Changelog
What’s new in Superbrain
Release notes for the beta. New builds reach testers automatically through in-app updates, no reinstall needed.
November 2026
- v0.7.7
- Added custom agents: a Markdown file in .superbrain/agents, ~/.superbrain/agents, .claude/agents or ~/.claude/agents adds a sub-agent the agent starts by name, or that starts only when you name it.
- Added custom agents that change files inside the workspace by default, with their edits in the chat's Keep and Undo; readonly: true, or tools that name only read tools, keeps one read-only.
- Added agents that start other agents, two levels deep, under one budget per turn.
- Added a row for each sub-agent in the chat: who ran, on which model, with what access, what it is doing or how it ended, and what it took.
- Added Stop All, which ends the chat's running sub-agents and leaves the main agent going.
- Added a receipt under each finished turn: the files it changed, its last test run after its last edit, and its tokens.
- Added approval cards for sub-agents in the chat, under the agent's name, while the other sub-agents keep running.
- Added the agent's own view of Superbrain's configuration, so "make this the team policy" or "where are my personal rules?" finds the right file and format at every scope.
- Added team policy and command patterns such as git commit * to what the agent can change for you, always at a card.
- Added secrets by name: the agent reads .env and other secret files with each value hidden, writes them back with the values kept and never into another file, and uses a key you paste in the chat by its name, so no value reaches the model.
- Added search outside the workspace, following .gitignore.
- Added OpenRouter credit to the endpoint page: the key's remaining credit shows as soon as it is typed, and a wrong key is refused there.
- Changed a turn that runs out of OpenRouter credit to send once more at the length the credit covers, and otherwise to name the limit with a button to the provider's page.
- Changed a custom endpoint's session cost to what the provider says it charged, where it was $0.
- Changed OpenRouter's reasoning to go back to the model that wrote it, so GPT models keep their reasoning between tool steps.
- Changed a file write outside the workspace to ask in Normal, as a command's does.
- Changed a pattern in your own policy files, such as Bash(git checkout *), to cover everyday commands only: what cannot be undone asks unless a rule names the command exactly, and the card says which pattern stood down.
- Changed the agent so a file write can tighten a policy file but never loosen it.
- Changed .superbrain/mcp.json and .mcp.json to protected configuration, which asks before an edit.
- Changed every plan to reset its usage weekly, on Monday at 00:00 UTC, with the plan balance as the only limit on spend.
- Fixed each chat's changed files to hold only its own edits: with two chats open, an edit from one showed in the other, where Reject could undo it.
- Fixed two chats writing one file at once, so neither takes the other's change for its own.
- Fixed Reject of a file the agent created leaving it empty: it now goes to the Trash, also after a window reload, unless you typed in it or kept it.
- Fixed writes from two chats running at once all going to the turn opened last, which broke Undo to here.
- Fixed an escaped quote in a bash or zsh command hiding the command after it, such as an rm -rf src between two echo commands, which ran in every mode as one echo.
- Fixed edits to the approval policy file running without asking in Auto when its path was spelled another way, such as .superbrain//policy.json, or reached through a link or git rm --cached.
- Fixed "Fix the bug. Do not edit the tests." being read as a request not to edit at all.
- Fixed a turn cut off by a reload reopening as "Stopped after 0s".
- Fixed a skipped step reading "Ran" after its row scrolled out of view and back.
- Fixed old secrets in git history, such as git log -p or git show of a deleted .env, reaching the model.
- Fixed example env files such as .env.local.example, and a Python virtualenv named .env, being treated as secrets files.
- Fixed a wait for OpenRouter credit held by running requests reading "Lost connection to the model".
- Fixed agents in .claude/agents being listed to the model twice, once with a tool it does not have.
- Fixed the weekly reset date naming the Sunday west of UTC; it shows the Monday the reset happens on.
- Fixed a request retried during a server update refusing the account for up to 10 minutes as if its balance were held.
- Fixed chat setup in the main window failing to sign in.
October 2026
- v0.7.6
- Added Auto mode: switch with Shift+Tab in the chat input or the mode button under it, and the agent runs commands and edits without asking while still asking before anything it cannot undo: 1.4 asks per 100 commands (3.2 in cmd) against 31.7 in Normal, on real agent commands in our tests.
