Changelog

Added

• Queue work for a computer that is offline: When the computer you are working on is offline, the web app lets you queue what you typed for it, and that computer starts it in the same workspace as soon as it is back online. Queued work can be cancelled until it starts, and it opens by itself once it does

Fixed

• Large text files you attach are read: A log or text file over 2MB is now read from its start instead of being refused, so the assistant can see what you attached
• The architecture graph draws again: The graph showed a blank canvas on the web; its nodes and links now appear

This release adds two things you can ask for by name — installing a skill, and searching inside documents — and makes long conversations summarize more completely.

Added

• Install a skill for Neuron. Ask Neuron to install a skill from a repository or a folder, and it fetches it, checks it, and puts it where Neuron reads skills from — not in another tool's folder, which Neuron never looks at. It works from the next message, and it stays installed when the extension updates.
• Search inside documents, not only code. Word, Excel, PowerPoint, OpenDocument and Jupyter notebook files can now be searched, so a question about a spreadsheet's numbers can find the file and open it at the right place. A large document works too, including one with pictures in it. PDFs can be opened, but their text is not searched.

Changed

• Long chats are summarized completely. When a conversation no longer fits, the older part is now read all the way through before it is summarized. Before, only the most recent part was read and the rest was left out without telling you.
• A summary that was cut off is finished rather than kept half-written, and if no summary can be written at all, Neuron tries your chat's own model before giving up.
• Neuron+ Adaptive Instructions never states the opposite of what you wrote. If your message says something was not used, a rewritten version can no longer describe it as if it had been. Your previous reply is also kept from adding details you never mentioned, and it is shown to fewer messages.
• The rewritten plan no longer asks you questions. The assistant was reading them as facts about your project, so they are kept out of the conversation. A new Draft steps tile shows a short outline on the card instead, for you to read.
• Smarter effort choices. Neuron now takes your recent turns into account when deciding how hard to think — including whether an earlier answer ran out of room — along with the files and attachments your request mentions. A helper that answered quickly before is asked first, so the decision is faster.
• Clearer message when web search is unavailable. A failed search now repeats the service's own reason and says whether the search itself or the service is at fault, instead of "Search failed. Try again later."
• Fewer repeated lookups, and replies in your language. Neuron is told to gather its lookups together, to reuse a result it already has instead of asking for it again, and to answer in the language you wrote in.

Fixed

• A document Neuron cannot read is reported instead of quietly missing. Such a file used to be skipped without a word, so it simply never appeared in search results.

Added

• See how your code fits together: A new architecture graph shows the structure of your project and how its parts connect, so you can understand a large codebase at a glance instead of reading file by file
• Your assistant can search documents and reference material: Documents and reference workspaces are now indexed alongside your code, so answers can draw on written material and not just source files
• Open your cloud computer for you: When no computer is online, the web app starts one for you and opens it, instead of leaving you on a dead screen with nothing to do
• A moving title while a session runs: A workspace that is busy now shows its title with a gentle shimmer, so it is obvious which session is still working

Improved

• Prompt improvements stay honest: The prompt booster never presents something told to it as already true, so a corrected message no longer comes back with the same mistake folded in
• The booster card shows more and asks less: The card shows which skill was chosen and drafts the steps for you to review, and it stops interrupting with clarifying questions
• One language rule for everything generated: Titles, suggested prompts, branch names, and summaries now all follow the same reply-language setting
• Host status recovers on its own: After a flaky network, the web app shows the connected computer again within a minute without a manual refresh

Fixed

• Relay stays in sync on connect: A newly connected control socket is always answered with the current state, even when no computers are connected yet
• Speech setup troubleshooting: Added a guide for when speech recognition setup gets stuck

Added

• Open your desktop sessions in the browser: Signed in with Neuron, the desktop app links this computer to your account so the Neuron web app can see and continue its sessions; Access sessions from the web is always on
• Read-only when the computer is away: A session on a computer whose relay is disconnected stays visible in the browser but cannot be sent to, and its history shows the most recent quarter of sessions until it reconnects
• Web app under a sub-path: The web build can be served under a path such as /cloudapp
• Relay under a sub-path: A relay endpoint can include a path, such as platform.sadata.dev:443/nr-relay
• Install skills into Neuron: Added an install skill tool that installs a skill into Neuron's own skills folder instead of other agents' folders

Improved

• Compaction follows Settings live: Changing the compaction mode or model applies to open sessions without restarting them
• Large conversations are summarised in full: Compaction reads a long history in passes instead of keeping only the newest part
• Summaries survive a failing model: When the chosen compaction model fails, Neuron retries with the chat's own model and then the on-device model
• Clearer web search failures: A failed web search now reports the service's reason and what to try next

Fixed

• Grep and glob false empty results: Fixed "No matches" when the path was a file, a missing path, or an ignored folder

Fixed

• Compacting a long conversation completes: /compact on a large session no longer sits on "Compacting..." for minutes when the summariser's model returns an empty reply, and it no longer drops a third of the conversation without saying so

New

• Library documentation for your assistant: Your assistant can now search, outline, and read official library and framework documentation directly in conversations, display cited documentation as chips, and automatically check docs before modifying unfamiliar code
• Interactive charts: Your assistant can now generate bar, line, and pie chart images directly in conversations with verified data and clear visual previews
• Date and time tools: Exact dates, current times, time zone conversions, and business-day calculations are now answered directly without guessing
• "Signal" activity overview: A redesigned, clean activity overview replaces older cards on the home screen, highlighting streaks, active days, output tokens, and shifts in your model usage over time
• Pixel mascot and live activity animations: An animated pixel mascot ("Octo") acts out live assistant activity on tool rows (reading, editing, building, subagents) and rests asleep when finished, accompanied by refreshed status indicators and clean activity labels
• Tokens-per-second speed indicator: Message footers now show live and final token generation speeds alongside token counts and duration

Improved

• Locked desktop window zoom: The desktop app locks to standard 100% zoom across navigation, launch, and reloads, preventing accidental trackpad pinch or shortcut zoom jitter
• Prompt booster summary tile: Summaries in the prompt booster card now stay complete without truncating early, neatly balanced alongside goal and scope tiles
• Clearer effort and model indicators: The message footer now displays the exact model and reasoning level used (such as Adaptive (High)), pinned to each completed turn so changing models later doesn't alter past message details
• Smarter automatic effort estimation: Your assistant gauges task complexity by reviewing previous context and workspace evidence, trying the fastest models first for instant responses
• Better prompt improvements: The prompt improver now lays out goals, scopes, and summaries in a balanced card layout, cleanly carries forward short follow-up requests, and focuses directly on the task
• Faster and unified web search and fetch: Web searches and page fetching now connect through unified platform endpoints with faster result handling and reliable authentication
• Lower cost for background tasks: Automatic background jobs like naming chats, generating branch names, and suggestions now explicitly run at the lowest effort tier to keep token usage down
• Prompt cache stability: Workspace files and paths are frozen per turn so the first request never unnecessarily rewrites conversation history or invalidates prompt caches
• Commit attribution: Commits made by your assistant now record the model name and effort tier in the commit details

Fixed

• Side-job cloud fallback: Chat titles, workspace names, and suggestions now reliably fall back to cloud providers when on-device models are turned off
• Persistent model labels: Model names in turn footers maintain clean, accurate display labels across app restarts and session switches
• Web fetch connection fallback: Web fetching now automatically falls back to local direct fetching when gateway network connections or IP requests are refused
• Documentation keyword matching: General words like "html" no longer mistakenly pull in unrelated documentation, and documentation attachments have bounded wait times
• Non-vision model safety: Image analysis tools are cleanly hidden when using models that do not support images
• Model routing recovery: If an automatic router selects an unavailable model, requests now automatically retry once

Improved

• Sub-agent and skill use is measured: Higher effort levels now spend effort on depth instead of delegating to many sub-agents or repeating skills, with a limit on sub-agents per turn by effort level
• Repeated checks answered from their earlier result: When nothing has changed since a passing check ran, the earlier result is reused instead of running it again

Fixed

• Sends during setup no longer look failed: A message sent while the prompt improver or other setup is still running now stays in the conversation, clears the composer, and is not sent twice
• Faster commits and fetches in the background: Neuron's commit helper no longer slows git operations that touch many references, so large fetches finish in normal time
• Sidebar shows sending: A workspace with a message still being delivered shows as running instead of idle
• Loader and test fixes: The loading animation no longer crashes when a theme color is not yet ready, and browser tools are correctly on by default

This release adds date and time answers you can trust, and charts you can ask for.

Added

• Ask for a date or a time and get the right one. Neuron can now answer "what day is it in Jakarta", "what time is that in New York", "how many working days until the 31st", or "what is 90 days from now" by calculating it rather than guessing. Your own time zone is used unless you name another.
• Draw a chart from your numbers. Ask for a bar, line or pie chart and Neuron checks the figures first, then draws it and shows it in the reply. Figures that do not add up are reported instead of drawn.

Changed

• The rewritten plan keeps to what you can see. It no longer suggests a thinking level of its own, so the level used for a question stays steady and matches the one shown on the answer. Plans are still written as outcomes rather than orders.
• Fewer useless tools offered to a model. A model that cannot read images is no longer offered the image-reading tool, which saves room in the conversation.
• Plainer explanations. Neuron is now told to explain what it did in everyday language, without internal jargon.

Fixed

• A model that is briefly unavailable no longer fails the request. When the automatic model choice lands on one that is not available at that moment, Neuron asks again once instead of stopping.

New

• Prompt improving is always on: Every request is restated before the assistant starts, so it works from a clear brief. The switch now lives in Settings and is locked on; you still choose which model does it
• See what was improved: A restated prompt appears as clear sections — goal, scope, limits and what "done" means — instead of quietly replacing your words. Your original is one tap away, and it is still what the assistant receives
• It knows what you attached: Attachments are named to the prompt improver — an image, a log file, an issue — so "work on this issue" becomes a real task instead of a vague one
• Long chats remember the work: Neuron keeps a running summary of the files changed and commands run, so a conversation trimmed for length no longer forgets what was already done
• Compact tool output: Large structured results, such as API responses, are stored far more compactly without losing a single value
• Image generation through your account: Image requests now use your signed-in models, with automatic failover if one is unavailable
• Clearer thinking: Thought headers now say what the assistant is doing — "Reading message.tsx", "Searching the code" — instead of showing file paths and patterns
• Clearer worktrees: A worktree shows one name — its branch — and how far it has drifted from its base, so you can see what is ready to merge and what will conflict
• Tidier change panels: Edit and Write panels follow your theme, show the change count, and keep the +/- markers lined up and out of copied text

Improved

• Prompt improving never gives up early: Your device and your account now work on the brief at the same time, and a nearly finished answer is kept instead of discarded when time runs out
• Fewer repeated charges: Neuron no longer rewrites its own cached work in the middle of a long run of tools, so unchanged content stops being billed over and over
• Lower memory: Speech recognition releases its memory when idle, and the built-in assistant no longer leaves a process running when it is switched off mid-load
• Voice input works on a fresh Mac: Speech recognition now sets itself up correctly the first time, and its helper shuts down cleanly when it hangs or its parent is gone
• First-run window: The setup window can be resized, and its layout adapts to the window instead of overflowing
• Steadier connection: Conversations are now routed so that follow-up messages reach the same server, and a brief network hiccup no longer clears the model list
• Better on-device results: The built-in assistant's working space was widened so longer requests, together with your previous reply, still fit

Fixed

• Improvements timing out: Fixed the recurring timeout — including the very first message after starting Neuron — and a case where a valid improvement was thrown away
• Improvements in your language: Requests in Indonesian now get Indonesian results instead of being translated, and the assistant no longer invents limits you never asked for
• Applies to your words: The improvement is held to your actual request, so a simple rename no longer gets a project-analysis treatment
• Effort nudges, not overrides: When improving a prompt, the suggested effort can move by at most one step
• "Continue" keeps the plan: Prompts like "lanjutkan" or "yes, do it" now carry the agreed plan forward instead of being treated as a fresh request
• Improvements that vanished: Several cases where a valid improvement was dropped, or the result disappeared after a refresh, now behave correctly
• Picks up a switch immediately: Turning on-device models on or off now updates every model list right away, instead of leaving stale entries behind
• Thought duration returns: Once a thought finishes, its header shows how long it took again instead of holding the last topic

New

• Better instruction improvements: Your instructions are now improved by the strongest available model, with the on-device model as a backup, so the result stays closer to what you wrote
• See the request sent to the assistant: A code button on each message opens the exact text the assistant received for that turn
• Smarter effort for each request: Neuron now reads how demanding a request is from the Hub's System One models, so vague but serious requests are no longer treated as quick ones
• Simpler settings: Permissions for Neuron tools are always on, and the agent-defaults dialog no longer opens by itself after an install or update

Improved

• Clearer improvements: Improvements are written as outcomes you can see or run, instead of orders to the assistant
• Speech setup that recovers: The first-open speech setup no longer stalls on an unstable connection; a stuck step is stopped and retried once automatically
• Clearer status while improving: The line now reads "Improving your instruction" while the improver works
• More accurate improvements: Improvements no longer name files that only share a word with your request

Fixed

• Web search: Searching the web works again after a change that broke it, and a failed search now says the service is unavailable instead of a confusing parse error
• Related-file guesses removed: Neuron no longer sends files that were only matched by name, which had pointed the assistant at unrelated files
• Bubble restored: The improvement bubble returns to its previous layout

This release keeps the rewritten plan honest: your own words always travel with it, and it no longer points the assistant at guessed files.

Changed

• Your own words stay with the plan. A rewritten request now ends with exactly what you typed, and Neuron is told that your words are the request and win wherever the rewrite adds something you did not say. Files guessed from the words of your request are no longer sent or shown — they often pointed the assistant at unrelated files. A rewrite also describes the outcome you want rather than giving orders, and a question, or a request to explain something, is left as you wrote it.
• A clearer message when web search is unavailable. If the search service answers with a web page instead of results, Neuron now says the service is down instead of showing a confusing error, so it stops retrying the same search.
• The line shown while a request is being rewritten now reads "Improving your instruction", or "Menyempurnakan instruksi" in Indonesian, matching the desktop app.

This release brings the desktop app's way of choosing how hard to think to VS Code: one level per question, kept steady for the whole answer.

Changed

• One thinking level per question, kept steady. Neuron decides how hard to think once, for the question you typed, and every step of that answer uses the same level. Before, the level could jump up and down while a single answer was being written. The choice comes from your account, reading your question together with the previous exchange; if that is not available, it comes from what your question refers to in your project. When an answer runs out of room, the next two questions get one step more. A slow or failed reading no longer turns the feature off, and the Neuron Companion log says why when no level came back.

Fixed

• The models that work in the background no longer appear in the Chat list. The one that judges how hard to think, and the ones that make images and video, are not chat models, so choosing one only led to an error. They are hidden now.

This release brings the request improver from the desktop app to VS Code, adds image generation, and moves Neuron to its new web address.

Added

• Neuron+ Adaptive Instructions, on by default. Before a Neuron model answers a question you type, Neuron rewrites it as a short plan: what you want, what is in scope, the limits, and how to tell it is done. The plan is added below your own words, never in place of them, and the assistant is told your words win where the two disagree. It runs once per question, not on every step, and gives up after 10 seconds — if it fails, your question is sent as you wrote it. Questions, slash commands and requests that are already clear are skipped. You can turn it off in settings.
  • You can see what it did. An answer to a rewritten question opens with a short quoted note showing the plan that was used: its goal, scope, limits and what "done" means, in Indonesian if VS Code is set to Indonesian. The note also names the skill it picked and any files you attached, and if nothing needed changing, one line says why. Background steps get no note, and the note is taken out of the conversation before the next request, so it is read once.
• Neuron: Improve Prompt. Rewrites a prompt — the text you selected, or whatever you type — and puts the result in the Chat box so you can review and edit it before sending. It works whether the automatic rewriting is on or off.
• Image generation in Chat. Ask for a picture and Neuron draws it with one of your account's image models, trying the next one if a model fails. The picture is saved in your project and shown in the reply.

Changed

• Neuron moved to its new web address. Sign-in, models, chat, memory, artifacts and web search all use the new address. An address you had set by hand to the old one is replaced automatically.
• Long chats keep an exact record of what was done. When older messages are shortened to fit, the summary now lists the files that were changed, each terminal command with its result, and your earlier requests — taken from the actions themselves, so it stays right even when the written summary does not.
• Large data results take up less room. Long lists of data returned by tools are written more compactly, with the parts every row shares written once. Nothing is lost, and Neuron checks the result before using it.
• Lower cost for repeated requests. The parts of a request that repeat are now sent identically every time, which lets the service charge less for them. Tool lists are always sent in the same order, so a server that reorders them after a restart no longer forces Neuron to send everything again.

New

• Prompt booster: Neuron now restates each request as a clear brief — goal, scope, limits and what "done" means — before the assistant starts, so it works from an unambiguous spec. On by default, with a toggle in the composer
• Setup for voice: The speech models and runtime now download during first-time setup with a visible progress bar, instead of quietly in the background

Improved

• Smarter code search: Search answers from exact matches first and only falls back to meaning-based search when needed, so results are faster and more predictable
• A more complete index: Preparing search now skips built and third-party files and follows your own ignore files, so more of your real project is covered instead of being cut short
• Much smaller setup: The models behind search are far smaller, bringing setup downloads down from around 800 MB to around 180 MB, and the first index now takes minutes rather than about a quarter hour
• Startup experience: The startup screen follows your light or dark theme, shows what is downloading, and reveals your workspace as soon as setup finishes

Fixed

• Dropped tool details: For the models that occasionally drop the details of a batch of tool calls, Neuron now remembers and asks for one call at a time — across new chats and restarts, instead of relearning it every session

New

• Customize page: Memory and Skills now share a single page, and Memory has moved out of Settings to a place of its own
• Smarter code search: Search now understands what code means, not just which words it contains, so you can find things by describing them
• On-device assistant: The built-in assistant that names chats, workspaces and suggestions has been rebuilt to give better results, entirely on your machine
• Diagnostics: A new window shows whether Neuron is working, with five plain-language checks and what to do when something is wrong
• Design studio: New tools for inspecting, previewing and sharing your designs
• Cleaner library: Artifacts and Skills have a tidier layout with filters, sorting, a details panel and bulk selection

Improved

• Skills in one place: Your own skills and the ones built into Neuron appear in a single list, each with a working on/off switch, and a whole set can be removed at once
• Downloads and speed: Model downloads are several times faster, can resume, and use a smaller, equally accurate version
• Stays responsive: Neuron no longer scans throwaway files while preparing search, so it stops spiking the processor every half a minute
• Smoother while indexing: Preparing search no longer makes the app feel sluggish, and the animated background now stops when nothing is running
• Browser control: Browser tools are always available and keep working even if the browser connection drops
• Reconnects on its own: If the local service moves, the app finds it again instead of getting stuck
• Memory stays on: Memory is now always on, so nothing is lost by accident
• Model picker: On-device models are clearly marked, no longer listed twice, and "Automatic" is always offered first
• Smaller footprint: The assistant carries less behind the scenes, leaving more room for your conversation

Fixed

• Audit records stay private: Audit no longer stores what your tools returned. Command output, file contents, pages and sub-agent logs are replaced by a short summary, so nothing sensitive is kept
• Your downloaded models: The app no longer overwrites or deletes them by mistake
• Flaky connections: A slow connection no longer breaks the model list — it retries quietly instead
• Text edits: Edits now match curly quotes, dashes and ellipses, and point to the exact line that differs
• Opening Customize twice: No longer crashes
• Installing a skill set: No longer fails when the set includes a skill Neuron already ships
• Titles and names from the on-device assistant: They no longer fall back to a cloud assistant, and correct answers are no longer discarded
• Duplicate errors: Tool results no longer show the same error twice

Added

• Skills library in Settings: A Discover tab browses the skills.sh community directory in the app, with search, sorting, categories, repo pages showing a skill's description, files and every skill in the repo, and one-click install of a whole repository
• Your skills: An installed list beside Discover, showing each skill's name and description with an on/off switch, editing, paging and an uninstall that removes every skill a bundle installed
• One place for installed skills: Skills now install into a single Neuron directory that the daemon reads, so an installed skill reaches the agent's / menu and its read allowlist instead of being copied into three agent homes the daemon never scanned
• Subagent cards: Delegated subagents render as their own cards in the chat
• Shell timeout setting: A shell command can run without a fixed timeout when the agent needs longer, and filesystem tools respect the sandbox's writable paths
• Content-aware shell output: Shell results render ANSI colour, diffs and code with the shared syntax highlighting

Improved

• Bundled skills rewritten: The shipped skills are versioned and split into focused submodules, and the new /wiki skill documents a codebase into a linked knowledge base
• Skill documents: Skill descriptions written as folded or literal YAML blocks now read correctly, and the skill file viewer shows code with line numbers and horizontal scrolling
• Compaction timing: The transcript folds on the prompt it measures, instead of counting the thinking budget it reserves, so a large reasoning budget no longer folds a conversation early; the fold row and context meter report the real sizes
• Speech: The default speech-to-text model is parakeet-redux

Fixed

• Installed skills missing from the / menu: The retired per-user skill store is now folded into the scanned directory at startup, instead of only when the Skills screen opens
• Bundled skills in Your skills: Shipped skills no longer appear in the user's own list, where editing or deleting them appeared to work and then silently reverted
• A disabled skill that still loaded: Turning a skill off now moves it out of the directory the agent scans, so off really means not loaded

