runner takes the same agent to CI and scripts.