- Added the approval mode to the chat footer, kept per chat across other chats and restarts, with a new chat starting in the mode you last worked in (superbrain.approvals.defaultMode).
- Added fact checks before a command runs in Auto: whether files are committed and unchanged, build output, protected branches, where links and junctions lead, and which processes the agent started decide whether it asks.
- Added script reading in Auto: a Python, JavaScript or TypeScript script that deletes, pushes, deploys or touches secrets asks and names its file and line, with imports followed three deep.
- Added asking in Auto before a program that is not a known development tool; allowing it twice, or a policy, lets it run.
- Added clearing the agent's own work without asking in Auto: files and folders its commands or file tools made in the chat, mktemp folders, build output and empty folders.
- Added a patch read before git apply or patch -i runs, so it asks only when the patch changes config such as .git or .vscode.
- Added approval policy files: ~/.superbrain/policy.json, .superbrain/policy.json, .superbrain/policy.local.json and a managed file, using Claude Code's Bash(...) rules, with completion and errors in the editor.
- Added an organization approval policy and an Auto switch, per plan and per account, applied above every other policy file.
- Added Show Approval Policy, Explain Approval, Allow a Recent Command Here, Review Project Approval Policy, Export Approval Audit Log and Forget Learned Approvals.
- Added Allow for this chat to the command card in Auto.
- Added learning from approvals: a command you allow twice in a workspace stops asking there, and a push to a protected branch offers "Allow, and stop asking here" after three; deletes, force pushes, history rewrites, deploys and secrets are never learned (superbrain.approvals.learning.enabled).
- Added a "Learned N commands for this workspace" row in the chat that opens to the commands and the policy that keeps them.
- Added changing approvals by asking the agent, such as "stop asking for npm run build", "always ask before git push in this chat" or "never run npm publish": stricter changes apply at once, a looser one shows a card unless you named the command, and the agent recorded 46 of 48 such requests in our tests against 7 of 24 without it (superbrain.approvals.agentChanges.enabled).
- Added rules for all your projects when you ask, such as "stop asking for npm test in all my projects".
- Added agentChanges to policy files, so a team can require a card for every looser change or allow only stricter ones.
- Added a hover on each command row Auto ran without asking that says why it ran.
- Added asking before git commit or git push while edits wait for Keep or Reject, with Keep and commit or Skip.
- Added Claude Haiku 5.5 on your own Anthropic key, with a 1M context window and 128K output.
- Added date ranges to web search, for results published after or before a date.
- Changed web search to be free on every plan, Forge included, and never counted against your usage.
- Changed web search to drop the limits of 10 searches per chat, 10 results per search and 20 domains per search.
- Changed Max to send each model's highest reasoning level on your own key: max for Claude and GPT-5.6, where Claude got xhigh before.
- Changed output limits to each model's own: Claude 4.6 and later answer up to 128K tokens, from 32K (Sonnet 4.6 from 16K), Haiku 4.5 up to 64K, from 8K, and DeepSeek V4 Pro up to 384K, from 64K.
- Changed MiMo models to the V2.6 line after Xiaomi retired V2.5: Pro, Flash and Pro UltraSpeed, each with a 1M context window, 131K output and image input.
- Changed Skip so a skipped command's files stay declined for the chat in both modes, and three skips in a row pause Auto.
- Changed Skip to tell the agent not to reach the same result another way, such as a script or a file tool.
- Changed sub-agents to follow their chat's approval mode, and to report a command that needs approval instead of opening a dialog.
- Changed MCP tools that do not mark themselves read-only or safe to ask before running in Auto, as the MCP specification says.
- Changed turning approvals off to still refuse what your organization's policy refuses.
- Changed Normal to ask before any write outside the workspace, whatever an Allow always covers.
- Changed Allow always choices to be listed in Show Approval Policy and Forget Learned Approvals, where you or the agent can remove them.
- Changed a policy file that stops parsing to keep its last good version in effect, and to say so.
- Changed your own policy file's asks and refusals to hold over an approved project file's allows.
- Changed a .superbrain/policy.local.json committed to the repository to be read as the project's file, which only tightens until you review it.
- Changed web fetch to follow the approval policy's network allow and deny lists, and to ask before a production host.
- Changed commands that print every environment variable, such as printenv, set, export -p and Get-ChildItem env:, to never be remembered by Allow always.