Added

• Spark-X2.5 on-device chat model: Neuron SelfAgent+ now runs conversations on device with a 256K context window, streaming replies, tool calls and visible thinking, while short helper jobs stay on Yi-Coder
• On-device models switch: Settings → Agents has a switch that stops the local model workers and sends their jobs to the Hub, and keeps the downloaded files so switching back on is instant
• Side-job defaults: Titles and branch names run on the on-device helper, commit and PR text goes to the Hub router, and next-prompt suggestions are now opt-in
• Agent defaults onboarding: A one-time dialog after the first model load explains the new agent defaults
• Add folder: The new agent tab and new workspace screen can add a folder to the workspace
• Plan as a slash command: /plan runs a single message read-only, in place of the old Plan mode

Improved

• Simpler mode picker: The Plan and Neuron Claw modes are gone; the picker offers Chat, Manual, Accept edits, Auto and Bypass
• Model memory: The on-device models free their memory when idle, and the reranker now powers memory recall
• Streaming replies: Replies write in continuously instead of a word at a time, and thinking appears as it streams
• Task display: Every task step shows in the chat with only the plan expandable, finished jobs keep their steps, and a task reads the same name everywhere
• Sidebar layout: Desktop sidebars and collapsibles opened and closed with animation, with the sidebar toggle and traffic-light padding fixed
• Structured prompts: Skills, MCP environment, workflow stages, sub-agent roles and user turns are wrapped in XML tags
• Prompt cleanups: Retired the automatic mode handoff, and removed the Splash engine and Plan mode plumbing

Fixed

• Web search date filter: A date-filtered search no longer comes back empty
• Spark thinking tags: The on-device model no longer leaves </think> in replies after a tool call
• Cancelled setup turns: A turn cancelled during setup no longer leaks its prompt
• Gateway auth errors: A 5xx that wraps an upstream auth failure is no longer retried
• Worktree folder names: Worktrees are named from the branch or prompt instead of a random slug
• First-message images: The image sent with the first message of a new chat or design now shows
• Duplicate auto models: The Settings model list no longer lists the same automatic model twice

Added

• Agents settings page: Settings → Agents now holds every agent default in one place, including default model and effort, permissions, background models and agent profiles
• Neuron SelfAgent+ on-device models: One switch runs local models that write session titles and workspace and branch names, using Yi-Coder 1.5B in place of Qwen 0.5B
• Automatic on-device memory: Relevant memories are recalled into new messages, and worthwhile exchanges are saved after a turn, all on device and skipped when the text looks like it holds a secret
• Background model choices: Pick separate models for titles, workspace names, commit and PR text, suggestions, image descriptions and conversation compaction
• Image policy: Choose whether images go to a separate vision model always, only after asking, or never
• Scoped permission grants: The Allow menu can grant a command prefix or edits in one folder for the session, and chained or redirected commands can only be allowed once

Improved

• Model picker grouped by maker: Long Hub catalogues are grouped under headings such as DeepSeek, Z.ai and Xiaomi, while search results keep their ranked order
• Automatic desktop updates: Updates now always install automatically on quit or when the app is idle, and the Settings switch is removed

Fixed

• Composer stuck on running: The composer no longer shows a turn as running after it has finished
• Collapsibles snapping shut: Tool groups and thinking blocks now animate closed instead of disappearing instantly

Added

• Memory cards in chat: Saved, recalled, updated and deleted memories now appear as structured cards with their content and search result
• Activity labels on tool calls: File edits, searches and fetches can show a short description of the work directly in the conversation

Improved

• Grouped tool-call summaries: Groups now distinguish reads, searches, file changes and shell commands, with file paths and line changes where available
• Agent task progress in chat: Todo updates no longer add a row for every task starting or finishing, and finished task lists collapse into one summary
• Background task output previews: Task output is cleaned of terminal color codes and progress redraws in the task panel
• Hub conversation audit: When audit is enabled for an account, the Hub records conversations and Neuron sends only tool results it cannot see itself
• Expandable chat sections: Tool groups, task lists, thinking, subagents and workspace panels now expand and collapse with motion that respects reduced-motion settings

Fixed

• Duplicate tool calls in memory workflows: Removed overlapping cloud-memory tools so each fact is saved once through local memory sync
• Todo plans lost on a refused update: Invalid task lists are checked before the current list is cleared
• Tool summaries mistaking option values for file paths: Commands such as head -n 20 README.md now classify the filename and option value correctly

This release adds web and GitHub research, image reading for text-only models, and a summary fallback for long conversations.

Added

• Look things up on the web from Chat. Two new abilities let any Neuron model research: one searches the web, the other reads a page you point it at. Pages on your own computer or private network are refused, and every redirect is checked again, so a web page cannot send Neuron somewhere private.
• Search GitHub from Chat. Neuron can search code, repositories, issues and pull requests on GitHub, using the GitHub account you are already signed in to in VS Code. There is no token to set up. The first time, VS Code may ask you to allow it.
• A model that cannot see can still read your images. Point Neuron at an image in your project and one of your account's vision models writes a report for it: the exact text, the layout, and any diagrams.
• Choose which models work in the background. In the Models tab you can decide whether images may be sent to a vision model (Always, Ask each time, or Never), which vision model reads them, and which model summarizes long chats. Every choice has a sensible default.

Changed

• Long chats are summarized instead of silently cut. When a conversation is still too long for the model, Neuron writes a summary of the older messages it has to drop — including the tools that were used — and keeps it at the start of the conversation. This happens only when messages would otherwise be dropped, and the same summary is reused until the chat grows again. If no summary can be written, Neuron falls back to trimming. VS Code's own shortening still runs first.
• The audit trail moved to Neuron Hub. For accounts with auditing turned on, Neuron Hub now records your chat conversations itself. VS Code no longer sends your prompts and replies anywhere for this; it sends only what the tools returned, using your Neuron sign-in. Messages VS Code writes on its own, such as chat titles and commit messages, are marked so they are never recorded. You can still turn all of this off in settings.

Removed

• The second Cloud Memory key. VS Code no longer stores a separate key for the audit trail, and it deletes the one that earlier versions kept.

Added

• Artifacts published to Neuron Hub: Published pages now upload to the Hub so links work beyond the publishing machine, falling back to the local link with the reason when the upload is skipped
• Artifacts gallery: New Artifacts screen with All, Yours and Shared tabs, search, list and grid views, type filters, and Private/Public switching per page
• In-app artifact viewer: Artifacts open inside the app with a toolbar for privacy, sharing and copying public links
• Sync local artifacts to the cloud: Pages that exist only on one host can be uploaded to the Hub from the gallery or the viewer
• Usage card on new agents: New agent tabs and the new-workspace screen open on a greeting and a usage card with sessions, messages, output tokens, a 26-week heatmap and per-model breakdown

Improved

• Usage scoped to the workspace: A new agent's usage card shows its own workspace first, with a switch to the whole host
• New-workspace onboarding: The new-workspace screen now matches the new-agent layout, with Import session, a dismissible tip, and project and branch chips
• Draft agent controls: Draft agent controls now match the live composer layout

Fixed

• Duplicate Cloud Memory entries: Each memory now reaches Cloud Memory once instead of once per tool call or under a second identity
• Images sent while steering: A screenshot sent during a running turn now reaches the model instead of being dropped
• Hover on the new-workspace screen: Fixed hover not responding on desktop

Added

• Share artifacts with signed-in Hub users: A new Hub users level sits between Private and Public, so a page can be shared with anyone signed in to the Hub without making it public
• Shared with you: Artifacts other users share with you appear in their own tab and open with an owner caption, and an Open link dialog accepts app links, Hub links and bare ids
• Agent task list: The agent keeps a todo list for multi-step work, shown in chat and in a new Tasks tab in the desktop sidebar, which opens when the agent starts a new list

Improved

• Daemon memory use: The daemon keeps archived agents' history on disk instead of in memory, and stores streamed replies compactly, cutting its resident memory from about 1.4GB to about 55MB on a machine with hundreds of agents
• Memory tools: The agent is offered only the local memory tools, which already sync to the cloud, so each fact is saved once
• Temporary files: When the shell sandbox is on, write and edit can reach the temp directories the sandboxed shell can already write, instead of pushing the model onto the shell

Fixed

• Daemon running out of memory: Fixed the daemon crashing roughly every 17 minutes on large workspaces by capping codebase index memory and keeping at most six indexes loaded
• Commit identity for agent commits: Commits the agent makes through its shell now carry the same committer and co-author trailer as the commit button, while the author stays the user

This release brings Neuron Hub artifacts into VS Code. The assistant can publish a page from your project, and you can browse, open and share your artifacts without leaving the editor. It also moves Neuron memory to Neuron Hub.

Added

• Artifacts tab. Run Neuron: Artifacts from the Command Palette, or press Artifacts in the Neuron sidebar, to open your artifacts in an editor tab. Switch between All, Yours and Shared with you, search by title, filter by type (documents, slides, dashboards, designs, pages), and choose a list or a grid of previews. The tab remembers the view you picked, and groups artifacts by when they were last edited.
• Open an artifact without leaving the tab. Click one to open it in the same tab, with a back arrow to the gallery. The page runs in a strict sandbox: it can run its own scripts, but it cannot reach VS Code, your sign-in, or the network. This also works when your Hub serves artifact pages from their own address — the tab learns that address from the Hub's replies, allows only it, and forgets it when you sign in with a different account.
• Three levels of who can open a page. Every artifact is Private (only you), shared with Hub users (anyone signed in to Neuron Hub who has the link), or Public (anyone with the link). Change it from the ⋮ menu or the Share menu in the viewer. Neuron asks before a change that lets more people in, and never before one that narrows access.
• Share with other Neuron Hub users. Copy link gives out the public link for a public page, and an app link for a page shared with Hub users — that one opens for people signed in to Neuron Hub in the desktop app. Use Open link… to open a link someone sent you. Once you have opened one, it appears under Shared with you until you remove it.
• Publish a page from Chat. The publish tool uploads a self-contained HTML page from your project to Neuron Hub. VS Code asks you to confirm every publish, since the page leaves your computer, and a new artifact starts as private. Publishing the same file again updates the same artifact rather than making another one, and a notification offers to Open it. Files outside your project folders, files that are not HTML, and files over 5 MB are refused.

Changed

• Memory now lives in Neuron Hub. The Neuron+ Cloud Memory service that Chat uses for long-term memory is now run by Neuron Hub and uses your Neuron sign-in. There is no separate sign-up and no second key, and memories you already had were moved over, so recall works as before. Signing out stops the service instead of leaving it running with a key that no longer works. The sidebar's Cloud Memory row tells you when you need to sign in again.

Removed

• Index Codebase, along with the code map, team wiki and published-skill tools that depended on it. They were served by the old memory service, which memory has moved away from.

Fixed

• Long sessions that could no longer continue: A conversation that hit tool_use ids were found without tool_result blocks now repairs itself instead of failing on every turn from then on. Some providers issue short numeric tool-call ids, and a long session can be handed the same one twice; because a tool result is matched to its call by id alone, the second call looked unanswered even though it had been answered. The repeated call is now given an id of its own, along with its result, so no tool output is lost

Added

• Animated chat and workspace title reveals: Types newly generated chat and workspace titles into tabs and sidebar rows with a smooth transition instead of a jarring instantaneous swap
• Optional questions in user prompts: ask_user questions can now be marked optional and omit free-text inputs for closed multiple-choice questions
• Clear finished background tasks: Completed background shell commands can now be cleared from the task list and turn footer

Improved

• Workspace-grounded adaptive effort grading: Resolves named paths against the workspace index and weighs terms with BM25 rather than matching regex keywords, with automatic escalation when turns hit token ceilings
• System prompt guidance for outline reads: Documented read outline mode in the system prompt so agents discover declarations and structure before reading full files
• Clearer git commit failure reporting: Captures trailing git hook stderr output to surface exact failure reasons when commits are rejected
• Cleaner turn footer metadata layout: Separated durable facts (effort tier, tokens used, task counts) from rotating activity labels and kept effort indicators visible after turn completion

Fixed

• Turn capability mode pinning: Pinned toolsets and capability modes to the active turn id so mid-turn mode switches do not cause tool dispatch rejections
• Socket leaks on stalled streaming requests: Aborts stalled Hub streams on watchdog timeouts to release connection sockets from the pool
• Composer attachment loss during send: Retained image attachments across the clear-and-encode window to prevent premature garbage collection and dropped pictures
• Duplicate tabs on slow agent creation: Handled delayed agent creation responses gracefully to prevent opening duplicate draft tabs or spawning twin agents
• Trailing assistant messages in transcripts: Automatically closed transcripts ending on an assistant message to prevent gateway prefill rejections
• Provider discovery lockouts after network drops: Retried transient connection errors during provider probing and refreshed cached failures after cooldowns
• Shell command working directory instruction: Clarified in tool schemas that shell commands already execute in the workspace root, eliminating redundant cd command prefixes
• Sidebar draggable list gesture lockups: Resolved gesture state livelocks when reordering items in the sidebar

Added

• Expanded code intelligence tools: Added directory listing, syntax diagnostics, symbol usage search, and GitHub code/repository/issue search for faster repository exploration
• Notebook editing tools: Read and edit Jupyter notebooks while preserving notebook structure and cell metadata
• Workspace task tools: Discover and run project tasks from workspace configuration
• Background task descriptions: Background shell commands now appear with human-readable titles and can be understood at a glance in the task panel
• Public Neuron tool reference: Added comprehensive documentation for the built-in tool surface

Improved

• Workspace-grounded adaptive effort classification: The agent now bases effort decisions on repository structure and named files rather than relying only on prompt vocabulary
• Reliable mid-reply turn recovery: Retries interrupted provider streams without replaying already-delivered text or re-running completed tools
• Cold-start navigation: Opening the app now consistently lands on Home with the sidebar closed
• Subagent input safety: Prevents interaction with a running subagent while its turn is active
• Attached image context: Marks images with the message they belong to so models can distinguish new attachments from earlier images
• Opt-in ignored-directory search: grep can explicitly search ignored paths such as node_modules and dist when needed
• Tool documentation and task guidance: Expanded the system prompt and public docs so agents discover the new tools and use background tasks correctly

Fixed

• Transient provider discovery failures: Provider snapshots recover after temporary network failures instead of remaining permanently unavailable
• Divergent Git pulls: Pull now reconciles divergent branches according to configured Git behavior instead of failing with exit 128
• Hub stream cleanup: Releases the underlying socket when a stream watchdog fires
• Composer steering state: Prevents stale steering locks and keeps attachments stable during send transitions
• Tool mode availability: Runtime tool catalogs now stay aligned with the modes actually served by the daemon

Added

• Claude Code-style background task indicator and modal: Track running shell tasks in turn footers and inspect output or terminate processes from a dedicated task modal
• Outline read mode: Added an outline mode to file reads that extracts declarations and signatures using Lezer parsers to compress token footprint
• Background cloud memory synchronization: Automatically mirrors local memory additions, updates, and removals to Neuron Cloud Memory in the background

Improved

• Upstream security and stability updates: Enforced fsmonitor restrictions on daemon Git commands, rejected weak X25519 keys, upgraded Electron to 44.4.3, batched GitHub PR status polling via GraphQL, and vendor pure-Lezer parsers
• Calibrated reasoning effort postures across all tiers: Tuned effort levels from low to max with distinct engineering disciplines, preventing scope calibrators from demoting max-effort classifications
• Model reasoning on command prompts: Surfaces the model's declared intent and reasoning when prompting for command approval

Fixed

• Subagent model inheritance: Delegated subagents and workflow stages now inherit the caller agent's model instead of defaulting to auto
• Sandboxed Claw mode enforcement: Claw mode now refuses or safely downgrades when required OS sandbox primitives are unavailable instead of running unsandboxed
• Symlink traversal in file mutations: Prevented write and edit tools from following symlinks that resolve outside the workspace root
• Subagent permission tier enforcement: Spawned subagents are now constrained by their caller's permission tier to prevent privilege escalation
• Mode catalog discovery: Ensured replacement-model profiles and runtime modes properly preserve mode options and unattended flags
• Browser automation tools default: Restored browser tools to disabled by default across protocol schemas, clients, and daemon configurations
• Skill trigger false positives: Prevented matching skills on casual mentions of the word prompt
• CLI Hub scaffolding paths: Fixed CLI scaffolding to write to .neuron/hub.yml to match deployment expectations

Improved

• Shell command intent in permission prompts: Carried tool descriptions from the model through to shell permission prompts so the purpose of each command is clear before approval
• Cache-stable system prompt ordering: Reordered system prompt parts to place volatile elements like effort tiers and granted folders behind stable session context, preserving 1-hour prompt cache breakpoints across tier changes
• Regex alternation in grep: Updated the grep tool description to recommend pipe alternation for multi-pattern searches in a single call

Fixed

• Model selection on newly created workspaces: Derived model selectability from the account catalogue rather than uninitialized daemon provider snapshots, preventing model picks from being discarded or stuck loading
• Premature context compaction during tool loops: Fixed prompt token inflation and calibration drift that caused small sessions to trigger context folding prematurely

Improved

• Router-based adaptive effort classification: Evaluates full prompts against the router model with increased token room, generous timeouts, and session-model fallback for reliable effort grading
• Early effort evaluation feedback: Displays the reading indicator immediately at the start of a turn while awaiting setup and classification
• Streamlined initial app setup: Bypassed the on-device base model download overlay during startup while the local engine is being reworked

Fixed

• Stray token spending on cancelled turns: Interrupted turns immediately drop execution ownership so abandoned iterations cannot incur token costs or block future turns
• Windows process management and file search stalls: Resolved directory traversal stalls in ignored paths, path separator mismatches, and process tree leaks on Windows
• Credential loss during session imports: Carried authentication credentials through session import requests so imported agents connect without re-authentication errors
• Internal tasks requesting reasoning budgets: Stopped automated background turns like chat titling from requesting thinking budgets or triggering effort classification
• Persistent reasoning disablement on routed models: Prevented isolated thinking budget rejections on the auto router from disabling reasoning for subsequent turns

Added

• Ambient background through startup: Splash, onboarding, and base-model verification now share the same field, with a spinner for verification instead of a misleading progress bar

Improved

• Ambient field follows the active conversation: The field keeps one shared clock across navigation and frames the window edges with temperature-based work signaling
• Reliable adaptive theme transitions: Light/dark changes now refresh themed content consistently across the app

Fixed

• What's New waits until the app is on screen: The changelog dialog no longer opens over splash or sign-in screens

Added

• Prompt history in the composer: Use Up and Down to recall the last 100 prompts, restore an in-progress draft, and keep normal multiline caret movement and command autocomplete behavior intact.

Improved

• Local next-prompt suggestions and chat titles: Hardened the on-device base-model paths against copying prompt scaffolding or conversation text. Invalid local output now falls back to structured generation instead of surfacing an echo or persisting a borrowed title.
• Stable adaptive effort per turn: Adaptive effort is resolved once per turn and held while tool execution continues, preventing system-prompt changes from evicting a large cached prefix.
• Cache-safe tool-result clearing: Clearing now tracks each preset's own compaction threshold and skips until the context window is known, avoiding early cache eviction during model changes or long-running sessions.

Fixed

• Local-model routing quality: Short greetings and sparse conversations no longer receive copied example titles or invalid next-prompt suggestions from the local model.

Added

• Workspace map on the opening turn: Sessions now begin with a compact map of the repository — root files plus one level inside each directory — so the agent knows where packages live without spending several searches working it out. Rendered once in the background, capped so a large monorepo never turns the map into its own context problem, and respects .gitignore.

Fixed

• Prompt cache no longer rebuilt mid-session: The conversation prefix is now append-only across turns, which is what prompt caching requires. Two separate causes were rewriting it — the cache anchor abandoned a large block whenever a turn ended, and the environment context block moved from one message to another as the conversation grew. Both left tens of thousands of unchanged tokens without cache coverage, billed as writes at twice the read price on the first request of each new turn.

Added

• Living ambient background across app modes: Added a fluid ambient background to chat, workspace, and design screens with dark and light palettes and a live beacon during active agent turns

Improved

• Automatic compression for image attachments: Shrinks oversized raster attachments once upon entry to reduce token usage while keeping prompt cache prefixes byte-stable

Fixed

• Duplicate image blocks from repeated attachments: Deduplicates byte-identical image attachments before prompt submission so duplicate pictures are not billed on every turn
• Premature tool result clearing on image attachments: Aligned token saver transcript measurement with the shared compaction serializer to prevent spurious cache eviction on sessions with screenshots

Added

• Claude Code-style floating accessory cards in the composer: Redesigned the tracked changes bar, running subagent track, and task list into standalone floating frosted glass pills with full rounded corners and borders, showing repository context, branch tags, and diff statistics cleanly separated from the prompt input box

Improved

• Continuous ambient background across navigation: Hoisted the ambient simulation clock and elevation state to module scope so orbs and working beacons flow uninterrupted without restarting or jumping when switching between chats, workspaces, and sessions
• Calmer idle ambient motion: Reduced idle animation speed to a peaceful resting cadence while maintaining lively elevation and luminous beacon signaling during active agent turns

Improved

• Global tool result clearing across all modes: Enabled proactive tool result trimming across build, plan, chat, and custom modes once transcripts reach 15% of the context window. Stale file reads, directory listings, and searches are safely compressed while protecting mutations (edit/write), keeping long exploration and planning sessions fast and avoiding expensive compaction spikes.
• Adaptive cache anchor TTL: Extended conversation cache breakpoint lifetimes from 5 minutes to 1 hour once transcripts exceed 20,000 tokens, eliminating repeated cold-cache reads when reading diffs or resuming after pauses.
• System-enforced medium token saver: Locked the token saver to medium by default at the system level and cleaned up redundant UI controls in the model selector, guaranteeing safe and consistent token savings across all user sessions.
• Enhanced Claw engine safeguards: Reinforced the autonomous Claw mode with automated sandbox isolation, post-edit AST syntax validation, and intelligent error trimming to prevent runaway tool loops.

Fixed

• Prompt cache stability: Disabled pxpipe rendering across presets after discovering that relocating static slabs into message payloads evicted system cache breakpoints and broke Anthropic prompt caching (causing repeated expensive cache writes). Static system prompts and tool schemas now remain byte-stable across consecutive turns, restoring 90%+ prompt cache read hits.
• Local base model worker resolution: Resolved local base model worker entrypoints during source runs, ensuring the on-device base model actively serves chat title generation, prompt suggestions, and adaptive reasoning classification instead of silently falling back.
• Adaptive reasoning effort calibration: Tuned adaptive effort detection to use the maximum of the classifier and linguistic keyword heuristics, preventing small model false negatives on non-English technical prompts while keeping lightweight greetings at Low.

Improved

• Prompt cache prefix stability in tool loops: Eliminated mid-turn prompt cache invalidations during multi-step tool loops. Tool result trimming now anchors to settled prompt turn boundaries (keepPromptTurns: 1), keeping active-turn tool results stable. Environment context blocks are pinned to the turn's root user message, and a 4th cache breakpoint is anchored at the settled prompt turn boundary—locking in 90%+ prompt cache read hits across tool execution loops instead of repeatedly rewriting hundred-thousand-token prefixes at 2× write cache cost.

Fixed

• Seamless daemon auto-restart on app updates: Resolved stale daemon persistence when updating the desktop application. Update routines now gracefully stop the running desktop daemon prior to app relaunch, daemon version metadata is persisted directly to the process lock, and startup routines enforce immediate daemon recycling whenever running daemon version mismatches the newly installed app binary.

Added

• Neuron Claw mode: Introducing a new high-agency autonomous mode (claw). It runs unattended without permission prompts and relentlessly self-heals failing checks—treating a failed test, typecheck, or linter as work still to do, and persisting until verification passes before declaring completion.
• Dedicated Design Studio workspace: A dedicated design and prototyping surface with an isolated preview canvas, responsive scaling, artifact type navigation, and integrated design controls.

Improved

• Dramatic token reduction in Code Mode: Streamlined Build mode by pruning 59 non-coding and administrative tools down to a focused 31-tool suite, cutting schema token overhead by ~73%. Trimmed system prompts, filtered the skills catalog to engineering tools, and enabled eager tool result clearing at 20,000 tokens to keep long sessions compact and fast.
• Language-agnostic adaptive effort calibration: Calibrated adaptive reasoning effort detection to dynamically distribute across all five tiers regardless of language, supporting casual slang, Indonesian, and complex architectural prompts.

Fixed

• Bootstrap setup overlay localization: Added missing UI strings for the first-launch local model setup and bootstrap overlay.
• Protocol schema preservation: Preserved adaptive effort classifier schemas across protocol messages and validation boundaries.

Added

• On-device AI model for instant local tasks: Neuron now includes a dedicated local base model running directly on your machine. It powers adaptive reasoning effort routing, next prompt suggestions, and automated chat title naming completely offline with fast response times and zero cloud token costs.
• Persistent reasoning effort indicator: Each turn now features a dedicated marker in the chat timeline showing the exact reasoning effort used, as well as whether your prompt was analyzed locally on your device.
• Seamless first-launch model setup: A clean full-window setup screen prepares the local model once upon update, keeping you informed with a progress bar and friendly status messages.

Improved

• Smart next prompt suggestions: The composer now offers instant, relevant suggestions for your next action right after the agent finishes working.
• Fast and accurate effort recognition: Casual remarks, greetings, and slang are now properly recognized at low effort, while complex multi-file engineering tasks automatically receive high reasoning depth.

Fixed

• Smoother desktop update distribution: Streamlined the automated release pipeline to ensure desktop updates for macOS and Windows publish immediately without waiting on inactive packaging runners.

Improved

• Instant adaptive effort feedback: When using Adaptive mode, the reasoning effort indicator now appears immediately as you send your message, giving instant visual feedback while the local model evaluates the prompt. Selecting a manual effort level bypasses detection entirely for uninterrupted sending.
• Friendly and transparent startup preparation: The initial setup screen verifies local model files and prepares the engine with friendly, non-technical status messages so everything is ready for your first turn.

Fixed

• Reliable background daemon shutdown and update restarts: Quitting Neuron or installing an update now cleanly and automatically stops and restarts the background service, preventing orphaned processes.
• Clean timeline marker ordering: Reasoning effort markers now appear directly beneath your message and persist reliably when reopening earlier chats.

Added

• What's New as a continuous release feed: the dialog is now a large vertical feed of release history (search removed) that auto-opens once after each app update, so a fresh build shows what changed at a glance
• The boost skill with Antigravity-style deep reasoning: prompt-master is fused into boost, adding a second mode modeled on Antigravity's /boost — a strategize → delegate → verify pipeline that drives real Neuron sub-agents for hard engineering problems, alongside the existing prompt-engineering mode for writing prompts aimed at other AI tools. boost now leads the root slash menu, ahead of /compact, /clear, /exit, and /skill
• Installed skills in the root slash menu: skills list below the main commands with top-down priority, and the neuron- prefix is stripped from their slash commands while both forms (/neuron-design and /design) still work
• Global skills load exclusively from ~/.neuron/skills: user-home skills are sourced from Neuron's own directory, isolated from other agents' conventions; workspace skills still come from .neuron/skills, .agents/skills, or .claude/skills
• Quota reset card in Account Usage settings: reset credits and the reset button now live in Settings → Account Usage alongside the in-banner reset, with the banner hiding the button cleanly when credits are unavailable

Improved

• LaTeX math symbols render as Unicode in markdown: $\rightarrow, $\leftarrow, $\le, $\approx, Greek letters, and other LaTeX tokens are normalized into clean Unicode characters across assistant messages, plan cards, height estimation, and clipboard copy — no more literal backslash-dollar soup without a math renderer
• antislop upgraded to the full upstream specification from miqdadbadjuber/anti-slop

Fixed

• Agent turns keep tool results after a failure cleanup: when an unresponded user turn is cleaned up after an error, the accumulated tool results are preserved instead of discarded
• Upstream 400s embedded in 503 gateway errors are surfaced properly: the true status passes through instead of a generic gateway error; empty turns self-heal and array schemas are sanitized on tool input

Security

• Dependency audit reduced from 103 to 71 findings (critical 6 → 1, high 42 → 22) by bumping same-major pins locked below their fixes and letting npm re-resolve transitive ranges: Electron 41.2.0 → 41.10.7, electron-builder, electron-updater, the ai SDK 5.0.78 → 5.0.255, vitest 4.1.x → 4.1.11 (browser-mode file-access RCE), esbuild, wrangler, the Cloudflare Vite plugin, concurrently (shell-quote), js-yaml, ws, and transitive ranges like axios, fast-uri, hono, lodash, and tar. The residual high/critical findings sit on major migrations (eas-cli 16 → 23, expo 54 → 57, wrangler/Cloudflare) or have no published fix and need separate scoped work

Added

• Streamlined skills catalog: Consolidated built-in skills into focused core workflows (/boost, /antislop, /release, and /design) with predictable command ordering.
• Automatic skill updates: Desktop automatically keeps built-in skills up to date in the background without requiring manual intervention.

Improved

• Smarter skill execution at higher effort levels: Specialized skills now activate automatically based on task context with optimal defaults.
• Protected core skills: Essential built-in platform skills are protected from accidental removal in the settings interface.

Fixed

• Stability and performance improvements: Resolved minor display and routing issues across the application.

Added

• Reliable What's New dialog: Release notes now reliably appear right after an update so you can immediately see what has improved in each version.

Improved

• Faster agent responses and lower token usage: Optimized internal prompt instructions and smarter skill loading cut token usage nearly in half, making the agent noticeably faster and more responsive.
• Natural language understanding for reasoning effort: The agent now accurately understands everyday conversational phrasing, casual Indonesian, and slang when deciding how deeply to think through your requests.
• Balanced skill activation at high reasoning levels: The agent now opens specialized skill instructions only when your task genuinely requires them, avoiding unnecessary prompt overhead.

Fixed

• More cost-effective Token Saver: The "Off" setting no longer loses prompt caching benefits, preserving full accuracy while keeping token costs predictable.
• Cleaner timeline sequencing: Status indicators and effort markers now appear in proper conversational order directly beneath your prompt.

Added

• Integrated prompt-master skill with natural intent routing: bundled prompt-master v1.8.0 with 13 prompt-engineering templates and 37 credit-killing antipattern rules under skills/prompt-master/
• Automatic natural activation for prompt generation: agent sessions automatically activate prompt-master on natural prompt creation, crafting, refinement, or optimization requests in English and Indonesian without requiring slash commands
• Prompt Master constraint discipline embedded in agent preamble: enforces explicit single primary objectives, strict file scope locking, and binary falsifiable acceptance criteria for high-precision task execution
• Comprehensive prompt-master research teardown: added architectural analysis and source triangulation under docs/research/prompt-master/

Fixed

• Prevented runaway next-prompt suggestion requests on idle navigation: guarded suggestion fetching so it only triggers when an active turn transitions from running to settled (settledTurns > 0), eliminating spurious LLM suggestion requests when mounting idle workspaces, switching agents, or navigating between tabs
• Graceful quota reset button display: hidden the quota reset button cleanly when reset credits are unavailable or unsupported

Improved

• Whitespace and tab tolerance in the edit tool: edit now transparently handles differences between tabs and spaces using deterministic tab-width conversions and sliding-window line matching. Files formatted with literal tabs (such as Go source or Makefiles) now match cleanly against space-indented edits from LLMs without failing
• Diagnostic diff feedback on edit mismatches: when an edit cannot find an exact match, the error response now displays the closest matching line and an inline diff, helping the model correct its search snippet in a single turn
• Tool input parameter alias normalization: common parameter name variations across built-in tools (such as filePath for file_path, content for text, and string-encoded numbers or booleans) are automatically normalized before tool execution
• Proactive delegation in Max Effort mode: turns running with maximum reasoning effort are directed to leverage specialized sub-agents (Explore, Planner, Reviewer) and full capability tools (web search, document tools, batch jobs) proactively for large, multi-faceted tasks
• Sub-agent visibility in Build and Plan prompts: create_agent tool capabilities are now explicitly documented in the build mode prompt, making models fully aware of sub-agent delegation
• Multilingual adaptive reasoning scoring: added Indonesian and international action triggers to calculateAdaptiveScore, ensuring non-English prompts are accurately classified with the appropriate reasoning tier and thinking budget

Fixed

• Fixed loop guard false positives during sub-agent and batch job polling: get_agent_status, get_agent_activity, batch_status, inspect_schedule, and schedule_logs are now exempted from repetitive tool-call loop guards so agents waiting on asynchronous tasks do not trigger false spin warnings
• Fixed active host serverId resolution in MCP settings and GitHub auth: resolved serverId from connected hosts in McpSection and useGithubAuth when local daemon server ID is unresolved or connecting to external daemons

Added

• Quota reset button and credits tracking in usage limit warning banner: when approaching or hitting account usage limits, the warning banner now provides an interactive reset button displaying remaining credits (e.g. "Reset (N left)")
• Direct Hub quota reset integration: seamlessly reset Neuron Hub quota from within the client with instant visual feedback, loading spinners, success toast notifications, and automatic refreshing of usage meters
• Hub auth credential bridge for quota management: surfaced Hub URL and API key access across desktop host and protocol stores to support secure, authenticated quota resets directly from the banner

Improved

• Redundant file read tracking: added RedundantReadTracker to detect when an agent repeatedly reads the exact same file spans within a turn, providing polite nudges without blocking execution
• Periodic progress checkpoints: turns report progress notices every 50 tool calls to keep execution transparent during extended tasks
• Decoupled token truncation handling: output truncation retries now run through dedicated max token resume tracking (MAX_TRUNCATION_RESUMES) rather than consuming turn loop budgets

Changed

• Unbounded tool-call execution for agent turns: removed the hard 160 tool-call ceiling (MAX_TOOL_ITERATIONS, extensions, and tool budget exhaustion) so long, multi-step autonomous tasks can proceed without synthetic interruptions

Fixed

• Fixed workspace empty state controls: default session mode and reasoning effort controls are now properly displayed on the new workspace empty state
• Fixed desktop auto-updater check latency: bypassed GitHub API cache headers and shortened background update check intervals to 1 minute for faster update delivery

Added

• Batch processing at half price: submit a whole set of files to OpenAI or Anthropic as one asynchronous batch job and pay roughly half the per-token price. Reach for it when the work is wide and shallow — auditing every service, writing tests across a directory, applying one rewrite to a package. The agent gets batch_submit, batch_status, and batch_results; the CLI gets neuron batch submit, ls, worker, and results. A batch is answered in minutes to hours, so you submit it and come back to it. It needs the provider's own API key (OPENAI_API_KEY or ANTHROPIC_API_KEY) in the daemon's environment; without one the job is queued locally and says so.

Improved

• A collapsed group of tool calls now says how many failed: "Ran 7 commands (1 failed)" instead of a summary that read the same whether every call worked or half of them errored. A call you canceled is not counted as a failure.
• An agent that gets a tool's name wrong now recovers in one step instead of guessing: a decorated spelling — read_ide, read_ide_ide, functions.read, readFile — is resolved to the real tool before the round is recorded, so the wrong name never reaches the transcript and teaches itself back. When no tool matches at all, the error names the closest one ("Did you mean \"read\"?") rather than only saying no. A session that spent an entire turn cycling through spellings of read is what this is for.

Changed

• The bash tools are now shell, shell_output, and shell_status: the tool has never been bash — it runs /bin/sh, and cmd.exe on Windows — and the timeline has always drawn it as "Shell". The old names still work everywhere, including in a workspace hook's matchTools, so nothing you have configured stops matching.

Added

• Four engineering and quality skills:
  • neuron-antislop: enforces purposeful design, real interactive behaviors (no dead buttons or placeholder links), mandatory tri-state views (empty/loading/error), WCAG AA contrast, and human voice copywriting without AI buzzwords
  • neuron-api: protocol design standards for REST, WebSocket, and type-safe RPCs, enforcing backward-compatibility invariants, wire schema purity, and structured error envelopes
  • neuron-db: database migration safety, expand-contract zero-downtime schema evolution, index placement discipline, connection pooling, and strict guards against destructive operations
  • neuron-git: pre-flight status checks, atomic conventional commits, branch isolation, and workspace hygiene to avoid orphaned test artifacts or dirty repo state

Improved

• Auto-derivation for memory titles: memory_remember and ncm_remember now automatically synthesize concise titles from memory content when omitted, eliminating schema validation errors when callers invoke memory tools without an explicit title
• Silent Cloud Memory fallback: if ncm_remember encounters a network or server issue, the agent session immediately and silently falls back to local memory without interrupting turns or producing spurious error notifications
• Workspace script tools bridged to agents: list_workspace_scripts and script management tools are now accessible to agent sessions outside plan mode

This release adds on-demand rate limit reset credits to the Limits tab, and makes chat audit records cleaner and more accurate.

Added

• Rate limit reset credits in the Limits tab. You can now see how many reset credits you have left and use one right from the Limits tab. When your usage is running low or you hit a rate limit, click Reset Quota to restore your allowance straight away instead of waiting for the usage window to clear.

Fixed

• Cleaner, single-entry chat audit records. When audit records are saved to Cloud Memory, your typed message is now saved once, when the reply is complete, instead of being repeated over and over during the steps of a tool run.
• Internal messages left out of audit history. Hidden system messages and retry instructions are no longer saved to Cloud Memory, so the stored history only shows your actual conversation with the assistant.

Fixed

• Fixed the context meter's hover panel showing nothing after a restart or a workspace switch: the breakdown of what filled the window, and the tokens the saver kept out of the prompt, were written to disk correctly but discarded when the agent was read back, so the panel stayed empty until some later turn happened to report them again
• Fixed the context meter disappearing entirely for the rest of a session when the model catalogue could not be reached on the first try — a single failed lookup was cached as final, leaving the ring without a capacity to measure against. This also restores the meter for models resolved per request, such as auto

Added

• A Design surface, alongside Home and Code, where a written brief becomes rendered files rather than a conversation
• The neuron-design skill behind that surface: it locks a route, clarifies at most once, binds a colour contract, and builds in one pass against contrast and layout rules that can be checked from the source
• Automatic updates now wait for an idle machine, checking your input, unsent drafts, dirty editors, running agents, pending permissions, background terminals, git work, and firing schedules before replacing the app on quit
• A suspend or lock-screen aborts an update in progress, so an install never lands on a machine going to sleep

Changed

• Every reasoning level now changes how the model works, not just how long it thinks: low answers directly, medium confirms then acts, high treats its first explanation as a hypothesis, and xhigh maps the problem before committing
• Max effort now buys depth instead of subagents: it previously required fanning out to subagents and rejected a turn that solved the problem in one context

Fixed

• Fixed a turn going silent after running a shell command, leaving "The model ended the turn without a closing message" under the tool row when the work had already been done
• Fixed two end-of-turn notices running together into one sentence
• Fixed the reasoning level being ignored above high: the Hub collapsed any thinking budget over 15k back down to high, so xhigh and max silently did the same work
• Fixed the model list and new workspaces failing with a bare "fetch failed", which hid the reason — a DNS failure, a connect timeout, and a refused connection all read identically
• Fixed Cloud Memory devices showing a generic device label instead of your git email
• Fixed an audit sync batch overflowing on long transcripts

Fixed

• Fixed new messages failing to save while a conversation was open elsewhere, which left a chat looking empty even though its full history was still on disk

Fixed

• Fixed a reopened conversation showing no messages, even though its full history was still stored and never lost

Improved

• Context meter displays upstream provider quota and capacity windows with absolute figures, window durations, and compact hover rows

Changed

• Tool-call safety ceiling raised to 160 calls with seamless auto-continue: turns in Autopilot now auto-extend their tool budget up to 160 tool calls instead of cutting off at 80 and requiring a manual "lanjut" message

Fixed

• Fixed an agent turn closing unexpectedly when interrupted by user steering
• Fixed transcript overflow causing Hub requests to fail on prompt length by automatically compacting transcripts and calibrating token estimates
• Fixed skill scanning across global user directories and symlinks, discovering skills in ~/.codex/skills, ~/.claude/skills, and user-configured paths
• Fixed Windows build tool failures across workspace scripts
• Fixed tool dispatch failing when tools carry the _ide namespace suffix

This release stops a "conversation too long" error from reaching you as a raw message when the service does not say how much room was available.

Fixed

• Over-long conversations now recover even when the error gives no numbers. Neuron already shortened and retried a conversation that a model rejected as too long, but only when the error included the limit as numbers. If the service worded the error differently, Neuron skipped the recovery and showed you the raw prompt is too long message. Neuron now recognizes that error by its wording as well as its numbers, and shortens the history based on the request it just sent, which is known to be too large. An error about the length of the reply, rather than the length of your conversation, is still handled separately and is never mistaken for this.

Added

• Chat transcripts are recorded durably to SQLite, with tools to read them back, so a session's history outlives the process that produced it
• Audit transcripts can sync to Neuron Cloud Memory, opt-in via NEURON_AUDIT_SYNC: progress is tracked durably, so a restart neither re-sends from the beginning nor drops rows, and a batch is only marked synced once the server confirms it
• document_patch gains paragraph operations for Word documents: insert_paragraph, delete_paragraph, and set_paragraph_text address paragraphs by position, counting those inside table cells
• document_patch gains sheet management for spreadsheets: add_sheet, delete_sheet, and rename_sheet, each doing the full bookkeeping a worksheet needs — the part, its content-type override, the workbook relationship, and the element whose order is the display order
• write_document exports a sheet to CSV and TSV with RFC 4180 quoting; a multi-sheet document is refused rather than silently writing only its first sheet
• Two more pure-computation tools: parse_csv renders CSV/TSV as a markdown table, handling quoted delimiters, escaped quotes, and embedded newlines; format_xml pretty-prints or minifies XML while preserving comments, CDATA, and DOCTYPE content verbatim

Changed

• Memory capture is now active rather than permissive: the agent saves what it learns on finishing a task, settling a decision, fixing a non-obvious bug, or discovering a constraint, instead of being merely permitted to
• A saved memory is written to both stores: local first, because it cannot fail on a network, then Cloud Memory immediately after, so knowledge follows you to your other machines rather than being stranded on one
• The subagent track sits directly above changed files as one frame, instead of floating as a separate card

Fixed

• Fixed max reasoning effort carrying over after you stepped it down: the delegation instruction lived in the transcript as a message, so a later high turn still read that it was running at maximum effort and kept fanning out subagents you had stopped paying for
• Fixed a profile's role mode overriding a subagent's Autopilot setting, which left spawned subagents stopping to ask for approval they had already been granted
• Fixed a model refusal being retried: a refusal is deterministic, so the retry loop only cost time and tokens before returning the same answer
• Fixed spreadsheet edits landing on the wrong worksheet after a sheet was added or removed: sheets were located by workbook order but indexed by filename order, which agreed only until the two diverged
• Fixed PowerPoint slides being read back in filename order rather than presentation order, which made reorder_slides and an inserted slide invisible when the deck was read again
• Fixed the max-effort cost notice appearing in the timeline before the model had run

This release adds background recording of chat audit transcripts to Neuron Cloud Memory, so conversations can be reviewed without slowing down chat.