- Changed the agent to stop confirming in chat what the approval card is about to ask, which asked twice before every irreversible action.
- Changed a stopped turn to keep its partial answer, read "Stopped after", and tell the next turn you stopped it.
- Changed command rows in bash and zsh to drop a leading cd into the workspace, so each row shows the command it runs.
- Changed tool results over their size limit to keep their beginning, or both ends for command output, never only the end.
- Changed the chat footer to show the approval mode where the disclaimer was, and its Settings link to open Superbrain's settings from the keyboard at 7.44:1 contrast, from below 4.5:1.
- Changed usage wording to included usage and extra usage, with a chat answered by Flux shown as Flux usage in the Context Panel.
- Changed the app icon to a 3D-rendered mark, and the empty editor's watermark to a letterpress relief that sits on any theme.
- Fixed the chat not following a streaming answer after the first one in a chat: the newest line fell up to 3,822px below the view in our tests, now it stays in view on every turn.
- Fixed the jump-to-bottom button never appearing: it now shows centred above the input when you scroll away, with a spinner while the agent works and the arrow on hover.
- Fixed the pinned question at the top of the chat covering the question it repeats, and repeating it again in a tooltip.
- Fixed answers not streaming while the agent had a plan.
- Fixed Kimi K3 dropping its reasoning after every tool step, on Kimi and on Bedrock: 5 of 10 SWE-bench Lite tasks resolved in our tests, from 2, with 42% fewer requests and 19% lower cost.
- Fixed Claude and OpenAI models losing their reasoning between tool steps in sub-agents and with preserved thinking off.
- Fixed every later request failing with "Invalid signature in thinking block" after a chat switched to Claude.
- Fixed OpenAI models reading a preamble as the final answer.
- Fixed sub-agent reports reaching the agent cut to their last 8,192 characters; reports of 15K to 23K characters now arrive whole.
- Fixed code search stopping for the rest of a session after 30 minutes idle; it restarts on the next search.
- Fixed chat summaries on Haiku 4.5 at Thorough being cut to 1,024 tokens; they now keep 10K.
- Fixed a replace-all across one long line of minified code freezing the window: 3.2 seconds for 100K matches in 4 MB, now 7 ms.
- Fixed the model pill flipping while typing in a chat panel narrower than about 330px: 1,316 flips across 55 widths in our tests, now none.
- Fixed a plan change, such as web search turning on, reaching Forge only after a restart; every turn now refreshes the plan.
- Fixed later agent edits to a file you had kept or rejected skipping Keep and Reject.
- Fixed a restored chat writing its older copy of a file over newer work, or offering to undo changes made after its edit.
- Fixed the delete tool removing a whole folder such as src without asking; it asks unless the folder is rebuildable output or the agent's own.
- Fixed edits to git hooks, editor tasks, .npmrc, .envrc, .mcp.json and approval policy files landing without asking through the edit tools; they ask in both modes.
- Fixed a leftover chat.tools.global.autoApprove, Bypass Approvals or Autopilot setting approving every tool without a card; only a card's own buttons approve.
- Fixed git options that choose another program or repository, such as --exec-path, --git-dir and --work-tree, running without asking.
- Fixed commands hidden in brace expansion, such as {rm,-rf,src}, being offered Allow always as an unknown program.
- Fixed sed scripts that run commands or write files, such as sed 'e rm -rf src', running under Allow always for sed.
- Fixed a carriage return hiding a second command, such as echo ok followed by rm -rf src, running as one echo.
- Fixed mv, ln, install, move, ren, robocopy and Move-Item not being judged where they write, so mv x.sh .git/hooks/pre-commit went unseen.
- Fixed writes through path forms Windows rewrites, such as a trailing dot, an 8.3 short name or an alternate data stream, being remembered as ordinary writes.
- Fixed commands run through another program, such as timeout, caffeinate, xcrun, flock, rg --pre, fd -x, tar -I, rsync, cmake -E, conda run and lerna exec, being judged as that program instead of the command it runs.
- Fixed reads of stored logins, keychains, shell histories and whole secrets folders such as ~/.ssh and ~/.aws running without asking.
- Fixed the git tool running a filter that a repository's own config names while comparing working files.
- Fixed a quoted regex alternation, such as grep "a\|b", asking every time in Git Bash.
- Fixed .env.example being treated as a secrets file.