Added

• Background chat audit transcripts. Finished chat turns, including your prompts, the assistant's replies, the model used, response time, usage, and any errors, are now recorded to Neuron Cloud Memory's audit log in the background. Recording never delays or blocks your chat. If the network fails briefly, the failure is ignored so your work is not interrupted.
• Turn recording on or off. You can enable or disable transcript recording at any time in VS Code settings (on by default).

This release gives Cloud Memory's code map something to answer from, and teaches Neuron about three memory tools it did not know about.

Added

• Neuron can now map your codebase. Cloud Memory can answer questions about your code's structure, such as what a file provides, what calls a function, what a change would affect, and how one part reaches another. But it never reads your files, so it could answer only from a map that nothing had sent. Run Neuron: Index Codebase from the Command Palette and Neuron reads your project on your computer and uploads its structure: the files, the names they define, which files use which, which functions call which, and any web routes it finds. Until you run it, these questions come back empty, which looks like "nothing depends on this". That is the most dangerous answer to get before a refactor.
• Indexing is manual, and stays manual. The map describes the shape of your code, so it is uploaded only when you ask for it. There is no background watcher and no automatic upload. Run the command again when the structure has changed. Each upload replaces the previous map, so files you deleted stop showing up as dependencies. Only the structure is sent, never file contents, comments, or text strings.
• Neuron knows about three more memory tools. It can now ask for everything known about one file or name at once, ask how one part of the codebase reaches another instead of tracing the steps by hand, and check which saved memories describe code that has since changed. Neuron already knew that memories can go out of date. Now it has a way to check which ones actually are.

Added

• Max reasoning effort now enforces orchestration instead of only asking for it: a max-effort turn that answers single-handedly is sent back once to delegate, or to say plainly why the task was trivial
• Token saver runs pxpipe image compression from low upward, not only at high, with the model allowlist deciding eligibility rather than the level
• pxpipe covers every model measured to read renders reliably, widening the scope beyond the library's conservative default

Changed

• Compaction reworked, with the context window calculation corrected: transcripts fold more predictably and the reported window now matches what the request actually carries
• An image refusal falls back to plain text for the same iteration, so a turn is never lost to a token-saving optimisation

Fixed

• Fixed a new agent being named by two separate model calls, which cost two round-trips and could disagree — the window and the tab beside it showed different names, in different languages
• Fixed the Setup tab staying open after setup succeeded; a failed setup keeps its tab, because that is the only place the error is shown
• Fixed grep results at high and ultra hiding matches behind a bare +25 more, which left the model no way to reach them and sent it grepping in circles — the overflow now names the line numbers it dropped

Added

• document_patch can size, clear, and freeze a spreadsheet, and add a slide to a deck: set_column_width and set_row_height size one column/row or a span, clear_range blanks a rectangle of values while keeping its styling, set_freeze_panes freezes the rows and columns above and left of a cell, and add_slide inserts a new slide with a title and bullets at any position — all in place, leaving the rest of the file untouched
• Four pure-computation tools that reach nothing on the machine: format_json pretty-prints or minifies and reports the parser's reason on invalid input, diff_text produces a real LCS unified diff so one inserted line reads as one addition rather than a cascade, hash_text digests with sha256/sha512/sha1/md5 as hex or base64, and encode_text converts base64, base64url, hex, and URL encoding in both directions with a clear error instead of a corrupted string

Added

• Chat and Code surfaces are available from the header switcher, with Design shown as disabled while it is under development
• Compacted transcripts are archived as JSONL before folding, preserving the verbatim history outside the live context

Changed

• New workspaces default to isolated Git worktrees when the selected project supports them
• Agent prompts include the current working directory, operating system, and date so each turn has up-to-date environment context
• Autopilot and Bypass modes use distinct accent colors, making their approval behavior visible in the composer
• Workspace shortstats now count only uncommitted changes

Fixed

• Fixed model changes made during an active turn applying to later requests in that same turn
• Fixed large resumed transcripts overflowing the runtime argument limit
• Fixed internal generation turns carrying the full coding-agent prompt and workspace context
• Fixed live agent title updates not reaching tabs opened from agent details
• Fixed the Code sidebar retaining the internal Home chat workspace after stale directory data
• Fixed sandboxEnabled being rejected in Neuron provider options

Added

• Document runtime and chat-legal productivity tools: native document inspection, patching, and OOXML support (DOCX, XLSX) along with safe productivity evaluation and unit conversion tools.
• pxpipe image context compression in token-saver: added high and ultra compression modes to automatically downscale and optimize image context payload without losing key visual structure.
• Dedicated document runtime architecture documentation and developer handover guide.

Fixed

• Fixed image attachment routing: images picked via 'upload file' in composer are properly routed to the image tray instead of as documents.
• Fixed folder path resolution: added folder now correctly shortens relative file paths for agent context.
• Fixed image size estimation in context meter and compaction: images are now estimated with a flat realistic token estimate rather than raw base64 string length.
• Fixed twin agent spawning: prevented duplicate agent instantiation on draft remount via client and server message deduplication.

Added

• Auto mode is a real Autopilot. Beyond accepting everything, it resolves ask_user questions itself, nudges the agent to run a verification command after it edits files without one, and tightens the tool budget — grounded in a read-only project read that never clears the verification ledger
• Image generation in the agent's toolbelt. The daemon offers generate_image in Chat and read-only roles, backed by Hub models auto-discovered from /v1/models with a 3x retry before falling back to Pollinations, and materializes remote image URLs into local attachments for reliable rendering
• Images work for text-only models. A vision translator bridges images to models without vision support
• Indonesian language support. System, English, and Indonesian are the available languages
• VS Code-style steering for the running agent. Interrupt and redirect the agent mid-turn
• Max reasoning effort turns orchestrate multiple agents, ultracode-style, and notify you about token and time cost
• ask_user Q&A shows as a persistent timeline card, and create_agent accepts a per-call model override

Changed

• OS-level shell sandboxing auto-enables for agent execution
• Agent updates always auto-download in the background
• web_search is renamed for GLM models whose upstream swallows it, and ask_user always offers free text
• Riskier modes get in-app confirmation modals instead of native alerts, with clarified Automatic and Bypass confirmations and a polished mode picker
• Top usage models shown in the usage popover, with provider usage refreshed while the agent runs so warning banners fire on time
• Saved tokens overlay the context window breakdown, aligned with the bar, and the token saver clears old tool results above the trigger at high and ultra levels

Fixed

• Fixed imported sessions losing their title on new workspaces
• Fixed a crash when composing — the client stays bound as this when sending
• Fixed a turn dropped while only reasoning had streamed — it's now retried
• Fixed leaked harmony tool-call markup from the live stream and timeline projections
• Fixed retrying an errored agent showing a 'failed to start' toast
• Fixed agent tools aimed at the agent itself — cancel/archive/kill are refused
• Fixed CJK word-spacing explosion in desktop markdown paragraphs
• Fixed image fallback UI when assistant images fail to load
• Silenced benign react-native-web console warnings in the desktop build

Added

• An edit_image tool for real image editing. Beyond generation, the agent can edit from an instruction, transform a style, inpaint a masked region, extend (outpaint) a canvas, upscale, and keep a subject consistent from a reference — reading image, mask, and reference files from the workspace and rendering the result as a local attachment. Hub credential resolution, image-model discovery, retry, and materialization are now shared with generate_image

Changed

• Lower memory and CPU footprint on low-spec devices — tuned V8 flags, Chromium cache limits, and the libuv threadpool; background Chromium overhead reduced and webContents trimmed on minimize; the event loop yields during codebase indexing; and the in-memory agent timeline is capped to prevent heap growth

Fixed

• Fixed leaked reasoning-tag markup in answers — <think> and its siblings across model dialects (DeepSeek/Qwen, Kimi, Cohere, Seed-OSS, and the harmony analysis channel) are stripped from the live stream and timeline, including a stray tag left by a truncated turn
• Fixed "Invalid mode" when creating an agent — a saved mode a provider no longer offers (such as the retired default) is dropped in favor of the provider's default instead of being rejected

Added

• Spawned subagents run hands-off by default. A neuron subagent created with create_agent now starts on Autopilot, so a permission prompt or ask_user question never bounces back to the lead and costs an extra turn; its reasoning defaults to the adaptive tier and its shell sandbox to off. The lead can still grant privileges per subagent through settings — an explicit mode, thinking level, or sandboxEnabled: true
• Steer a running subagent, safely. When a subagent another agent dispatched is running, its composer locks to steer-only: the text box stays live and a submit interjects the active turn, while the stop button, mode selector, and model selector go dead so a bystander cannot interrupt or reconfigure a subagent its dispatcher owns
• The built-in daemon heals itself after an app update. When the desktop app updates but the long-lived daemon is still running the old bundled binary, it auto-restarts once to pick up the new one — no manual restart, no stale shell behavior

Changed

• The shell sandbox is opt-in. It now defaults off per agent — the composer toggle sits to the right of the model picker with a tooltip that reflects the action (Enable/Disable), and the NEURON_AGENT_SANDBOX=0 kill switch is unchanged
• A tighter, clearer system prompt. The agent preamble was reformatted into labelled sections and its wording condensed without dropping any rule, with sharper anti-spin and "know when to stop" guidance

Fixed

• Agents that spin in circles now stop. A loop guard fingerprints each iteration's tool calls and catches a straight repeat, a ping-pong between two calls, or a longer cycle — nudging the model back on track, then ending the turn if it keeps repeating, instead of grinding to the iteration cap and burning tokens
• Home screen model lists refresh with a client Hub credential, relayed through the provider snapshot so signed-in models appear without probing every provider on every mount

Added

• Permission-tier modes modelled on Claude Code. Choose how autonomously the agent acts: Plan (read-only exploration and planning), Build (the default mode that prompts for tool and execution approvals), and Bypass permissions / Auto (runs standard tools and operations autonomously). Saying "plan first" reliably drops every auto-approving mode back to Plan
• Office and OpenDocument attachments are read now. Attach a .xlsx, .docx, .pptx, .ods, .odt, or .odp and the daemon extracts its text — spreadsheets come across as tab-separated rows with a heading per sheet, decks with a heading per slide — so the model can answer about the contents instead of only the filename. The legacy binary formats (.doc, .xls, .ppt) are named as unsupported rather than inlined as a blob
• Turns auto-continue at the tool-call cap instead of stopping, allowing long multi-step workflows to keep progressing without needing manual "continue" prompts

Changed

• Neuron Hub quota now flows through daemon listProviderUsage. The Account usage popover and Approaching limit warning banner both read from the daemon with ephemeral request-scoped credentials, preserving shared cache isolation across accounts
• Retired the legacy Automatically mode in favor of Build as the default, aligning session behavior with explicit permission tiers

Fixed

• Fixed imported sessions losing their title. Adopted conversations from Claude Code keep their custom titles, AI titles, or prompt previews on workspace tabs instead of falling back to workspace branch names like "main"
• Fixed quota limit warnings on Neuron Hub account windows. The banner now surfaces approaching or exhausted limits accurately based on the higher of request and token usage
• Fixed truncated bash outputs losing data. Large tool outputs now spill cleanly to temporary files so the agent can inspect full outputs without blowing context boundaries
• Fixed harmony tool-call dialects leaking into model transcripts, preventing unexpected tool-call formatting errors
• Fixed a conversation dying with HTTP 400 — unexpected tool_use_id found in tool_result blocks. An orphaned tool_result left at the front of the transcript when its tool_use is trimmed away is dropped before dispatching the request

Added

• The token saver trims the prompt before it is sent, as five levels in the model popup: prompt-cache breakpoints and duplicate tool definitions at low, recognised tool output trimmed from medium, and shorter answers asked for at high and above
• Turns now report what the saver kept out of the prompt, on the context meter while the turn runs and on the turn footer once it ends — a ~ figure, because the saver measures what it did not send rather than what a gateway will bill
• create_agent takes a profile, so a delegated agent runs on the provider, model, mode, thinking level and features configured in Settings instead of a guess

Changed

• The saver defaults to medium rather than the lossless low. low alters nothing a model reads, which also means it saves nothing on the turns that cost the most. medium only touches output whose shape is recognised and only once it is large — an MCP tool's JSON, or anything the detector cannot place, is passed through untouched
• The saving control starts collapsed in the model popup, with the current level on its header, so the model list is not pushed down by a control most people set once
• Neuron is the only provider. The manifest already offered nothing else; this removes the unreachable code behind it, and an install still naming a custom ACP agent is warned about and skipped rather than failing to boot
• Compaction now fires later and keeps less, so a compacted conversation actually feels lighter instead of carrying a third of the window forward

Fixed

• Fixed the saver never compressing what read returns — the largest tool result in an editing turn. Its numbered body matched a rule inside the git-status detector, which handed the text back unchanged, so it went through whole at every level
• Fixed web search and fetch results going through uncompressed, which no filter recognised; a turn spent on the web saved nothing at any level
• Fixed a turn ending on a message that only described the next step, which left the work unstarted until you typed "continue"
• Fixed a tool call running with no arguments when the gateway opened the block and never streamed the body; required arguments are now checked before the call is dispatched
• Fixed a batch of tool calls reporting a missing argument when one gateway delivered the whole batch's arguments in a single body — the failure now names where the arguments went and asks for one tool at a time, instead of sending the model round the same batch to fail identically
• Fixed hitting the tool-call ceiling ending a turn silently, and stopped bookkeeping tool rows from burying the actual work in the transcript
• Fixed a release becoming available before every platform had built it. Each platform uploads to the same release, which was created already published — so the first one to finish offered that version on its own, to users whose platform had no build yet

Added

• Import a session from another coding agent. Adopt a Claude Code conversation into a Neuron agent and carry it on — the transcript seeds the agent so the next turn continues where the other tool left off, on your Hub account. Open it from the sidebar's Import session or the home screen. Old tool calls come across as history the model reads, not replays
• Every tool argument now shows an example in its description, so a model has a concrete value to copy — `**/*.ts` for a glob, 120 for a line offset — rather than a bare type. The same for the agent, schedule, terminal, and browser tools

Changed

• The sidebar puts Add project in its header — each project already lists New workspace inside it, so the global row was the odd one out — and shows which account you are signed into, with an avatar generated from the name, in the footer
• Local memory is the primary store. Its tools are offered before Cloud Memory's, and the saved-memories list leads with it. A Cloud Memory that cannot be reached is no longer drawn as a failed call: the fact is already saved locally, so the row says what happened without reading as data loss
• create_agent takes its title from the start of the prompt when one is not given, instead of refusing the call — a model that never sent a title could not create an agent at all
• dev:desktop rebuilds the server before it launches, so a server change is never left running against a stale build; dev:desktop:fast skips it for app-only work

Fixed

• Fixed a batch of tool calls losing their arguments. When a gateway delivered a whole batch's arguments as one body, one call got them and the rest failed empty. The body is now split back to each call, and a gateway that keeps doing it is asked for one tool call per message — recovering the round instead of sending the model back to fail the same way
• Fixed MCP tool errors arriving as a wall of validation JSON. A rejected call now reads as one sentence naming the argument and what to send
• Fixed web_fetch pasting an error page into the transcript on a 404 or 500; it reports the status and stops instead of carrying two hundred lines of a not-found page
• Fixed read answering with a raw ENOENT — it now names the missing path and points at glob, and it opens files in a workspace reached through a symlink, which it had been refusing
• Fixed the token saving showing only after a turn ended. It now shows on the footer while the turn runs, and the count and saving land on the reply as it streams rather than only after a restart
• Fixed the "Failed — retrying" footer freezing for the rest of a turn; it clears the moment the retried attempt resumes
• Fixed a collapsed tool group not showing whether it is still working — it shimmers on its own summary while a tool inside it runs
• Fixed a schedule duration in days (2d) being rejected, which the tool's own description had offered as an example

Added

• Add folder puts another directory in the agent's working context, so files outside the workspace can be read without copying them into the prompt; added folders remain visible in the composer and can be removed individually
• The model selector now includes the selected model's reasoning effort, with a provider-driven slider that only shows levels supported by the active model

Changed

• Adaptive reasoning continues to choose effort locally from the prompt, while the selected concrete tier is now kept aligned with the model and provider request that actually runs
• Changing the model, reasoning effort, or additional folders during an active turn applies on the next turn, rather than interrupting the work already in progress

Fixed

• Fixed higher reasoning tiers being shifted upward on budget-based providers, where high could arrive as xhigh and xhigh as max; tier budgets now match the gateway's bucket boundaries
• Fixed reasoning requests ending with an incomplete reply, by keeping a generous answer allowance independent of the selected thinking tier and recovering from provider output ceilings
• Fixed Gemini reasoning configuration using the wrong control for the model family, so Gemini 2.5 uses thinkingBudget, Gemini 3 uses thinkingLevel, and unsupported none/top-tier values are handled safely

Fixed

• Fixed the app failing to load against a host with enough stored agent history, where the daemon closed the connection while sending the agent list and the app reconnected into the same failure

Fixed

• Fixed Neuron thinking blocks running together when the Hub stream omits block-start boundaries
• Fixed long-running turns failing after a Neuron Hub connection reset with retryable diagnostics instead of a bare TypeError
• Fixed cancelling a retry from sending another Neuron Hub request

Added

• bash_status lists the background commands a session started, with whether each is still running and its exit code, so a lost id no longer means a process you cannot reach

Changed

• Memory saves to this machine by default, and Cloud Memory syncs it to your other machines — a cloud outage no longer means the fact is saved nowhere
• A Cloud Memory failure is no longer reported mid-conversation, since the fact is already stored locally; Settings still shows the real error
• Removed task, whose hidden subagent duplicated create_agent while running work you could not see, cancel, or inspect
• Removed the Cloud Memory graph and wiki tools, which answered questions codebase_search and grep already answer against your working tree
• Renamed the memory tools to memory_* and ncm_*, and run_workflow to agent_run_workflow
• The usage limit banner can no longer be dismissed and appears at 85% instead of 75%, so the warning stays until usage drops back under the threshold
• The usage limit banner spells out what is left and when it resets, choosing its unit from the distance to the reset
• The error card leaves the screen when you press Try again, instead of announcing a failure above the turn that is already retrying

Fixed

• Fixed pasting a large image failing the whole message, such as a range of cells copied from Excel — Neuron now names the size and the limit instead of surfacing the gateway's error
• Fixed a conversation failing on every later turn once an oversized image had entered it
• Fixed consecutive thinking blocks running together into one caption on models that emit a summary per step
• Fixed a thinking block staying marked as in progress after a later one had started
• Fixed a git repository opened directly keeping the raw first line of its prompt as its workspace name

This release keeps long chats responsive: counting the size of a conversation now always happens on your computer, and it is faster on dense content.

Removed

• Counting a conversation's size never uses the network. VS Code asks how large a conversation is many times while you type and while a request is prepared. Neuron could send each of those counts away to be measured, which sometimes made a chat wait before anything was sent. The size is now always measured on your computer.

Fixed

• A large result returned by a tool no longer freezes the size estimate. Long text is now measured in a single pass instead of repeatedly. A very long result used to take tens of seconds to measure.
• Long names are no longer mistaken for encoded data. Only long unbroken runs of code-like characters containing digits are treated as dense. Ordinary names and prose are counted as before.

This release stops a long conversation from being cut off for being too large, and recovers by itself when it happens anyway.

Removed

• Counting a conversation's size now happens on your computer only. VS Code asks how large a conversation is many times while you type, and Neuron could send each of those counts away to be measured. That could leave a chat apparently stuck before anything was sent. The same problem was fixed once before, and the remote counting has now been removed for good so it cannot return. It was no more accurate anyway: both sides used the same estimate.

Fixed

• Conversations with images no longer go over the size limit. An image travels as text and takes up far more room per character than ordinary writing, and Neuron was counting it as if it were prose. It could be out by up to a quarter, so a conversation looked like it fit, VS Code never shortened it, and the model still refused it. Images and other dense data are now counted at their real size, while ordinary text and code are counted as before.
• A conversation that is too large is now shortened and sent again instead of failing. When the model refuses because the conversation is too long, Neuron reads how far over it is, shortens the older part to fit, and tries again. Before, you just saw the refusal.

This release fixes Cloud Memory returning a "not authorized" error when you sign in to the Neuron desktop app or use several devices with the same account.

Fixed

• Signing in to the desktop app no longer signs VS Code out of Cloud Memory. Both apps used the same account name, so signing in on the desktop replaced the one key the server kept, which immediately made VS Code get "not authorized" on its next memory request. Each device now registers with its own name, so each editor and desktop install has its own key while all your devices still share the same synced memory.
• Several computers on the same account no longer kick each other out. Two laptops running VS Code or the desktop app each keep their own key, so using Neuron on one machine never interrupts another.

Every current Claude model reasoning correctly for the first time, an empty composer that knows what to ask next, and delegated work that runs in the background instead of holding the turn open.

Added

• The empty composer suggests what to ask next. Once a turn settles, the agent that just finished has usually left an obvious next move — run the tests it wrote, look at the file it named, undo the part that did not work — and typing that out is the tax on every turn after the first. It now reads the conversation and offers one line where the placeholder was. Tab writes it into the box; typing anything at all ignores it. Written in rather than sent, because a suggestion is a starting point, and it is only ever offered in an empty composer between turns. The old placeholder is erased and the suggestion written in its place a character at a time, because a swap between two frames is easy to miss entirely — the composer is not what you were looking at when the turn ended.
• Delegate and carry on. task takes run_in_background, so a survey that takes minutes returns an id immediately and the agent gets on with its own work instead of sitting idle for the whole of it. task_output reads the report when it is wanted, and when one lands the agent is told once — alongside whatever it is already being sent — so a delegated search is neither waited on nor forgotten. The waiting form is unchanged and still the right default when the answer is the next thing needed.
• Independent tool calls run at the same time. Ask for three reads, two greps, and a delegated survey in one message and they all start together rather than queueing behind each other. Anything that edits a file or runs a command still goes one at a time, in the order it was asked for, so a write cannot race another write and permission prompts still arrive one by one.