- Fixed the workspace trust prompt opening hidden under the sign-in screen and taking keyboard focus; it now waits for sign-in, and the folder stays Restricted until you answer.
- Fixed keyboard focus and menu commands reaching the window behind the sign-in screen; Help > Report Issue, reload and close still work there.
- Fixed the chat panel staying at 170px after the window had been narrow.
- Fixed chat header actions clipping instead of folding into the ... menu.
- Fixed the window title keeping "Settings" after Settings closed.
- Fixed long commands being clipped on the approval card; the command wraps, and Expand shows every line.
- Fixed "N files changed" being clipped in a narrow panel; Keep All and Reject become icon buttons there.
- Fixed a code block's toolbar being cut off at its top edge.
- Fixed the context orb not opening with Enter or Space.
- Fixed the signed-out avatar sitting off-centre.
- Fixed custom endpoint form messages showing only while the field had focus.
- v0.7.5
- Added Agent Skills: a SKILL.md folder in .superbrain/skills, ~/.superbrain/skills or a parent folder up to the repo root, loaded when it fits the task or when you type /name.
- Added the skills you keep for other agents, from .agents, .claude, .cursor, .github and ~/.copilot (superbrain.skills.compatibility).
- Added Claude Code's SKILL.md fields: when_to_use, paths, argument-hint, effort, user-invocable and disable-model-invocation, with $ARGUMENTS, $1 and ${CLAUDE_SKILL_DIR}.
- Added allowed-tools, which pre-approves a skill's commands for its turn, never git push, publishing, deletes, file writes or download-and-run.
- Added disallowed-tools, which blocks a command however it is spelled (git -C . push, /usr/bin/git push), in sub-agents too.
- Added context: fork, which runs a skill in a read-only sub-agent and returns only its report.
- Added superbrain.triggers, phrases in SKILL.md that load the skill when your message has them.
- Added six built-in skills: debug-tests, test, launch, init, create-skill and create-rule, with /init, /create-skill and /create-rule.
- Added a "Using skill" row in the chat with the SKILL.md and the commands it pre-approved.
- Added skill policies for teams: superbrain.skills.allow, deny, bundled and allowedTools, lockable by Group Policy or a macOS profile.
- Added SUPERBRAIN.md and SUPERBRAIN.local.md in any folder, else AGENTS.md or CLAUDE.md, applied while the agent works in that folder.
- Added ~/.superbrain/SUPERBRAIN.md for every project, and personal skill writes there that show their content, ask each time and are refused in an untrusted workspace.
- Added Migrate Project to Superbrain for Claude Code, Cursor, Windsurf, Copilot, Gemini CLI, Cline, Roo, Kiro, OpenCode, Junie, Zed and Aider, with preview and Undo: 48 of 48 items in our tests, against 41 by hand, at 30 to 40% lower cost.
- Added web search on Free and Prevue: 10 results per search, 20 allowed or blocked domains, a day, week, month or year filter.
- Added a limit of 10 web searches per chat, sub-agents included, each counted against your plan (superbrain.webSearch.enabled).
- Added custom endpoints on every plan: LiteLLM, Portkey, OpenRouter or any https Chat Completions, Responses or Anthropic Messages API.
- Added live endpoint checks that detect the API and test the address, key, models, tool calls and context window, each failure with its fix.
- Added a switch and a test call per endpoint model, typed model ids, and extra request headers.
- Added reasoning effort for endpoint models, with a picker in the chat: 35% fewer turns and 20% lower cost per task in our tests.
- Added Direct, Cloud and Custom headings to the model picker on Forge.
- Added code search for C, Kotlin, Swift, Dart, PHP, Shell, PowerShell and SQL: 34 of 51 script questions answered in our tests, from none.
- Added code search for Svelte, Vue and Astro components: 14 of 22 component questions answered in our tests, from none.
- Added code search for .mjs, .cjs, .mts and .cts files.
- Changed symbol lookups to return the declaration as written, with parameters, types and return type, 4 to 36 times smaller than the function.
- Changed symbol lookups to return every overload of a name: 393 to 3,113 overload sets per test project answered as one function.
- Changed an edit by a name several overloads share to be refused with the list, never landing on a guess.
- Changed code search indexes to be 62% smaller, 4.2 GB to 1.6 GB across 27 projects in our tests.