Fixed

• A quick thought showed no time at all. One that finished fast enough to arrive in a single update was treated as one that had never been timed, so it read as a bare "Thought" next to slower ones that reported their seconds. It says "a second" now; only a thought nothing watched stays untimed.
• The running-turn line never named the reasoning level. It was meant to read "Still thinking with high effort"; it always fell back to the plain wording. The provider resolved the level and announced it, the wire carried it, the app was ready to show it — and the daemon threw it away in between, because the accepted-turn boundary publishes a start of its own and dropped the provider's as a duplicate. That duplicate was the only one that knew.
• A long wait sat on one sentence. The dots kept moving and the words never did, which reads as a hang whatever the dots do. Waiting, thinking, and the long wait each work through several phrasings as they run — naming the reasoning level when the turn reported one — so how far the line has got is itself a hint at how long it has been going.
• Recents kept sliding around. The chat list borrowed the agent list's ordering, which lifts anything running to the top: a chat jumped up the moment it started answering and dropped back when it stopped, moving the row you were aiming at. Recents is ordered by recency alone now, and a running chat says so with its dot instead of by moving.
• The reasoning picker no longer offers levels the model cannot take. Some gateways refuse a reasoning level for a model their own catalogue advertises it for, and the daemon only finds out by asking. Until now nothing said so: the picker went on offering five levels, every one of them did nothing, and the turn that discovered it looked exactly like a turn that had run at the level you chose. The turn now says so once, and the control stays on screen reading "Not supported" with the reason behind it — rather than disappearing, which explains the missing level no better than leaving it live did.
• Every model is now asked to think in the shape it accepts. There are two, they are mutually exclusive, and each is a 400 on the models that take the other: a fixed token budget on Claude 4.5 and earlier, and adaptive thinking with an effort level on 4.6 and later. The daemon sent the budget form to everything — a hard refusal on every current Opus and Sonnet, survivable only for as long as the gateway rewrote it. Current models get adaptive thinking and the effort level you picked, older ones keep their budget, Opus 4.5 gets both because it wants both, and a level a model does not have steps down to the largest it does rather than falling back to the default. A model that refuses one shape and names the other is believed and remembered, so the correction costs one request rather than one per turn.
• One refusal could switch reasoning off for a whole session. A model that answered 400 to a turn carrying a reasoning budget was written off as one that cannot reason, and every later turn went out with no budget at all — while the picker kept offering every level and each of them did nothing. A 400 has more than one cause: a ceiling too large for the model says the same thing, and now says it to the part that lowers ceilings instead. Before writing a model off, the same budget is tried once under a ceiling small enough that the ceiling cannot be the objection. A ceiling a model does refuse is remembered, so later turns open where the last one landed rather than paying for the same refusal again.
• Picking Max did the same thing as Ultra. The reasoning levels were sized to fit under an output ceiling pinned to the weakest model the Hub fronts, which left the top two a quarter apart — and since the Hub turns a thinking budget into a reasoning effort, a gap that small stopped being a difference at all. The ceiling is sized to what the larger models actually take now, so Ultra and Max are a real distance apart and Max is more than twice what it was. A model that will not accept the larger ceiling says so and the turn steps down to the next published limit, which no longer costs it the turn.
• A turn that thought itself out of room stopped, and only moved again when you told it to. Spending the whole allowance reasoning leaves no answer to resume from, so continuing was refused outright — re-sending the identical request would have looped. Typing "continue" was doing by hand the one thing that made the retry different. It is asked again automatically now, without a thinking budget, so the model has the full ceiling to answer in. Once per turn: a second empty reply ends it rather than looping.
• Cleaning up after a command that never started killed the daemon. When a spawn failed — a hook with a typo, a working directory that had been deleted, too many open files — the tidy-up signalled a process that did not exist, and the way the runtime reports "no process" turned that into a signal to every process in the daemon's own group. The daemon and every agent under it went down together, with no error anywhere. Nothing is signalled now for a command that never ran.
• A subagent that needed permission hung forever. Its request had nobody to answer it — the child session's events reach no one but its parent — so the agent that delegated sat waiting until the user gave up. Subagents are now told plainly that they cannot ask, which turns a hang into a report. This is the "still thinking" that never ended.
• Stop now stops, and could previously leave the chat unusable. Stopping a turn aborted the agent's own work and nothing else: the subagents it had started carried on spending tokens on answers nobody wanted, and a turn that was waiting on one never learned it had been canceled. The session stayed marked busy, so every message after it came back "already has an active turn" — a chat with no way out but restarting the daemon. Stop now ends the delegated work too, and the session lets go of the turn whether or not the thing it was waiting on can hear the cancel. Whatever a subagent had already reported stays readable.
• The subagent limit did nothing. The cap of four concurrent subagents could never be reached, because tool calls ran one at a time and a second delegation never started until the first had finished. It is enforced now that they genuinely overlap, and says how to free a slot when it refuses.

Added

• Long-running delegation now stays out of the way. task is aimed at surveys, builds, and tests that take minutes: run them in the background, keep working, and receive the report when it lands instead of spending a turn waiting. A turn no longer ends while delegated work has an unread report; it waits for the result or names the work still running.
• The composer stays alive while a turn runs. Its placeholder rotates through ten localized prompts such as “Still working — you can queue up what's next” and “Type your next message — it'll send when this finishes”, so a long wait reads as progress rather than a frozen input.

Changed

• Delegation now distinguishes task from create_agent. task is the throwaway, report-back path for long work; create_agent is the visible agent path for real sidebar subagents. Agents opened by a parent are archived with that parent, and permission requests can be answered by the agent that opened them.
• Desktop update notifications are actionable without interrupting work. A published release appears before its download finishes. Automatic updates show a disabled Downloading Update button, then Install & Restart when the download is ready. With automatic updates off, one Download & Install Update click downloads the release and restarts immediately. The release check polls every five minutes, and the callout has one action instead of splitting the update flow with a second changelog button.
• Workspace agent tabs receive generated titles. The generated name now reaches the tab, switcher, and browser title instead of leaving the first line of the opening prompt as a permanent provisional label.
• task is now for the long jobs you can walk away from. It is described, and prompted, as the way to hand off something that takes minutes — a survey across many files, a long build or test run — and carry straight on while it runs, rather than as a general way to ask a question. Waiting for the answer is still available for when the very next step needs it, but it is no longer the default reading. A turn that would otherwise finish while delegated work is still running now waits for it. Previously the report had nowhere to arrive: anything unfinished was cancelled when the session closed, and a result that landed after the last tool round was never delivered — so the minutes spent on it bought nothing. Work still running after ten minutes is named rather than silently dropped, and stays readable with task_output.
• Clearer guidance on which kind of subagent to use. task and create_agent are both delegation and were described as one thing, so asking for "subagents" could produce either. The distinction — a throwaway that reports back, against a real agent in the sidebar — is now stated, along with the rule that anything opened must be closed.
• The line under a running turn keeps moving. It now has eight phrasings per phase instead of four, changes every six seconds instead of twelve, and cycles back to the start rather than settling on the last one. Previously a long turn used up its wording and then sat on a single frozen sentence for the rest of its life — the exact stalled look the rotation exists to avoid, on precisely the turns that most needed to look alive. All nine languages.

Fixed

• Parallel tool calls no longer collapse when a gateway omits block indexes. Concurrent calls used to share one pending slot, concatenate their JSON, lose most calls, and produce Unexpected non-whitespace character after JSON. Unindexed blocks now receive independent synthetic slots.
• Malformed tool arguments no longer burn extra turns when the intent is clear. Concatenated objects, duplicated objects, trailing prose, and a single-object array wrapper are recovered at the gateway boundary. Truncated or ambiguous bodies still fail, but now explain the shape and the exact retry needed.
• Required tool argument failures are actionable. Missing, empty, and wrong- type values now tell the model what to send instead of repeating a schema message that encouraged identical retries. This covers workspace search, fetch, and ask_user inputs.
• Window-controls overlay refusal no longer breaks theme updates. Unsupported overlay operations are caught and logged once per window instead of failing every subsequent theme synchronization.
• Shell, hook, and background-process cleanup now stops the right process group. Compound commands and late-starting children no longer survive a cancelled turn or leave the daemon waiting on a pipe that should be closed.
• The update notification is one button, and it appears as soon as a release exists. It used to carry both "What's new" and the install action, which split attention and made updating look like the secondary choice. There is now a single Download & Install Update; the changelog stays in the help menu, where it was already. With automatic updates switched off nothing is fetched in the background — the release is announced the moment it is found, and the press is what downloads it. The wording follows: a release that has not been fetched now reads "is available to download" rather than claiming to be ready to install. The release poll also went from every thirty minutes to every five, so a new release is noticed while it is still news. When automatic updates are enabled, that same notification remains visible while the background download runs, but its one button is disabled and says Downloading Update. It becomes actionable after the bytes land, changing to Install & Restart. The app does not restart when the download finishes; the installed update waits for the normal quit path, so an active agent or terminal session is not interrupted. With automatic updates disabled, the button stays active and starts the download itself.
• Asking for several things at once could lose most of them. When a reply requested five tools together, the calls arrived overlapping and — from gateways that do not number their content blocks — were all filed under the same slot. Their arguments ran together into one malformed body, so four calls vanished and the fifth failed with "Tool input was not valid JSON". A fan-out of subagents was the shape that hit it hardest, since that is exactly a batch of simultaneous calls.
• Subagents were left open after the work was done. An agent that opened others was expected to close each one itself, so a run that forgot — or failed partway, or was interrupted — left them in the sidebar with their sessions still alive, and the user closing work they never started. Closing an agent now closes the ones it opened.
• The agent was told to use a tool it did not have. When a subagent stopped to ask permission, the notification said to answer with respond_to_permission — a tool that was withheld, so following those instructions produced Unknown tool and the subagent stayed blocked with nothing able to release it.
• A failed delegation showed a row titled "Task task". The title fell back to the literal word, so the calls that went wrong were the ones explaining themselves least. A row now takes its name from the work it was given, or says plainly what was missing when there is nothing to go on.
• Window controls no longer break the theme sync on Windows and Linux. Recolouring the minimise/maximise/close buttons throws outright on any window where Windows Controls Overlay is not active, and nothing was catching it — so on such a window every theme change failed, for as long as the app was open. The colours are now attempted and the refusal noted once per window; the buttons fall back to system colours instead of taking the rest of the update down with them.
• Changing the reasoning level no longer alters the turn already running. Moving the selector mid-reply took effect from that turn's next request onwards, so a turn could reason at one depth and finish at another while the line underneath still announced the level it began with. The change now applies from the next message, which is what the selector always appeared to promise.
• The model list no longer offers models you cannot talk to. Embedding and reranking models — BGE-M3, BGE-Reranker-v2-m3 — were listed alongside the chat models, but neither has a chat endpoint: one returns a vector, the other returns scores, and picking either produced an error instead of a reply. Image and video generation models are held back for the same reason. The list goes by the kind the Hub reports for each model rather than by guessing from its name, so a chat model that happens to be called something like "embedding-tuned" still appears.
• Adaptive graded every request as if you had sent an empty message. The effort shown under a turn was worked out from what you actually wrote, but the effort sent with the request was worked out from an empty string — so the footer said "high" while the model was asked for something lower. On one message from a real session: medium on screen, low on the wire. It also explains the stopping. The room a turn gets to answer in is sized from its reasoning level, so a demanding request graded as if it were blank was given less room than it needed, ran out mid-answer, and stopped — which is why it kept taking a "continue" to finish a thought. Both numbers now come from the same place.
• Commands with two parts could hang forever, and their timeout could not save them. cd build && npm test is two commands, so the shell forked instead of becoming the command — and stopping it killed the shell while the work carried on in a process nothing was holding. That process kept the output pipe open, so the agent went on waiting for a command that had already been told to stop: no output, no error, no end. The 120-second limit ran on time and killed the wrong thing. Commands are now stopped as a group, along with everything they started. The same flaw ran through two other places. A dev server started in the background could outlive the session that started it — the one thing that feature promises never to do — leaving a process the user could no longer see or stop. And a preToolUse hook, which runs before _every_ tool call, could hang the agent outright rather than just failing its own command. A late-starting process could still slip out of the group in the moment between being created and joining it, so the group is swept twice more after a stop. Measured: sleep 2 | cat survived 5 times in 8 before, 0 in 10 after.

An agent that stopped in the middle of a job and waited to be told to carry on.

Fixed

• The agent stopped mid-task and only moved again when told to continue. A turn ended whenever the model asked for no tools — which is what a finished turn looks like, and also what a turn cut off by the output limit looks like. The model would spend its allowance writing the sentence introducing its next tool call, get cut off before the call itself, and the turn ended there: half a sentence, no error, nothing on screen to say why. Typing "continue" worked because the next turn starts with a fresh allowance. The loop now reads the stop reason before deciding a turn is over, and resumes from the unfinished sentence. Only when the cutoff left text behind — a truncation that produced none appends nothing to the transcript, so resuming would re-send an identical request until the iteration cap.
• A tool call the stream cut off was dropped silently. A tool_use block only counts once its closing frame arrives; one interrupted before that was discarded, leaving text and an empty tool list — the same shape as a finished turn, and so the same silent stop. The call is now reported as truncated, which the model can act on. It could do nothing with one that vanished.
• The reply allowance shrank as reasoning grew. The output ceiling was the thinking budget plus a fixed 8,192, so the room left to actually answer was the same whether the model was asked to think for 5,000 tokens or 100,000. A turn that reasoned hard and then had to describe a multi-file change ran out mid-sentence. The ceiling no longer derives from the budget: every turn gets the same 32,000, four times the old allowance.
• A rejected output ceiling killed the turn. The Hub reports each model's context limit but not its output limit, so the ceiling cannot be known in advance — and a ceiling a model will not accept comes back as a malformed request, which is not retried by design. It is now lowered and re-sent, down to the allowance every turn used to get, without spending a retry.

Updates are visible as soon as published, /compact updates the context meter, and ask_user accepts stringified option arrays.

Fixed

• A question with options was refused as having none. Asking the user to choose between several approaches failed with "questions must be a non-empty array" while the fully-formed list sat on screen. Models and the gateways in front of them both encode a nested list as text often enough that it has to be expected; the text is now decoded before the tool sees it. Only lists and objects are — a path that reads as a number stays the text it was sent as.
• A published release could stay invisible. Three separate things hid one: the gradual rollout was applied to _checking_ rather than to installing, so a machine whose turn had not come was told no update existed at all; the banner only appeared once the download had finished, so nothing announced a release during the transfer; and Settings and the banner kept separate state, so checking in one told the other nothing. A check now reports whatever the release channel actually holds. The rollout still staggers the install that happens by itself on quit — pressing Install never waits.
• An update found by hand vanished seconds later. Checking for updates in Settings would report a new version and then, a few seconds on, report there was none. Checking by hand deliberately skips the gradual rollout — you asked, so the wait does not apply — but the background poll that follows was asking again as though it were automatic, being told this machine's turn had not come yet, and clearing what you had just been shown.
• The context meter ignored a compaction. Folding a conversation — by /compact or automatically — summarised it correctly and then left the ring showing the size it had been before, until some later turn happened to refresh it. The conversation had genuinely shrunk; only the number was stale. The new size is now published as soon as the fold completes, and the marker in the thread says what it dropped from and to.
• A browser tab opened that could never be used. On a runtime with no desktop browser, opening a tab reported success and left one behind titled example.com that only said the browser is desktop-only. Every later action on it failed. It now says so instead of opening the tab.

One-agent delegation, scoped search tools with context and case-insensitive match, and Cloud Memory wiki knowledge tools.

Added

• Scoped search tools. glob and grep now accept a path parameter to narrow the search directory directly rather than filtering afterwards. grep adds ignore_case (without requiring invalid (?i) regex flags), context lines surrounding matches (using - vs : separators), and files_with_matches. Both tools clearly report when results are capped.
• Tail file reading. read supports negative offset to count back from the end of a file (e.g. -50 for the last 50 lines), and explicitly indicates remaining lines when a read is truncated.
• Cloud Memory Wiki & Skill tools. Integrated wiki_search, wiki_read, wiki_write (permission-gated), and skill_list into the Neuron provider.

Fixed

• Asking for a subagent ran a five-stage workflow instead. Handing one self-contained job to a helper — research, an investigation, a second opinion — had no way to happen: the only delegation tool available was the full Team Lead / Explore / Planner / Builder / Reviewer pipeline. An agent can now create a single subagent, task it, and close it. The pipeline is still there for multi-stage implementation work, which is what it was for.
• Workflow subagents were never closed. A finished workflow left all five stage agents running, and one that stopped early — a failed stage, or a stage that asked for permission — stranded every agent it had already started. They are now closed however the workflow ends.
• A workflow's findings never reached the answer. The stages each wrote a summary, but the calling agent received them as a raw data payload rather than as text, so its reply read as though the subagents had reported nothing. It now gets their reports as prose, in the order they happened.

Resilient timeline parsing for older histories, fixed screenshot double-paste on first attempt, accurate live version display in Settings, and adjusted macOS dock icon sizing.

Fixed

• One unreadable entry made a whole conversation unopenable. An agent whose history contained a single entry this version did not recognise refused to open at all, reporting a validation failure instead of the transcript. Entries that cannot be read are now skipped, so the conversation opens with a gap rather than not opening.
• Pasting a screenshot attached it twice the first time. Every paste after the first behaved. The two copies the browser hands over for one image were being told apart by a timestamp it stamps on each of them as it goes, which differed by a millisecond on the first paste and matched afterwards.
• The app icon looked smaller than its neighbours. It filled slightly less of its canvas than the apps beside it in the Dock; every size has been regenerated to match.
• The version in Settings did not change after an update. Installing an update replaces the application, but the screen kept reporting the version it had been installed as — the number was read from the interface bundle, which an update leaves alone, rather than from the application actually running.

Adaptive reasoning that no longer stalls mid-answer, conversations that survive an interrupted tool call, a running-turn effort label that means what it says, updates you hear about before they download, and a redesigned error card.

Added

• A clearer agent error card. A failure now shows a title, a plain-language subtitle, the model that produced it, and a retry, with the details a tap away.

Changed

• "Still thinking" now names the level the turn is actually running at. The effort in the label followed the selector — so it read "adaptive", and changed under you if you opened the selector mid-turn. It is now the concrete level the running turn resolved to, fixed for the life of the turn.
• You hear about a desktop update before it downloads. A found release used to pull ~170MB in the background before anything appeared on screen. The update is announced as soon as it is found now, and the transfer waits for Download & Install — unless automatic updates are on, which still prepares and installs on quit.

Fixed

• Adaptive reasoning ran out of room to answer, and stalled outright at the top levels. The reply allowance was a flat ceiling, and the thinking budget was spent inside it — so a hard turn that reasoned a lot had almost nothing left to write with, and at the two highest levels the budget exceeded the ceiling entirely. The gateway refused that, and the daemon made it worse by shrinking the ceiling until the turn died. The ceiling now grows with the budget, so the answer keeps the same generous room at every level.
• An interrupted tool call could break a conversation for good. A turn that stopped between the model asking for a tool and the tool running left the call unanswered, and every later turn was then rejected with a "tool_use without tool_result" error. The transcript is repaired on the way out now, so a stranded call is answered rather than left to poison the thread.
• Images were re-sent to models that cannot read them. A picture stayed in the transcript after the conversation switched to a text-only model, and the gateway refused the whole request for it.

Agents that can run things in the background, delegate a piece of work to a subagent, and be shaped by workspace hooks; a sign-in loop fixed at its source; and a codebase index that finally says what it's doing.

Added

• Commands that keep running. bash takes run_in_background, so a dev server, a watcher, or a long migration starts and hands the turn straight back instead of holding it open until the timeout kills it. bash_output reads what one has written since you last looked, says whether it is still running, and stops it on request. When one finishes, the agent is told once — alongside whatever it is already being sent — so a build that completes while it works on something else is neither blocking nor forgotten. Everything a session started is stopped when that session closes.
• Delegate a piece of work with task. A subagent runs inside your turn and hands back its report, without its reading landing in your conversation. That is the difference from creating a second agent: this one answers a question rather than opening a tab you have to manage. The tools it runs appear under the row as it works and stay on it once it finishes, as a record — not just a description that said "Delegate: …" for as long as the work took.
• Workspace hooks. A hooks block in neuron.json runs shell commands around the agent's tool calls. preToolUse decides whether a call happens at all — a non-zero exit blocks it and the hook's own output becomes the reason the model is given — and postToolUse runs after and reports, which is where a formatter or a type check belongs. "Format after every edit" written into a prompt is advice the model may drop; written here it is enforced.

Improved

• Status labels retype themselves when the words change. Thinking becoming responding, or a thought closing into how long it took, used to swap in place on a line nobody is watching. The words are typed in from the point they stop matching the ones before them, so a count going from three to four retypes the count rather than the sentence. The trailing dots on every in-progress label now count up and start over in step with each other, and hold still for anyone who has asked their system for reduced motion.
• A long thought no longer pushes the answer off the screen. An open reasoning block grew without limit while it streamed, so the reply kept sliding down and the reader chased it. The block now stops at a fixed height and scrolls inside itself, staying on the newest line while the thought is still being written — and staying put the moment you scroll up to read back through it, until you return to the bottom.
• A finished thought is marked with a brain, and a fast one no longer reads "0s". The collapsed reasoning block led with a disclosure arrow whether or not it was still thinking; it now leads with a brain once the thought is done. A thought that took under a second used to round down to "Thought for 0s", which read like a broken clock — it now rounds up and reads "Thought for a second".
• The turn footer counts tokens, not "output". A running turn read "1.2k output" and a finished one "Worked for 19s · 1.2k output". Both now read "1.2k tokens", and hovering a finished turn keeps the count beside the end time instead of trading it away for one.
• The model list says what each model can do. Every row used to read "Neuron Hub · anthropic" under the name — the provider that serves all of them, and the vendor already shown by the icon beside it. Rows now carry capability icons instead: whether the model reads images, and whether it offers reasoning levels.
• The XHigh reasoning level is now called Ultra. A display change only — agents and saved settings keep working, and a level chosen before the rename still reads correctly.