- Changed code search to give disk space back: a 0.7.4 index went from 5.0 GB to 1.7 GB in 11 minutes on macOS in our tests.
- Changed code search to hold 47% less memory after a save on a large project, 125 MB to 66 MB.
- Changed updating from 0.7.4 to upgrade code search in place, reading each file once more in the background.
- Changed edits to SUPERBRAIN.md, rules and folder instructions mid-chat to reach the agent without breaking the prompt cache.
- Changed project skills in an untrusted workspace to stay unloaded, with /name saying why, and skill edits to apply without a restart.
- Changed git stash, stash pop, branch <name> and tag <name> to ask first, and branch -f, -M and -C to warn like reset --hard.
- Changed a saved "Allow always" for open to no longer apply.
- Changed custom endpoint models to need at least 64K of context, and the model picker to list every provider, without "+N More".
- Changed issue reports to state the Superbrain version, not the VS Code engine's.
- Fixed code search missing edits to one-line functions and signatures: 174 of 9,183 edited functions in our tests, now none.
- Fixed the agent getting another function's code after a file changed: 367 of 87,564 reads in our tests, now none.
- Fixed edited docstrings never reaching code search, for 5 to 9% of edited Python functions in our tests.
- Fixed code search losing symbols that share a name in one file: up to 18% in Swift, 15% in C# and C++, 11% in Rust and 10% in Kotlin, now none.
- Fixed most C++ functions missing from code search: 3.5 times as many found across 14 C++ projects.
- Fixed caller lookups missing a call in chains such as get_user().save(): 5% more calls found across 27 projects.
- Fixed calls traced into another language's code, 1% of calls in our tests, now none; 11% more calls reach their function.
- Fixed TypeScript imports written as ./x.js not being followed: 0.1% in our tests, now 99.9%.
- Fixed the code gathered before an edit running over its limit: up to 96K tokens, now at most 9.5K and 31% smaller on average.
- Fixed that code leaving out 18% of what it should hold, such as Vue and Svelte components and C++ classes, now 1%.
- Fixed code search answering one question differently on two builds of a project: up to half the questions, now none.
- Fixed one file stopping code search for a whole project (deep nesting, a named pipe, a non-UTF-8 name): 1 of 125 projects.
- Fixed saves of large or unusual files holding up code search: up to 301 seconds in our tests, now at most 16.
- Fixed symbol lookups failing on non-ASCII text such as Chinese comments, which stopped 3 of 4 test runs on such projects.
- Fixed UTF-16 files, the default in Windows PowerShell and SQL Server Management Studio, reading as empty.
- Fixed an update, a rollback or a busy index leaving code search empty for a session or rebuilding it.
- Fixed files deleted while Superbrain was closed staying in code search, and accented macOS names appearing twice.
- Fixed a project in a folder named test, or folders such as attestation/, being read as tests.
- Fixed a change summary added to read-only answers, which cost the next turn its cache in 10 of 14 test sessions.
- Fixed a restored chat switching to its last request's model instead of the one you picked after it.
- Fixed git branch -D, stash drop, tag -d, remote changes, kubectl get secrets, npm config set, gh repo delete and gh pr merge running without asking.
- Fixed git options that start a program (--upload-pack, --exec, rebase -x), and look-alike programs in your project or C:\Windows\Temp, running without asking.
- Fixed the agent writing through a link in your project to files outside it, such as ~/.ssh.
- Fixed a locked or denied keychain deleting a custom endpoint's key; it now says so, with Try again.
- Fixed web fetch returning the end of a page's list when every section matched.
- Fixed Help > Report Issue opening a missing page; it opens Superbrain's tracker with your version filled in.
- Fixed the issue reporter sending drafts to Microsoft; it now finds similar issues on Superbrain's tracker.
- v0.7.4
- Added Amazon Bedrock on Forge: connect your AWS account with an AWS sign-in, a Bedrock API key or AWS_BEARER_TOKEN_BEDROCK, and run the Claude, Kimi and GLM models AWS offers your account.
- Added Claude Sonnet 5.5 and Opus 5.5, each with a 1M-token context window.
- Added new Bedrock models as soon as AWS offers them to your account, with no update needed.
- Added AWS's prices for your region and your AWS quota beside each Bedrock model.