Fixed

• Signing in on Windows and Linux did nothing at all. Approving "Login with Neuron" in the browser returned you to an app that never responded — no error, no progress, just the sign-in screen. Windows and Linux hand the app its sign-in result by relaunching it with the result attached, and the app mistook that for a command typed at a terminal and quit on the spot, so the window still waiting for the result never heard back. It now recognises a sign-in result as a sign-in result, including when the app was not already running.
• Signing in bounced straight back to the sign-in screen. Approving "Login with Neuron" in the browser returned you to the app, which immediately asked you to sign in again, over and over. Neuron was still calling your account at its old address, which now forwards to the new one — and a sign-in key cannot survive that forward, so every request arrived as though nobody had signed in. The app took the refusal at face value and sent you back to the sign-in card. Neuron now uses the current address, moves a device that was signed in at the old one across on its next start, and no longer signs you out over a forward. The same forward was emptying the model list and refusing agent runs.
• A refusal during sign-in is no longer read as being signed out. Your account issues one key shared by every device, and the Hub switched that key off for the moment between your approval and the app collecting it — so the app's very next request was refused. The key now stays on throughout.
• Signing in on one device signed you out on another. Because that one key is shared, approving a login on your laptop knocked out your phone until the login finished. Nothing is switched off mid-login now, so your other devices keep working while you sign in.
• Signing out of one device signed you out of all of them. Sign-out disabled the shared key on the Hub, so leaving on your phone also ended the session on your laptop. Sign-out now only forgets the key on the device you used it on. To sign every device out at once, disable the "Neuron Companion" key from the Hub admin panel.
• Codebase search and local memory were switched off for every agent. The daemon built both and then handed neither to the agents that use them, so codebase_search was never offered as a tool, no project was ever indexed, and local memory had nothing to write to. The only visible trace was the context panel reporting that searches were falling back to grep — which they were, for every search anyone has ever run. Both now reach the agent, including agents on a profile that extends Neuron.
• The codebase index did not start for every agent. It began when a Neuron agent was built, which left out an agent reopened from an earlier session and every agent on a custom provider. Opening an agent now starts its index whatever provider it runs on.
• The codebase index said nothing about whether it was running. Not in the context panel, not in the daemon log — so an index that built successfully over zero files looked exactly like one covering the whole repository, and "is it even working?" had no answer anywhere in the product. The panel now reports a ready index and how many files it holds, and both ends of a build are written to the log.
• Read showed two columns of line numbers. The file content sent to the timeline already carried the numbers the model needs to cite lines, and the timeline drew its own gutter on top of them. The model still gets numbered lines; the timeline gets the file.
• The last message could hide behind the composer. Anything that grows the composer without touching the input — the tracked-changes bar, a queued message, the usage banner — left the stream anchored where it was, so the newest line ended up half-covered. The whole composer area is measured now, not just the input box.
• The turn's token count no longer disappears when the turn ends. The gateway reports output tokens once per message, in a final frame some models never send — so on those models the figure you watched for the whole turn was replaced by nothing at the moment it became permanent. The count is now kept, and marked with a ~ when it was sized from the reply text rather than reported. A reported count stays exact and unmarked.
• A tool badge shows whether it is still working. "Ran 1 command, read 3 files, and searched 5 times" looked the same whether the run had finished or still had a call in flight — the only sign was a shimmer across the label, which is easy to sit through without noticing. The tool icon is now replaced by the same working indicator the turn footer uses, pulsing off the same clock, and comes back the moment the work stops. A badge that failed still shows the failure first.
• A slow turn says so even when it cannot name a level. Escalating from "Thinking…" to "Still thinking…" used to require the turn to have reported an effort level, so a provider that reports none left the footer on the plain label for minutes — the question the escalation exists to answer going unanswered. The wait is what escalates now; the level is named when there is one.
• The turn footer named the effort selector, not the turn. "Still thinking with adaptive effort" was reading the setting for your _next_ turn, so it said "adaptive" where the turn had actually resolved to a level, and changing the selector mid-turn rewrote what the turn already in flight was reported to be doing. The footer now names only the level the running turn resolved to, and a turn that reported none says a plain "Thinking…" instead of borrowing one.
• The two highest reasoning levels did the same thing. The picker offered five, but the top two asked for more thinking than a reply could fit alongside, so both were cut back to the same amount — choosing Max over XHigh changed nothing. The five levels are now spaced to fit, so each one is a different setting, and every level keeps the same generous room to answer in.
• Escape no longer stops a running agent. Escape is what you press to dismiss a menu or a suggestion list, so it kept cancelling turns nobody meant to cancel — and a stopped turn looks much like a finished one. Interrupting is now Cmd+Shift+X (Ctrl+Shift+X off macOS), and it is rebindable under Settings → Keyboard shortcuts.

Cloud Memory gained a knowledge layer, a team wiki and skills your team publishes, and Neuron now knows when to use it.

Added

• Neuron reads your team wiki. Cloud Memory now holds two kinds of information: memory, which records what happened in a session, and a wiki, which holds the product rules, the terms used in your field, and runbooks that last beyond one session. When you ask why something works the way it does, or what a term in your field means, Neuron searches the wiki instead of guessing from the source code.
• Skills your team publishes are used automatically. A skill published from the Cloud Memory dashboard becomes available to Neuron in its next session, with nothing to set up. The list can change between sessions, so when a task looks like a familiar procedure, Neuron checks what is available rather than assuming. A skill returns the written procedure. It never runs anything on your behalf.
• Documenting for the team waits until you ask. Neuron saves ordinary memories on its own, but publishing a wiki page is something every teammate reads, so it happens only when you ask for it. Ask and Plan modes cannot publish at all.

Fixed

• A published skill is no longer blocked in Ask and Plan modes because of its name. Read-only modes hide anything whose name suggests it changes your project. So a skill named something like create_release was hidden even though it only returns text. Cloud Memory's read-only tools are now judged by what they actually do, not by their names.

Finding code costs less, long conversations stop dying, and the provider list is down to the one this build actually ships.

Added

• Codebase search. codebase_search finds the parts of a project relevant to a description and returns file paths with line ranges — pointers to open, not the code itself. A grep leaves every matching line in the transcript for the rest of the session; a pointer costs a path and two numbers. Ranking is BM25 over an inverted index rather than embeddings. Measured against a hashed-vector index of the same repository, the top five results for "cosine similarity ranking" were a gateway test, package-lock.json, and three unrelated skill files — the function of that name did not appear at all, and the five scores spanned 0.012. The index builds when an agent opens on a directory and follows the repository from there through the file observer, so an edit costs one re-read of that file rather than a pass over the project. git ls-files --exclude-standard decides what is in the project, which is how .gitignore is honoured without reimplementing it.
• /compact folds the conversation on request, and anything typed after it steers what the summary keeps.
• Skills in the composer. Every skill installed for a workspace is now reachable from the / menu, under /skill so the root list stays the handful of commands that act on the session.

Changed

• One provider. AGENT_PROVIDER_DEFINITIONS is Neuron Hub only. Any ACP-compatible agent still plugs in as a custom provider; the CLI clients remain in the tree but nothing can reach them. Existing agents on a dropped provider keep their records and say so when opened rather than failing at boot.
• Home greets you by name and by the hour, and the composer slides down to meet the thread's instead of cutting to it.

Fixed

• A conversation past the context window stayed dead. Every turn re-sent the whole transcript and the transcript only grew, so once it exceeded the window the gateway answered 400 — and kept answering 400 to every turn after, because nothing shrank it. The transcript is now folded at a threshold before the request rather than after the refusal. The fold never cuts between a tool call and its result, keeps the message that opened the conversation, and says what it did on the timeline.
• The context window meter reset itself mid-conversation. A cached prompt bills as one or two tokens, and reading that as occupancy emptied the meter over and over.
• Turning automatic updates off also stopped Neuron telling you an update existed. The setting now decides whether a new version is installed when you quit; you are told about it either way.
• Thinking arrived after the answer it belonged to, and a reopened conversation could report "Worked for 6h 53m" for a greeting.

This release stops one oversized screenshot from breaking a whole conversation, works around models that limit their own reply length, and temporarily turns off image-based shortening of long conversations.

Changed

• Image-based shortening of long conversations is off for now. Requests are sent in their plain form while this feature is paused. Long conversations may cost a little more than before, and nothing else changes. The caching that saves most of the cost still works. It will be turned back on after another review.

Fixed

• One oversized image no longer breaks the whole conversation. A screenshot larger than 8000 pixels on either side is rejected by the model. Because images are sent again with every later message, one such image could make a long conversation impossible to continue, with an error that pointed to a message number only. Such an image is now replaced by a short note that gives its size and asks for a smaller one. Everything else keeps working.
• Models with a reply limit no longer fail the whole turn. Some models accept a shorter reply than Neuron's model list says, and asking for more was rejected with an error about a setting you never chose. Neuron now reads the limit from the error, retries the same request within it, and remembers the limit for that model. This matters most if you set Reasoning Effort to a high level, since Smart Adaptive asks for less and usually stayed under the limit by chance.

First stable release. The version number is a declaration of intent, not a feature list — everything a user sees shipped in 0.0.2; what changed since is the release pipeline underneath it, which had never actually produced a signed, notarised, downloadable build until now.

Added

• Desktop releases are published to a separate public repository (sadatech/neuron-ide-app-releases) rather than to this private source repo. The electron-updater feed reads from there, so the update channel never needs a credential that could also read the source. Cross-repo uploads use RELEASES_REPO_TOKEN, scoped to that repo alone.
• macOS and Windows desktop builds run on a self-hosted Apple Silicon runner: arm64 natively, and macOS x64 plus Windows cross-built from it.

Changed

• Linux desktop builds are paused. AppImage cannot be cross-built from macOS, so they need a hosted Linux runner. The job now reports _skipped_ rather than _failed_ on a normal release, so a Linux gap no longer withholds the update manifest from the platforms that did build. Still reachable on demand via a desktop-linux-v* tag.
• CI is paused to manual dispatch only; all of its jobs need hosted runners.

Fixed

• Fixed macOS builds shipping unsigned. The exported .p12 carried no Developer ID Application identity, so electron-builder fell back to ad-hoc signing and Apple rejected every binary in the app at notarisation. The reported symptom was a build that ran for half an hour and produced nothing.
• Fixed cross-platform native binaries being silently omitted. npm installs the optionalDependency matching the _host's_ os/cpu, so cross-building for another platform shipped without local speech support; --os/--cpu alone do not override this, --force does.
• Fixed a race in the packaged-app smoke check: it read the daemon's pid file as soon as the file existed, which is before the writer has flushed it. Latent on native hosts where the window is microseconds; hit every time under emulation.
• Fixed the packaged-app smoke check attempting to launch a binary built for a different OS. It compared build arch against the host but never build platform, so an arm64 host cross-building win32-arm64 tried to execute a Windows binary on macOS.

No user-facing change. This release exists to exercise the auto-updater, which has never actually been observed working: 1.0.0 is the first build whose app-update.yml points at the releases repo, so it is also the first build that can be updated _from_. Everything below is release plumbing.

Changed

• Dry runs no longer notarise or request secure timestamps. Both are round trips to Apple that a build with discarded artifacts has no use for; measured against this certificate, timestamps alone cost 0.499s per signed file across 415 files. Releases keep both, and must: notarisation refuses a build without a secure timestamp, and such a build stops launching once the signing certificate expires.
• Paused the workflows inherited from upstream. Their deploy targets (neuron.sh, the neuron-app Pages project) belong to the upstream project, and the rest need Docker, nix, or secrets this repository does not have. Each keeps its manual trigger and records why it was paused.

Fixed

• Fixed the two macOS legs failing at random with errSecInternalComponent. Signing runs security list-keychains -d user -s ..., a read-modify-write of a setting owned by the user account rather than the build, so two concurrent legs raced it and the loser's signing keychain dropped out of the search path mid-build. They are now serialised.
• Fixed finalize-rollout failing its checkout. actions/checkout defaults to cone mode, which only accepts directories, so listing package.json in sparse-checkout aborted the job. Latent since the fork: every previous release had skipped this job, so its first execution was also its first failure — and the reason no release had ever published update manifests.

Added

• Added a chat-only Home surface: chats are agents with no workspace, files, or shell, sharing one hidden chat workspace, with their own sidebar section, status buckets, and auto-generated titles.
• Added web_search and web_fetch to the Neuron provider, offered in every mode including Chat, with the fetch trust boundary enforced in the daemon.
• Added an MCP client to the Neuron provider so agents can reach MCP servers the user configured, with lifecycle handling, tool namespacing, per-server permission, and a tool budget.
• Added Neuron Cloud Memory management to Settings: cloud and local memories share one Saved memories list with the account named above it, five per page, each row previewing its text and opening on demand, and cloud entries archivable from the row.
• Added a machine-local memory subsystem ($NEURON_HOME/memory.json) with agent tools, atomic private writes, corruption quarantine, and a Settings section, alongside Cloud Memory as the shared authority.
• Added file attachments to the composer: uploads are extracted to text at prompt-build time, including PDFs, with unsupported types refused rather than silently dropped.
• Added a tracked-changes bar above the composer showing changed files with Keep, Undo, and open-diff actions, capped at five rows with the rest scrolling.
• Added git staging to the checkout pane, with a dedicated commit box and stage/unstage controls.
• Added a context-window breakdown panel behind the composer ring, naming what filled the window rather than only how full it is.
• Added workflow-only built-in roles (Team Lead, Explore, Planner, Builder, Reviewer) in @getneuron/protocol with role-aware prompts, read-only policies, and automatic profile seeding.
• Added native sequential multi-agent workflow runtime (run_workflow catalog tool) that spawns and coordinates real, visible child agents with parent labels through Team Lead, Explore, Planner, Builder, and Reviewer stages.
• Added Automatic and Manual model selection for built-in role profiles in Settings with segmented controls, automatic model materialization (modelId: ""), and picker filtering to keep built-in roles exclusive to workflows.
• Added real-time per-turn mode routing in the Neuron daemon for Automatically mode (mode-router.ts), dynamically granting full workspace tools (Build) by default and restricting to read-only tools on explicit plan/investigation intent.
• Added hardened Plan ↔ Build mode handoffs in the composer with strict deliberation: Build → Plan transitions automatically, while Plan → Build requires explicit confirmation unless approving an existing plan.
• Added live context-window token meter and usage tracking for Neuron Hub models by parsing upstream context_length / limit fields and streaming used tokens (contextWindowUsedTokens) per request.
• Added account-wide rolling usage window and rate-limit summaries directly in the Context window meter popover, showing real-time 5-hour and 7-day quota usage, progress bars, and scheduled reset times backed by /v1/usage/window.
• Added persistent storage and automatic recovery for agent lastUsage (contextWindowMaxTokens and contextWindowUsedTokens), keeping context meter data visible across desktop app restarts without waiting for a new prompt.
• Added an ask_user tool so an agent can put a question to the user mid-turn instead of guessing.
• Added Sentry error reporting to the app, with a self-hosted collector and real stack frames.
• Added dedicated Hub Account and Usage settings pages with consolidated OAuth PKCE authentication state.

Improved

• Reasoning is drawn as reasoning rather than as a tool call: a muted header and a left rail instead of a bordered card, open while the model is thinking and collapsing itself once the answer arrives. The setting is now Expand reasoning while thinking, on by default.
• A model that does not read images no longer receives one. The composer marks the attachment unsupported instead of sending a file the model will ignore.
• The reasoning-effort picker is hidden for models that have no reasoning levels.
• Adaptive reasoning now scores prompts in any language. The old scorer read English keywords and a length tier that started several screens in, so a long Indonesian request was answered with no reasoning at all.
• Auto-commit messages are generated with the Reviewer profile's model.
• Bundled Neuron Skills trimmed from 92 to a curated set, all under the neuron- prefix, with Remove replacing Uninstall.
• Cloud Memory is presented as Neuron+ Cloud Memory with a description of what it does instead of its transport and URL.
• Token counts render their K/M suffix at full size.
• Improved workspace auto-title and branch generation by propagating the active chat's selected provider, model, and thinking options into FirstAgentContext.
• Improved desktop application icon assets across macOS ICNS and PNG formats, precisely scaled to the frozen 81% safe area.
• Polished Context window and Account usage popover layout with single-row aligned headers, full-width progress bars, and subtle divider spacing.
• Consolidated daemon auto-shutdown handlers on desktop dev exit to cleanly stop background processes.
• Improved model picker behavior to open direct provider selectors when only one provider is configured.
• Unified task list status indicators and styles across workspace tabs.

Fixed

• Fixed turns ending with no reply. max_tokens was a flat 8,192 while the thinking budget ran to 100,000, so the model was told to think for longer than it was allowed to speak — it spent the whole allowance reasoning and never reached an answer. The ceiling now follows the budget, and a turn that still ends empty says the output limit was reached instead of only that it ended.
• Fixed Retry sending the message a second time. The prompt is resent under the id it was first sent with, so the row is not duplicated and the model does not read the request twice.
• Fixed Failed to load agent on a whole conversation. A tool called with no arguments emitted a detail with no input, JSON.stringify dropped the key, and the client validator rejected the entire timeline response over one row. Validation failures now also name the field that failed instead of showing an empty error.
• Fixed a recurring 400 INVALID_ARGUMENT on Neuron Hub turns: a transcript carrying an empty message or two consecutive same-role messages is repaired on the way out, and a model that refuses a reasoning budget is retried without one and remembered.
• Fixed the reasoning setting appearing to do nothing — a finished thought was dropped from the stream entirely, leaving the setting with nothing to act on.
• Fixed Cmd/Ctrl+R reloading the web page instead of reloading the desktop window.
• Fixed workspace auto-title metadata generation failing with Sign in to Neuron Hub or empty fallbacks by deferring title scheduling until after initial agent authentication and retaining in-memory Hub credentials for internal metadata tasks.
• Fixed workspace auto-title generation being skipped when the desktop client precreates directory workspaces in a separate initial request.
• Fixed Automatically mode displaying tool-less limitations by eliminating empty chat branches and defaulting per-turn routing to capable Build mode with approval gates.
• Fixed missing Context window meter after app restart by restoring persisted token usage on initial agent hydration.
• Fixed layout clipping and text truncation in the desktop context quota popover by ensuring full-width container expansion.
• Fixed provider usage query gating so usage windows load reliably even when provider identities are resolving.

This release makes long conversations much cheaper, stops several ways a reply could fail completely, and fixes Cloud Memory asking you to sign in on its own.

Changed

• Long conversations cost much less. Neuron now keeps the repeated part of every request exactly the same from one message to the next, which lets the service charge a fraction of the normal price for it. In a test of three messages in an agent session, the cost went from $0.1993 to $0.0352. The biggest cause was the tool list appearing in a different order each time. One changed item threw away the saved data for everything after it, which cost more than any other issue found.
• Code search returns shorter excerpts. A search result now points to a file and a range of lines, instead of sending about a hundred lines per match. The same eight results, in the same order, now cost about a third as much. Since a conversation carries its history forward, the saving adds up with every later message.
• A smaller skill library. Nine guides aimed at academic reading, research funding, patents, course reading lists, and hiring have been removed. None of them are needed while you are writing code, and every guide was sent with every request. The library went from 87 guides to 78, and the extension download is 900 KB smaller.
• More room kept free before the chat fills up. Neuron's size estimate is accurate for ordinary writing, but it undercounts dense content — structured data returned by a tool, for instance — by up to 18%, which could push a busy conversation over the model's limit. Neuron now keeps extra room in reserve to cover that, at the cost of about 7% of the usable space.

Fixed

• Cloud Memory stops asking you to sign in on its own. Setting up a replacement key deleted the old key first. A network hiccup during that step left the device with no key at all, so VS Code opened its own sign-in prompt. A later attempt would quietly succeed, which made it look like the problem fixed itself. The replacement is now obtained before the old key is removed, and a single failed sign-in check no longer throws away a working key.
• A large tool result no longer ends the turn. Reading a big file, or several at once, could produce a message too large to send, leaving you no way forward except starting a new chat. Tool output is now limited per result and per message, and when the budget runs out, Neuron says so instead of failing.
• Long conversations no longer break when they are shortened. Making room for a new message could leave the conversation starting in the wrong place, which the service rejects. It could also drop Neuron's own instructions, so the longer a conversation ran, the less guidance the model had.
• Reasoning settings are no longer sent to models that do not support them. A model without reasoning was still given a thinking level, which it either rejected or ignored while cutting its answers short to match. Support is now read from Neuron's own model list.
• Saved chats stop growing without limit. The full output of every tool was saved in the chat history, which VS Code loads in full each time the window opens. A project with many tool calls could grow to hundreds of megabytes. Now only a preview is kept. The model still sees the same content.
• Model output can no longer corrupt the chat display. A rare character in a file or a recalled memory could replace part of a rendered reply with the word "undefined".

This release moves Neuron to its new platform address, stops Cloud Memory from repeatedly asking you to sign in, and fixes the model lists and the Context Window bar.