- Added a choice of where AWS runs your Bedrock requests: any region, or only inside one geography such as Europe or the United States, which AWS prices 10% higher.
- Added the regions AWS runs your requests in for each routing choice.
- Added the other models AWS offers that Superbrain doesn't run yet to the Bedrock page.
- Added "Bedrock" in front of the model name in the chat for models running on Bedrock.
- Added one-click fixes when Bedrock refuses a model: request quota, sign in again, or continue the chat on a model that runs.
- Added your Amazon Bedrock connection and usage per model to your account on onesuperbrain.com.
- Added "Conversation compacted" in the chat when Superbrain compacts a long conversation.
- Added /compact to compact a conversation now, and a Compact button in the context panel once context is 80% full.
- Changed Claude Fable 5.1 to need this version, so a long conversation on it keeps working after Superbrain shortens it.
- Changed the chat to tell you when your Bedrock model can't run, such as when Bedrock isn't connected or your AWS sign-in ended, with the fix, instead of switching to another model.
- Changed compaction to save the full earlier conversation for 7 days, so the agent can look back at exact details.
- Changed compaction to reuse the cached conversation, so a summary reads 97 to 99% of its prompt from cache in our tests.
- Changed the context meter to show the token count the model's provider counts; on Claude it read about 40% low.
- Changed GLM and Bedrock requests to follow your proxy settings, including a proxy that asks for a username and password, and your system's proxy and certificates.
- Changed sub-agents to run one at a time while your provider is limiting your requests.
- Changed the model button on Forge to read "Choose a model" instead of "Auto" when no model is chosen, and a message sent then to ask you to pick one.
- Changed Set API Key to list only the providers your plan offers.
- Changed the message for an unexpected failure to offer support instead of a new chat.
- Changed the chat on the Free plan to show Kilo.Flux and Concise, the model and depth your requests run on, instead of an Upgrade button.
- Fixed long tasks stopping when a conversation outgrew the model's context window: Superbrain now shrinks the request and keeps going, and a task that outgrew a 200K window finished in full in our tests.
- Fixed a compacted conversation being compacted again on every later turn.
- Fixed answers stopping when the connection dropped mid-reply: Superbrain resumes where the answer was cut, up to 3 times in a row.
- Fixed a chat failing with "An unknown error occurred" during a short network outage instead of waiting up to 5 minutes for the connection to return.
- Fixed a window freezing while signed out.
- Fixed a dropped connection during compaction losing the conversation summary.
- Fixed a newly opened workspace showing Flux for under a second before your chosen model, which could also replace your choice.
- Fixed the closing summary of a long task paying to cache the whole conversation again; it now reads about 98% of it from cache.
- Fixed the cost shown leaving out replies that dropped part-way; they now count, within about 2% of the provider's figure in our tests.
- Fixed chat replies without tools being left out of the session cost.
- Fixed a stopped turn's cost going missing from the chat's total.
- Fixed a locked or refused keychain deleting your saved API keys.
- v0.7.3
- Added reading of PDF, Word (.docx, .doc), Excel (.xlsx, .xls), PowerPoint (.pptx, .ppt) and OpenDocument (.odt, .ods, .odp) files, both when the agent opens them and when you attach them.
- Added text recognition for scanned PDFs and images (.png, .jpg, .gif, .bmp, .webp, .tif, including multi-page TIFFs), using your system's own OCR in the languages you have installed (tesseract on Linux).
- Added a map of pages, headings, slides and sheets for long documents, so the agent reads only the part it needs.
- Added search inside PDF and Office documents.
- Added images by path: a model that can see looks at a screenshot or chart in your project, and one that cannot reads its text.
- Added showing the model images that tools return, such as an MCP server's screenshot, which were dropped before.
- Added text recognition for images attached to a model that cannot see, in place of the "does not support images" notice.
- Added a scratch folder for each chat, outside your project, where the agent keeps its own scripts and data.
- Added a built-in runtime, so the agent can run scripts on a machine with neither Node.js nor Python installed.
- Added letting the Debug agent write and run its own test scripts in that folder while leaving your project untouched.
- Added a setting for the shell the agent uses on Windows, macOS and Linux.
- Added a "Waiting for <model>…" line when a model has not started answering after 20 seconds.
- Added a clear message for password-protected Office files.
- Changed the agent's shell on Windows to PowerShell 7 when installed, then Git Bash, then Windows PowerShell, which replaces Command Prompt for users without Git.