Changed

• Neuron moved to a new platform address. Sign-in, model lookup, embeddings, and chat all use neuron-platform.sadata.dev. An address you had set by hand to the old host is switched over automatically, so nothing keeps trying a host that no longer responds.
• Endpoint selection is hidden for now. Neuron currently runs on a single endpoint, so there is nothing to choose. The picker is gone from Options, and the background endpoint checks no longer run. They will come back when more endpoints are available.
• The Context Window bar starts with reserved space. The reserved amount is fixed, so placing it at the left gives the growing "used" part a stable edge to read against. Each number under the bar now has a colour swatch that matches the part it describes.

Fixed

• Cloud Memory stops asking you to sign in again. The sidebar's status check threw away a rejected key while VS Code was still connected with it, and nothing told the editor to switch to the new key. The next memory request then failed sign-in, and VS Code opened its own sign-in prompt. Every status check could trigger this again. The editor is now told to reconnect whenever the key changes, including at sign-in and sign-out.
• Cloud Memory is tied to your account, not your computer. If your first chat message came before sign-in finished, or the account lookup failed once, the device registered a memory that no one else could see, and never checked again. So memories did not follow your account to another computer. The device now moves to your account's memory as soon as the account can be read, and registers again when you sign in as someone else.
• A hotel or office sign-in page can no longer break Cloud Memory. Some networks answer every request with their own sign-in page. That page was mistaken for proof the memory key worked. The key was passed to VS Code, which then rejected it and opened the sign-in prompt.
• Neuron+ Agent no longer flickers to "(unavailable)". One slow or failed model list update emptied the list, so the Explore and Planner pickers marked a good saved model as unavailable until the next check. The pickers now keep the last list that loaded, and only mark a model unavailable when Neuron's list really no longer includes it.
• The Context Window bar shows what you are actually using. The "used" part was drawn at zero height and never appeared, the session-peak marker never showed, and the empty part of the track used the editor's colour for a filled bar, so the bar looked full no matter the percentage. Used, reserved, peak, and free space now look clearly different.

Added

• Added terminal reattach support and better worktree terminal handling
• Added global keyboard shortcut help in the app
• Added sidebar host filtering and improved agent workflow controls
• Added reliable Neuron Hub credential propagation when restoring stored agent chats after restarting the desktop app.
• Added account-scoped Cloud Memory signup and project-aware key rotation to prevent duplicate account keys.
• Added direct macOS Apple Silicon and Intel DMG links and synced Desktop 0.0.1 build details on the landing page.
• Added explicit Plan/Build handoff: planning requests switch Build to read-only Plan mode, while implementation requests from Plan require cross-platform approval before switching to Build; planning intent wins over implementation words, negated requests are guarded, mode-switch failures fail closed, stale mode events cannot revert the handoff cursor, and concurrent submissions share one handoff operation.
• Added voice capability status reporting in the client
• Added background local speech model downloads with runtime gating
• Added adaptive dictation finish timing based on server-provided budgets
• Added relay reconnect behavior with grace periods and branch suggestions

Improved

• Improved connection selection and agent hydration reliability
• Improved timeline loading with cursor-based fetch behavior
• Improved first-run experience by bootstrapping a default localhost connection
• Improved inline code rendering by auto-linkifying URLs
• Improved worktree setup visibility by streaming setup progress
• Improved terminal streaming reliability and lifecycle handling
• Preserved explorer tab state so context survives navigation better

Fixed

• Fixed stale protocol/client build artifacts that prevented the desktop app from compiling.
• Fixed Linux checkout diff watch behavior to avoid recursive watches
• Fixed stale relay client timer behavior
• Fixed unnecessary git diff header auto-scroll on collapse

This release ties Cloud Memory registration to each user's account email on all platforms, and fixes broken links between the 87 bundled skills.

Added

• Cloud Memory is tied to your account. When you set up Cloud Memory, Neuron looks up the email of the account you are signed in with and derives a fixed ID from it. Every session and device signed in to the same account now shares the same memory and recall history.

Fixed

• Links between skills work again. Links from one skill to another were broken in several of the 87 bundled skills, including the grants, dossier, pulse, litreview, patent, syllabus, research, and named-persona review skills. Each link now opens the correct section instead of a missing file.

This release unlocks the full context window for large models, adds an update check so old versions do not show confusing errors, and fixes chat streaming, size budgeting, and the order of tool steps.

Added

• Update check for chat models. When a newer extension version is available, chat models stay in the picker with an (update required) label. Requests are refused with a clear message asking you to upgrade, instead of failing partway through with unclear network errors. The sidebar keeps working normally.
• Full context window for large models. The artificial 256K limit is removed. Models with larger windows (1M or 2M tokens) now use their full size.

Fixed

• Gemini "invalid argument" errors with tool steps. Back-to-back messages from the same speaker, created while repairing tool steps or replaying sessions, are now merged into one turn. This keeps the order the service expects.
• An accurate size estimate for the instructions and the skill list. The instructions Neuron adds to every request, and the list of available guides, are now counted against the model's limit. Before, they were left out of the count, so a request that looked like it fit could still be rejected by a model with a smaller window.
• No more duplicated text in replies. The stream now stops cleanly at the end of each message. This prevents repeated text if a gateway retries or sends the same reply through two providers.
• Trim notice no longer shows by mistake. The "older turns trimmed" notice now appears only when messages were really dropped, not when tool steps were just repaired.

This release stops the agent chat from quietly switching to a different model when the one you picked is briefly unavailable.

Fixed

• The agent chat no longer switches models behind your back. If the model you chose is briefly missing from the list, for example during a refresh or a gateway failover, the turn now waits or shows a clear error. Before, it silently ran on whatever model was first in the list. That meant the cost was billed to the wrong model while the picker still showed your choice.

This release adds a library of expert guides that Neuron uses on its own.

Added

• Built-in skills, ready from the moment you install. Neuron now includes 87 expert guides covering backend design, code review, refactoring, documentation, testing, security, cloud, and research. There is nothing to set up or download.
• Neuron picks the right guide for you. You do not need to name a guide or remember what exists. When your request matches one, Neuron reads that guide and follows it. The rest stay out of the way.
• Your own guides come first. A guide folder in your project under .neuron/skills/ is used as it is, and replaces any built-in guide with the same name.

This release makes Cloud Memory the one memory Neuron uses, and fixes four issues that were quietly hurting chat, indexing, and code search.

Changed

• Cloud Memory replaces the editor's local memory files. Neuron no longer offers its built-in memory tool in its chat view. Notes go to one shared store, instead of being split between a local file and the cloud, where neither part could see the other.
• Cloud Memory instructions are now specific. Neuron is told to look up context at the start of a task and to save what was decided at the end. Before, it was left to guess when memory might help.
• Ask and Plan modes can no longer write to Cloud Memory. Both modes promise not to change anything, and Cloud Memory is shared with every other session, so only Agent mode can save, update, or archive a memory.
• Longer wait before a chat request counts as stalled. A model working through a hard problem can stay quiet for minutes. That was mistaken for a dead connection.

Fixed

• Chat requests no longer time out while the model is still working. The inactivity timer ignored the keep-alive signals sent during a long answer, so the request was cancelled in the middle of thinking.
• Instructions no longer disappear from shortened requests. When a conversation was shortened, Neuron's own instructions, including the rules for running tasks and using memory, were dropped from that request.
• The code index no longer stays on "Not indexed". Background indexing gives way to chat, but cancelling the very first pass threw away all of its work. Anyone who chatted every few minutes never finished a first index. The first pass now runs to completion.
• Code search works again. The tool was registered under a different name from the one Neuron advertised, so it was never actually available in Chat.
• Cloud Memory no longer shows "Offline" while working normally. A healthy server answered the status check in a way that was read as a failure.

This release fixes silent Cloud Memory sign-in for existing Marketplace installs.

Fixed

• Cloud Memory stays out of your way. The extension sends its automatically created key in a header the server understands, instead of a sign-in method that VS Code could mistake for a browser login.
• Cloud Memory sets itself up in the background. On first use, the extension registers the device, stores the key securely, and connects. It does not open a browser or ask for sign-in details.
• Old Cloud Memory keys are fixed before startup. A saved key is checked first. If the server rejects it, a new one is created before VS Code connects, so VS Code does not fall back to a browser sign-in.
• Old keys are checked before being moved over. A revoked key from a workspace can no longer skip the automatic repair and push VS Code into a browser sign-in.

This release tidies the Neuron sidebar and keeps Cloud Memory setup in the background.

Changed

• Reasoning Effort is now shown above Context Window. The control moved out of Options and into the status card, so the current thinking level sits close to the context usage bar.
• Cloud Memory repairs its own sign-in. A rejected saved key is discarded, so the device can quietly create a new one instead of staying stuck on an error.

This release brings Neuron+ Cloud Memory into VS Code through the editor's own tool-server support.

Added

• Neuron+ Cloud Memory, set up inside VS Code. Neuron's long-term memory is now available to Chat without installing anything separately. It sets itself up the first time it is used, your sign-in is stored securely by VS Code, and the connection is always encrypted.

This release cuts the cost of long conversations by turning bulky context into images, and fixes how the system instructions reach the model.

Added

• Shorter long conversations. When a model that can read images is selected, Neuron turns the large, repeated parts of a request (the instruction text, tool descriptions, and older history) into compact images before sending. An image takes far fewer tokens than the same text, so long sessions stay within budget for longer. Exact values and recent messages stay as text, and shortening is skipped whenever it would not save anything.
• Guidance for the code index. Neuron is now told to search your project index before editing or answering questions about unfamiliar code, to search using real names from the source, and to open the cited file and lines before relying on an excerpt.

Fixed

• VS Code's editor instructions no longer look like something Neuron said. VS Code has no separate slot for instructions, so it sends them as the first message from the assistant. Neuron passed that message on as conversation history, which made it look like the model had written its own instructions. Those instructions are now sent in the proper place for system instructions.
• Prompt caching now works. The cache marker was placed in the wrong spot, so the repeated start of each conversation was never cached and every message was charged in full.
• Image attachments no longer fail in long conversations. Once a request had enough images, the service switched to stricter rules and rejected ordinary screenshots over 2000 pixels. Neuron now stays under that size.

This release makes Smart Adaptive respond better to how complex a task is, and gives native Neuron Chat requests a consistent set of working rules.

Added

• Sentry test command. Run Neuron: Test Sentry Error from the Command Palette to check that error reports reach the configured Sentry project.
• Error reports without private data. The extension sends error reports automatically, with your prompts, workspace, requests, user details, and tool data removed. This works regardless of VS Code's telemetry setting.
• Smart Adaptive thinking level chosen from the task. Smart Adaptive now looks at how long your message is, whether it is about coding or debugging, its structure, how long the conversation has gone, and past tool use. From this it picks a thinking level between None and Max, without an extra request.
• Adaptive output length. Smart Adaptive scales the length of each reply with the chosen thinking level, while staying within the model's limit. Harder tasks get more room for the answer.
• Built-in working rules for Chat. Neuron requests now include rules to stay within the requested scope, keep unrelated changes intact, plan complex work, check the result once, stop after a failed check, and treat file and tool output as information, not as instructions.
• Faster repeated context. Requests now ask the service to keep the repeated start of each conversation, which reduces cost and wait time in longer conversations.

Changed

• Smart Adaptive no longer uses one fixed Medium level. Simple questions can skip extra thinking, while more involved ones get progressively more thinking and a longer reply.
• Error reporting can retry after a failed start. The error reporting is marked as ready only after it has started successfully.
• The Sentry test reports the result accurately. The test waits up to five seconds for confirmation and shows the event ID if the report is still on its way or timed out, instead of always claiming success.

This release adds a second, more opinionated way to talk to Neuron, and a higher thinking level for the hardest problems.

Added

• A new "Ultimate" thinking level. The previous top level is now called Powerful, and Ultimate sits above it as the new highest choice in the Options tab. Each level now has a small icon in the picker.

This release removes two settings that duplicated work VS Code already does, and reworks the Context Window indicator so it shows what the current conversation actually costs.

Changed

• The Context Window row shows your latest request, not just the highest reading. The percentage used to show the highest value ever reached for a model, so one large request kept the bar at 100% for the rest of the session. The main reading now shows the latest request, and the session peak appears as a thin marker on the bar.
• The striped part of the bar is labelled clearly. It used to say "reserved for response", but it also covers tool and request overhead and the safety margin. It is now called "protected" space. The exact split (reply, overhead, safety) appears when you hover, along with the current prompt, the session peak, and whether older messages were trimmed.
• The limit shown is the limit Neuron actually uses. Some models advertised their full size to VS Code (for example 1M), while Neuron capped requests at 256K. VS Code then built prompts that were trimmed on every turn. The advertised limit now matches the enforced one, and the model's full size is still shown in the tooltip.
• Endpoint and Reasoning Effort share one card in the Options tab, stacked as two rows separated by a thin line, matching the Context Window and Codebase Index rows.
• Endpoint checks pause while one endpoint is pinned. Health checks exist to keep Smart Adaptive on the best endpoint. When you pin one yourself, the periodic checks stop, since they can no longer change anything. The endpoint list then shows each gateway's reported status instead of a latency figure that stays frozen from when you pinned it. Switching back to Smart Adaptive re-checks every endpoint right away. A failed request still triggers a re-check while pinned, so a dead endpoint cannot leave you stuck.

Removed

• Context Size and Smart Compact History are gone from the Options tab. Both duplicated work VS Code Chat already does. Smart Compact's on-device summary also broke down in exactly the long conversations it was meant to help: past a certain length it shrank the history to a short excerpt and left out tool calls and their results, so agent sessions lost almost everything. Neuron now follows each model's own context window (still capped at 256K), and a conversation that is too big has its oldest messages trimmed as a safety measure. The /compact chat command, the summary cache, and the related saved settings were removed with it.

Fixed

• The Neuron icon in the Activity Bar is no longer a solid grey block. The 1.14.1 change used the full-colour logo there, but VS Code shows that icon as a silhouette in the theme colour. Because the logo is a rounded square, the result was a plain square. A dedicated one-colour icon now shows the Neuron mark, and it brightens and dims with the active state like other icons. The full-colour logo is still used on the Marketplace page.
• The context bar can now reach the end of its track. The used and reserved parts were squeezing each other as a prompt neared the limit, so the bar understated usage exactly when the context was nearly full. The track is now split into a usable area and a reserved end, which always add up to 100%.
• A rough token estimate no longer inflates the session peak. The on-device estimate was recorded as the highest reading before the service reported the exact prompt size, which could keep the peak high for the rest of the session. The service's reported count is now the one that counts. The estimate is shown only as a preview.
• A slow request can no longer overwrite the newest reading. When several requests overlap, an older response finishing late can no longer replace the indicator with out-of-date numbers.
• Long conversations without tool calls no longer fail. Shortening a conversation could leave it starting with an assistant reply, which the service rejects, since it must start with your message. Shortening now always starts the history correctly.
• The Context Window bar matches its percentage. The percentage measured usage against the usable budget, while the bar measured against the budget plus the reserved reply space. So a full prompt showed "100%" while the bar stayed at about two-thirds and could never fill. Both now measure the same thing.
• The peak context reading resets on sign-out. The per-model high point survived sign-out, so a new session could show a percentage based on a previous account's conversation.

Changed

• Extension icon updated to the new Neuron logo. The new logo replaces the old generic speech-bubble icon in the Activity Bar and on the Marketplace.
• Changelog and README updated for version 1.14.1. Notes on the 1.14.0 performance work (batched streaming, bounded caches, safer shutdown) and the 1.14.1 empty-message fix are included below.

Fixed

• Compacted conversations no longer send an empty message list. After VS Code replayed a /compact reply, or after broken tool-step history was repaired, the outgoing history could end up empty and the request was rejected. Neuron now detects this, shows a helpful retry message, and never sends the invalid request.

This is a performance and reliability update that keeps Chat, the Neuron sidebar, and background indexing responsive during long sessions.

Changed

• The release is now 1.14.0. This update combines prompt streaming, panel lifecycle, background indexing, gateway monitoring, and caching into a safer, lighter setup.

Fixed

• Background requests stop when their owner closes. Usage checks, gateway discovery, endpoint checks, Marketplace checks, embedding requests, timers, watchers, and cancel listeners are stopped when the panel is hidden, the window loses focus, you sign out, or the extension shuts down.
• Hidden panels receive fewer updates. Dashboard refreshes pause when VS Code is in the background or the sidebar is hidden, and old results are not shown after a newer refresh or after the view closes.
• Chat requests no longer resume indexing too early. Automatic indexing waits until all active prompts finish. Cancelled or pending indexing work is kept and resumed safely afterward.
• Stored state stays bounded. Per-model limits and context readings cannot grow without end if a gateway keeps changing model IDs. Summary and model-list storage also stay bounded.
• Closing the panel does not cancel a shared model request. Closing the panel can no longer cancel a model request that is shared with VS Code's native Chat.

Performance

• Streamed replies appear in small batches. Rapid pieces of text are grouped before they reach VS Code Chat, which reduces overhead while keeping the text in order and tool steps in their place.
• Preparing a prompt does less repeated work. Fitting the context, estimating size, converting messages, and mapping models reuse work already done instead of scanning the same history and model list again and again.
• Background indexing is less disruptive. Indexing pauses between batches, slows down automatic work, waits while Chat is preparing or running a prompt, and resumes only when there is work left to do.
• Less idle background work. Model changes, model list saves, dashboard refreshes, gateway checks, and progress updates are merged, so unchanged information does not trigger repeated screen, disk, or network work.

Added

• Context Window now shows peak usage per model instead of per request. Each model records the highest token count seen in a session, so the panel shows a steady "this model has used up to X%" instead of jumping with every message. Each model keeps its own peak. The peak resets only when you switch models or restart VS Code.

Fixed

• The shortened summary no longer repeats its heading and content. The on-device summary used to include its own heading, and then another heading was added around it. That sent the same content twice and wasted room. The heading is now added only once, and the summary itself stays clean. The summary also keeps fewer entries (3 important ones instead of 6, with shorter excerpts) so it stays tighter.

Changed

• Context Window and Codebase Index share one status card. The two indicators are now stacked rows in a single card above Log out, instead of two separate boxes. Each percentage or status stays beside its title instead of wrapping onto its own line, which gives more information in less space.
• Codebase Index has no button anymore. Indexing has always run automatically after you open a project and sign in, so the manual Index Workspace button repeated that without adding anything. To rebuild the index right away, use Neuron: Index Workspace (Semantic Search) from the Command Palette.

Fixed

• Tools with no inputs could break a whole chat request. Tools that declare no inputs were sent without any input description, and the service rejected the entire request once one of them was included. This showed up as a missing field 'parameters' error. Tools with no inputs now send an explicit empty description.
• Smart Adaptive was thinking at High level, not adaptively. With no explicit level, adaptive thinking defaults to High on the service side, so the "all-rounder" setting was running almost at maximum. Adaptive now sets the level to Medium and lets the model decide how much to think on each message.

Changed

• The Context Window block is a single-line summary. The used / limit tokens line was removed. The heading and the percentage share one row above the bar, which saves a line in the sidebar. The exact token counts, the reserved reply space, and the model's window are still shown when you hover, and are read out as the bar's value.
• Removed an unused indexing fingerprint. A size and date fingerprint was saved for every file but never compared. Indexing still checks the content of each file, because some file systems and remote services keep the same size and date even when the content changes. Without that check, the index could go stale. The 1.13.7 note describing this unused path has been corrected.

Fixed

• Neuron models could still disappear from the Chat model list. The background check added in 1.13.8 rarely ran. It checked only every two minutes and refreshed only a list between four and five minutes old, so checks at four and six minutes skipped that window completely, and the list expired without being renewed, which was the exact problem the check was meant to stop. The check now refreshes the list as soon as it reaches the refresh age, with no upper limit, so an old list is always renewed. The timing is based on how long the list is kept, so the two cannot drift apart. If a refresh fails, the existing list stays and the next check tries again.
• Usage windows that reset every minute no longer show an old reset time. Windows that reset on a fixed schedule are sent by the service with a fixed: prefix (for example fixed:1m). The check for short windows did not recognize this prefix, so the one-minute window kept showing a reset time that went stale within seconds. Sliding windows such as 1m were not affected.

Changed

• The Context Window panel matches VS Code's own context indicator. The token count and percentage sit on one line under the heading, and the bar shows a striped part for the space reserved for the model's reply, with a matching legend.

Fixed

• Context Window no longer understates the prompt size. With prompt caching, the service splits the prompt into three counts: new input, cached input read, and cache written. Only the first was being read, so a message that used the cache could show almost nothing while the real prompt held over 100K tokens. All three are now added together, as Anthropic's documentation describes.
• The Context Window percentage matches the room Neuron actually uses for trimming. The panel divided by the full model size, while messages were trimmed against a smaller budget after the safety margin, the reserved reply space, and per-request overhead were taken out. Both now use the same number, so the percentage shows the share of usable space that the prompt really takes.
• The Context Window bar no longer shrinks its parts when nearly full. The used and reserved parts were squeezed as the bar neared full, so it understated usage at exactly the point where the warning matters most.
• Neuron models no longer disappear from VS Code's Chat model list. A background check refreshes the model list before its five-minute storage expires, so the list stays filled without a manual refresh. The check stops cleanly when you sign out or the extension shuts down. (The timing was too narrow to work reliably, and was fixed in 1.13.9.)