- Changed coding tasks to cost 40 to 52% less, error-handling changes 47 to 57% less and multi-turn sessions about 83% less, with the same or better results in our graded tests.
- Changed questions about documents and images to be answered correctly 26 of 26 times in our tests, up from 16 and 6 on the two models we test with, in about 6 seconds instead of 39.
- Changed a question about an attached 60-page PDF to cost $0.013 instead of $0.48, answered in 4 seconds instead of 45.
- Changed a question about an Excel workbook to take 4 steps instead of 14, and one about a PDF 2 instead of 5.
- Changed new chats to reuse the cached start of the conversation, cutting the cost of a first question by about two thirds.
- Changed attached files to reach the model as their content, so it no longer opens them again.
- Changed edits to return faster, most of all on files no language checker covers (a Python edit from 0.8 to 0.1 seconds).
- Changed edits to report their own errors, cutting separate error checks in a coding session from about 5 to 1.
- Changed reads of very large files to load only the part asked for: a 128 MB log reads 7 times faster with 20 times less memory, and a file over 500 MB can now be read.
- Changed large CSV and TSV files to open as a 50-row preview with their row count.
- Changed UTF-16 text files, such as Windows PowerShell output, to read as text.
- Changed long tasks to run until they are done or stop making progress, instead of ending at a fixed step limit.
- Changed the command time limit to 2 minutes by default and 10 at most, up from 60 seconds and 5 minutes, so installs and builds can finish.
- Changed long command output to keep its beginning and end, up to 30,000 characters, instead of only the last 8 KB.
- Changed the agent to suggest only the test commands your project declares, so it runs the right test runner first.
- Changed tool calls that change files or run commands to run one at a time and in order, while reads still run together.
- Changed the agent to rewrite a whole file once it has read it, where before it could only patch an existing file.
- Changed read-only git commands written with -C or --no-pager to run without asking.
- Changed git config changes to ask first, and a git setting that can run a program to ask every time.
- Fixed a reply waiting about 30 seconds to start when the code search model was not ready; it now starts in about 7.
- Fixed a stalled reply holding the chat forever; it now stops after 10 minutes with no data.
- Fixed a failing provider delaying Flux's switch to another one.
- Fixed Stop not ending a running command; it now stops the command and everything it started.
- Fixed a command that starts a server in the background holding the turn until its time limit.
- Fixed a command meant for another folder running in the project folder.
- Fixed a file meant for outside the project being written inside it.
- Fixed the Explore agent writing report files into your project and its git status.
- Fixed the agent saying a file was created when the write had been refused; it now says which changes were not made.
- Fixed several ways a command could run more than the permission prompt showed, including here-documents, zsh process substitution, eval, find -exec, xargs and git options.
- Fixed an allowed command asking again when it redirected its error output with 2>&1.
- Fixed the agent claiming tests pass after an edit that no test had run against.
- Fixed the agent re-running tests on a turn that changed no code.
- Fixed a passing node --test run not counting as a passing test.
- Fixed the agent being talked out of throwing an error the task asked for.
- Fixed edits to files read together in one batch being refused as unread.
- Fixed a workspace override of the Debug agent turning its thinking off.
- Fixed attached images over 5 MB being refused instead of resized, and very tall screenshots being rejected by the provider.
- Fixed an attached BMP image failing the request on every retry.
- Fixed Gemini models mixing up the saved output of earlier commands in a long session.
- v0.7.2
- Changed first-run setup to two screens: sign in, then a summary of the plan your account is on.
- Added links to the Terms of Service and Privacy Policy on the sign-in screen.
- Changed "Add your API key" in usage warnings and the account menu to appear only on plans that use your own key.
- Fixed setup telling every account it had free credits and needed no API key, including on Forge.
- Fixed the guided first task not starting when you open a folder during setup.
- Fixed "Choose a model" doing nothing on Forge before an API key is added, along with the model picker shortcut and /models.
- Fixed the model picker opening next to a different chat than the one you clicked in.
- Fixed a model chosen from the … menu of a narrow chat not being applied.
- Fixed Set API Key saving a MiMo Token Plan key that then failed every request.
Loading earlier releases…
Showing 2 of 6 months of releases.
Superbrain is in beta · 45 releases since June 2026