Fixed

• Sending a message no longer waits on a network token count. VS Code's token counting and the pre-send size check now use the on-device estimate, so pressing Enter starts the request immediately. Before, it could wait up to 1.5 seconds for the token-count service. That service is still available as an optional helper outside the send path.
• Sending a message no longer reloads the model list. Silent requests from VS Code, including the ones made as you press Enter, now read only the saved model list. The service is asked for the list when you open the model picker or refresh it by hand, which keeps sending clean.
• Warm chat requests no longer read the keychain before sending. The sign-in token is kept in memory after its first secure read, and kept in step on sign-in and sign-out. Normal sending can use it straight away, without another keychain lookup.
• Background indexing no longer slows chat when reusing a large index. Files that have not changed skip splitting and embedding, using a saved content fingerprint. Earlier pieces are grouped once instead of rescanned for every file, large files respect the batch size, and long passes pause now and then to keep the editor responsive. Reuse still checks each file's content, not just its size and date, because some file systems and remote services keep both unchanged when the content changes. Skipping that check would leave stale results in the index.

Added

• Optional exact token counting. VS Code's token counts can use the service's counter, with a short time limit and a small cache. If the counter is unavailable, the on-device estimate is used.
• Request IDs and safe diagnostics. Each chat request carries an ID, and the extension Output panel logs basic timing and error details. Prompt text, tool data, and keys are never logged.
• Context usage survives closing the panel. The latest Context Window reading is saved and restored when the sidebar is reopened.
• Faster indexing of changed projects. Files that have not changed reuse their existing index entries. Changed, new, and deleted files are updated safely.

Changed

• Groundwork: Chat's internals are split into smaller parts. Fitting a conversation to the window, shortening history, converting messages, handling images, and reading the reply stream each live in their own place now rather than one shared file. Nothing you see changed; this is what made the fixes in this release possible.
• Neuron uses the size details the service reports, when it sends them. Model information may include how much a model can hold, and Neuron now uses those figures rather than only what it has worked out itself. When they are missing, what Neuron has learned is used instead.
• The Context Window indicator is always visible above the dashboard tabs. It shows the latest prompt usage and stays in view across Limits, Tokens, Models, and Options.

Fixed

• Retry behavior for chat is clearer and tested. When a request is rejected for being too large, Neuron retries exactly once with a tighter limit. Rejections for invalid or unauthorized requests are not retried.

Added

• Usage limit rows show the exact count on hover. Hovering over the percentage in a quota row swaps it for the underlying count, such as "0/60 requests" or "68.6M/200M tokens", and changes back when you move the mouse away. This uses the same measure (tokens or requests) that the row already shows.
• Short usage windows no longer show reset times. Windows of one minute or less hide the reset time, since it goes stale almost at once. Longer windows still show it.
• A clearer sidebar layout. The top five models list now sits under Tokens, and the choice of model for Explore and Subagent and for Planner now sits under Models.
• Shortcuts in the footer and a model refresh. You can open the Neuron web panel or the model status page from the sidebar. The quota refresh button now also refreshes the Neuron model list in VS Code Chat and the Models tab.
• Shortening long chats runs entirely on your computer. Shortening no longer uses Copilot, Neuron, or any other model or network request. It pulls out the key context locally and falls back to plain trimming when needed.
• Updates are checked every 5 seconds. When a newer version is found, a full-screen notice appears with a shortcut to the extension page.
• The loading screen stays until data is ready. Partial account or version information no longer shows a half-loaded panel while account and usage data are still loading.
• Explore, Subagent, and Planner model choices are in Options. Choose a model that supports tools for each built-in agent. Auto now uses Neuron's own automatic model when it is available.
• Neuron models can be picked for both Explore and Plan agents. Models that support tools are now recognized correctly, so eligible Neuron models appear in both choices.

Changed

• The sidebar's background refresh pauses when the window is not focused. The 30-second refresh, which made three network requests each time, used to keep running while VS Code sat in the background. That wasted CPU, battery, and network on slower machines. It now stops when the window loses focus and restarts with one refresh when focus returns.
• Automatic indexing no longer pushes the CPU to 100%. Indexing used to run in one long burst that kept a CPU core fully busy and used a lot of memory. It now pauses between batches so the editor stays responsive. Automatic indexing also takes short breaks between batches, so a background refresh cannot take over a CPU core. The manual Index Workspace command still runs at full speed.
• Automatic indexing fully stops watching files after its first pass. Before, after the first pass, the extension kept watching the whole project for changes for the rest of the session, which slowed down low-end machines. The watcher is now shut down once the automatic pass finishes, and it starts again only for a new session (a new sign-in, a different project, or re-enabling the setting). Idle cost now drops to zero.
• More space below the Log out button. The gap between Log out and the version line at the bottom has been widened.

Fixed

• Automatic indexing could run forever on projects with a live dev server. If a file that gets regenerated (such as a framework's cache or compiled view) kept changing during an indexing pass, the indexer started another pass right away, and never finished. The completion check that shuts the watcher down was never reached, so the CPU stayed at 100%. Common generated folders (Laravel storage and bootstrap/cache, .turbo, .cache, .parcel-cache, .angular, coverage, tmp, target) are now skipped by indexing and file watching. A limit on back-to-back automatic runs also guarantees the loop always ends, even for folders not yet on the skip list.
• The 1-minute usage window no longer disappears from the sidebar. Quotas limited by request count (such as fixed:1m, which reports no token limit) were filtered out of the Usage Limit card, because the card only understood token limits. The card now shows request-based windows too, so they appear alongside token-based ones.

Added

• Codebase indexing can now stay up to date automatically. When turned on (the default), Neuron refreshes the search index on your computer shortly after the extension starts. Changes to source files are grouped into one background refresh. Automatic runs never send source code to a remote service for embedding. Dependency and build folders are skipped. If files change during a run, one follow-up refresh follows. Closing the project, signing out, or turning this off cancels pending work. A failed refresh keeps the last working index, and you can turn off automatic runs in settings while keeping the manual command.

Fixed

• Summarizing a long chat no longer stalls a prompt behind a chain of requests. Even very long histories now use one limited summary request that keeps the oldest goals and the latest state. If it cannot finish within two minutes, Neuron falls back to plain trimming. The summary length scales from 1K to 4K tokens depending on the model's window, and stays within small windows and output limits.

Fixed

• Code search results no longer cut off in the middle of code. Results are now shown as whole blocks. Instead of cutting a result at a random point, Neuron leaves out the next lower-ranked result entirely. When a single result is too big, only that result is trimmed, at a line break, and its code block is always closed properly.
• Backticks in code can no longer break search results. Each result's code is wrapped in a block marker longer than any run of backticks inside it, so code samples that contain triple backticks cannot end the block early.
• Malformed search results no longer fail the whole search. A missing or invalid score is shown as "score unavailable", missing file or line details show a readable placeholder, and unexpected content is converted safely.
• Search truncation is reported exactly. Instead of a vague "results truncated" note, the output ends with a clear count of results shown, results left out, and snippets trimmed. This note always sits outside the code blocks and within the size limit.

Changed

• The sidebar no longer loads an unused embedding-model picker. Codebase indexing stays on your computer by default and uses a remote fallback only when needed.

Fixed

• Streamed tool calls keep all their input. Arguments given when a tool step starts are combined with the pieces that follow, instead of being discarded. Final pieces are also processed even when the connection closes without a trailing line break.
• Tool steps and results run in a valid order. Tool results come before any text from the user, and several tool calls and their results stay paired through history repair and trimming.
• Remote indexes keep the model that built them. A saved index records the exact embedding model used, so a changed model list cannot mix incompatible vectors. Older remote indexes without this record are rebuilt safely.
• Code search output is limited. The number of results is kept between 1 and 20, and the output is capped at 48K characters so a large or minified file cannot flood the chat.
• Panel and extension resources are cleaned up properly. Timers, listeners, and watchers are shut down together with the part of the extension that owns them. Changing project folders loads the matching saved index.
• Sign-in no longer waits forever for the token exchange. The sign-in callback times out after 30 seconds, and update-check timers are cleared on every path.
• Usage still appears when parts of the service are down. Token usage shows even if the profile or quota endpoints are unavailable.
• Installing a local build works again. The helper that packages and installs the extension now runs its preparation step correctly.
• Saved index timestamps reflect the finished run. Storage failures now show as errors instead of leaving the index stuck.
• Broken tool steps are removed as a pair. When trimming, summarizing, or replaying history, a tool request and its result are removed together. This prevents an orphaned result from reaching the service.
• Very long requests are caught before sending. A message that is too long is reported locally with a clear explanation, instead of being sent to fail. Tool definitions and request overhead are now counted in the budget too.
• Shortening works properly. Summaries are checked against the final budget. A cancelled summary does not continue into a normal request. A failed progress update no longer repeats the same summary request.
• Unknown model sizes use a safe default. Models without size information no longer get an invented 200K history budget, and reply limits are capped for small known windows.
• Summaries are kept per model and stay stable. A summary made for one model is never reused for another, and new messages reuse the existing summary instead of rewriting it each time.

Performance

• Gateway availability updates by itself. A single monitor in the extension checks the status of the selected endpoint regularly, runs a full speed check only now and then or after a confirmed failure, backs off when offline, and stops cleanly on sign-out. The sidebar is updated only when availability, the selected endpoint, or speed changes.
• Smaller production files. The build now compresses the code in two stages, which cuts the main file from about 112 KB to about 51 KB and the sidebar file from about 19 KB to about 12 KB. Public names and protocol details are left unchanged.
• Sidebar refreshes do less background work. The 30-second refresh now fetches only the dashboard data. Endpoint discovery and health checks, and release lookups, run only during a full status refresh. Overlapping refreshes are merged, independent requests run together, hidden views pause, and Git and Marketplace lookups are saved for the session.
• Model lookups are saved and shared. Simultaneous requests for the model list share one call, successful results are kept for five minutes, and signing in or out or refreshing the list clears the saved copy. Indexing still asks for fresh embedding models.
• Code search avoids sorting the whole index. Only the top matches are kept, and vectors that are already normalized use a faster calculation. In a synthetic test with 20,000 chunks of 512 values, the median search time fell from 29.19 ms to 21.52 ms (about 26% faster).
• Indexing progress updates are merged. The counts stay exact, but updates to the sidebar are limited to about ten per second, with the final result sent right away.

Changed

• Endpoint status is clearer. The Adaptive badge sits beside the endpoint's speed figure, and draining gateways show a warning icon.

Fixed

• Choosing an endpoint affects only chat and prompt requests. Sign-in, usage, model lookup, and embedding requests always use the main Neuron gateway, so picking a regional endpoint cannot disturb the sidebar or your account data.
• Smart Adaptive respects gateway availability. Gateways reported as draining are greyed out in the picker and never receive new chat requests or get picked by Smart Adaptive. A saved gateway that starts draining returns to Smart Adaptive automatically.
• Gateway health is clearer and more reliable. A failed health check is retried once. The Endpoint view now distinguishes DNS problems, timeouts, network failures, and HTTP errors, instead of calling every failure simply "down".
• The Codebase Index panel stays readable in narrow sidebars. Its status moves to its own line when space is tight, and the indexing tooltip wraps properly.

New

• Choose your endpoint. A new Endpoint picker in Options lets you choose which Neuron endpoint your chat and prompt requests go to: Neuron+ Smart Adaptive (chosen automatically) or a specific regional gateway. The list of gateways comes from Neuron.
• Smart Adaptive picks the fastest endpoint for you. Neuron checks each endpoint's health and speed in the background and sends Smart Adaptive to the fastest one that is working. Options shows a live list with each endpoint's status and response time, and marks the one in use.

Changed

• Shortening a long conversation won't get cut off. Summarizing history (including /compact) now has a much longer time limit, so large conversations finish instead of stopping early with "Cancelled."
• Indexing covers the whole project. The old limit of 5000 files has been removed, so large projects are fully indexed. Very large individual files are still skipped.

New

• Faster, model-aware Context Size. The Context Size options are now three simple levels that scale to the model you use: Smart Adaptive (the model's full window), Balanced (about half), and Fast (about a quarter). Smaller levels send less history with each message, which is the main reason replies arrive faster. Choose Fast for the quickest replies, or Smart Adaptive for the most context. Old numeric settings are converted to a sensible level automatically.
• Summarize the conversation on demand with /compact. Type /compact as a chat message and Neuron summarizes the conversation so far, streams a short recap, and remembers it. The next reply keeps the main points without resending the whole history. This is useful for trimming a long session before asking something new.

Fixed

• The codebase index no longer shows "Ready" with 0 files. File counts are now saved and restored, so reopening a window shows the real numbers. Switching folders no longer carries over the previous folder's status.
• The full-screen loading screen appears only when you first open the panel. Refreshing usage, or the panel's own background refresh, now updates in place instead of briefly blanking the whole panel.

Changed

• On-device indexing is always on. Indexing runs entirely on your computer by default, and no embedding model is needed. If an on-device pass ever fails, Neuron automatically uses an embedding model for that run. The separate on/off switch has been removed, so it just works.

Fixed

• Long replies no longer time out. Chat used a fixed two-minute limit for the whole reply, so answers that legitimately took longer (deep reasoning or long output) were cut off with "Neuron request timed out or was cancelled." The limit now depends on inactivity: a reply stops only if the connection goes silent, not just because it is taking a while.
• Indexing no longer flashes and fails with "Cancelled." Starting indexing twice in quick succession (a double-click, or the sidebar refreshing mid-run) could make two runs collide and show a false cancellation. Overlapping runs are now ignored, so the run in progress finishes cleanly.

New

• Index your codebase without an embedding model. Indexing now runs entirely on your computer. Neuron builds a searchable map of your files locally, so nothing is sent to the server and no embedding model is needed.

Changed

• vendor/ is no longer indexed. Third-party dependencies under vendor/ (from Composer and similar tools) are skipped during indexing, just like node_modules, so search results focus on your own code.

New

• Choose whether long chats are summarized or simply trimmed. A new Smart Compact History switch in the Options tab decides what happens when a conversation outgrows your context size. Leave it on (the default) to have Neuron summarize older messages into a short recap that keeps their main points. Turn it off to drop the oldest messages instead. That costs less, but the older context is lost.
• See when Neuron is summarizing. When a long chat is summarized for the first time, a progress notice (Neuron: Compacting conversation history…) appears while it works, like indexing does. It is clear that something is happening, rather than the reply looking stalled.
• A cleaner panel while it loads. Opening the sidebar shows a full-screen loading indicator while your account and usage load, then reveals the panel once everything is ready. The empty dashboard no longer flashes briefly.

Fixed

• Long chats no longer re-summarize on every message. Auto-summarizing now covers a stretch of history once and reuses that summary on later messages, instead of rebuilding it from scratch each time. This removes a large hidden cost, since a big summary request used to repeat on every turn, while keeping your context intact.

New

• Long chats keep their context instead of losing it. When a conversation grows past your chosen context size, Neuron now summarizes the older messages and keeps that summary in the conversation, rather than silently dropping the oldest turns. Your recent messages stay intact, and a short status-bar notice tells you when this happens, so long sessions stay coherent without starting over.

Fixed

• Chats with tools no longer break when the context is shortened. Trimming or summarizing old history always keeps a tool call and its result together, so a result is never left without its call. This makes tool-using chats reliable at every context size.

New

• Choose how much conversation Neuron remembers. A new Context Size setting in the Options tab controls how much chat history is sent with each request. Leave it on ✦ Neuron+ Smart Adaptive (the default) to follow each model's own context window, now capped at a sensible 256K so large-context models do not run up cost or wait time. Or pick a fixed size (128K, 256K, or 1M) for tighter, cheaper context. Each option has a short description.
• The codebase index is easier to find. The Codebase Index section now explains why it matters. An info tip beside the label describes what indexing does, and a short prompt under the button suggests running it once so Neuron can search your code during chat. The Index Workspace button also looks like a real button now.

Improved

• Codebase indexing is more reliable. Each run now checks which embedding models are available and picks a working one automatically. It skips models the server lists but cannot actually run, and recovers on its own if your previous choice was removed or renamed. When something goes wrong, the message gives the real reason instead of a generic error, and the Neuron Companion output panel records the details for troubleshooting.

Fixed

• "What's New" opens correctly after an update. The changelog tab could open blank because the file name's capitalization did not match what the extension looked for. It now finds and opens the file reliably, and never opens an empty tab if the file is missing.

Fixed

• Codebase indexing now always uses the top embedding model. Once an embedding model was chosen, it stayed selected even after the server ranked a different model first, so re-indexing looked stuck on an old choice. Every Index Workspace run now fetches the model list again and uses the current top-ranked model. The saved choice is kept only as a fallback if the list cannot be fetched at all.

New

• See what's new after every update. When Neuron Companion updates, this page opens automatically so you can catch up on the changes, like VS Code's own release notes. You can reopen it any time from the Command Palette with Neuron: What's New.
• Smarter thinking by default with Neuron+ Smart Adaptive. New chats let the model decide how deeply to think about each question. Quick answers stay fast, and hard problems get more thought. You can still pick a fixed level (Low, Medium, High, or Max) anytime in the Options tab, and each choice now has a short description.

Improved

• Thinking applies only to models that support it. Smart Adaptive and the thinking levels only affect models built for deeper reasoning. Simple chat models are left alone, so no time or tokens are spent on a setting they would ignore.
• Clearer error messages. When something goes wrong, such as an expired sign-in, a usage limit, or a problem on our side, you now see a short plain message that says what happened and what to do next, instead of a wall of technical text.

Fixed

• No more stray text in replies when using tools. In some tool-based chats, a small internal marker could leak into the conversation and show up as odd text. It is now filtered out, so replies stay clean.

Fixed

• Thinking levels now really change how the model thinks. Choosing a level in the Options tab had no real effect before, since every choice behaved the same. Low, Medium, High, and Max now each genuinely change how much the model thinks.

Improved

• More reliable indexing. If the codebase index briefly cannot reach the server, it keeps working with your current setup and switches to a newer option automatically once it is available.

New

• The sidebar keeps itself up to date. While you are signed in, your profile, usage, limits, and index status refresh by themselves every 30 seconds. You do not need to reopen or reload the panel.
• Build info in the footer. The version number in the sidebar now shows a short build ID next to it, which helps when reporting a problem.

Fixed

• Indexing no longer looks stuck. If something went wrong while building the codebase index, the progress bar could keep moving while nothing happened. Errors now appear clearly, so you know to try again.

New

• Codebase indexing just works. If you have not picked an embedding model, indexing chooses a suitable one for you automatically and remembers it.
• Update notices. A version label now sits under the Log out button, and an "Update available" badge appears when a newer version is out. Click it to go straight to the update.

Fixed

• The panel no longer breaks in some setups. Fixed a case where the sidebar could fail to load, and the "Update available" badge now appears only when a newer version really exists.

New

• Search your codebase by meaning. Choose an embedding model in Options, run Neuron: Index Workspace, and Neuron models in Chat can find relevant code by what it does, not just by exact words. This is especially useful in large projects. A status card with a progress bar keeps you updated, and indexing runs only when you ask it to.
• Smarter handling of long conversations. Very long chats are trimmed automatically to fit the model, and your most recent message is always kept.

Fixed

• Failed tool steps are reported correctly, so the model can tell a failed action from a successful one.
• Clearer usage limits. Accounts with unlimited usage no longer show a stuck 0% bar.

New

• Refresh usage on demand. The Usage Limit card has a refresh button in its corner.

Fixed

• Thinking levels behave as expected. Higher levels no longer quietly drop to a lower one, and the picker no longer jumps back to Medium after you choose Adaptive, Extra High, or Max. A mislabeled level was also fixed.

Improved

• Simpler usage limits. All your quota windows are now combined into one Token Quota card, with its own progress bars and reset times.

New

• Usage limits at a glance. A new section shows your request, token, and cost limits as easy-to-read progress bars.
• Tabbed sidebar. Usage Limits, Token Usage, Models, and Options are now organized into tabs, so the panel stays tidy.

Improved

• Shorter models list. The Models view shows your top 5 models to keep things focused.

New

• Reasoning effort picker. Choose how hard Neuron models think, right from the sidebar. VS Code's own Chat has no setting for this. Your choice is remembered across windows.
• Plan badge. Your plan (for example Free or Pro) now appears next to the "Neuron" title in the sidebar.

Fixed

• No more stray code in replies. Some models occasionally showed raw tool-call text in Chat. That is now handled cleanly behind the scenes.

Improved

• Quieter, release-ready build. Internal diagnostic logging was removed. Images, chat, and token counts keep working as before.

Fixed

• Accurate token counting. Your session token usage no longer stays at 0. It now reflects text, chat messages, tool activity, and images.

Fixed

• Correct context size. The session context counter no longer shows 0 tokens, so VS Code can manage long chats properly. Model tooltips now show the full context window too.

New

• Paste images into Chat. Neuron models can now see and read images you paste, in common image formats.

Improved

• More reliable chat and tools. Every model now uses Neuron's native chat path, which makes streaming replies and multi-step tool use more dependable.

Fixed

• Images are no longer dropped. Unsupported image data now fails with a clear message instead of silently sending text only.

Improved

• Model lineup in the README, so you can see the Free, Lite, Basic, and Pro / Pro+ options at a glance.

Fixed

• No more repeated tool calls. The model now remembers when it has already run a tool and seen the result, instead of repeating the same action.

New

• Streamed replies. Responses appear word by word in VS Code Chat.
• Tool support. VS Code tools work with Neuron models.
• Usage dashboard. A sidebar shows your token usage (Today, 7 days, and All time), progress bars, and a breakdown by model.
• Friendlier error messages instead of raw technical output.
• "Sign in to use Neuron" prompt when you are not signed in.

New

• Sign in through your browser, securely.
• Neuron models in VS Code Chat. Pick them right from the native model picker.
• Sidebar panel for signing in, signing out, and checking status.