Inferpal 1.8.0

With LM Studio, Qwen3.8 now spends its reasoning where you are waiting for it: a /check review comes back two to three and a half times sooner, and the small tasks Inferpal runs for you — titles, commit messages, summaries — no longer wait for it to think. /check reviews your whole change again, and a repository's MCP server is asked about again when the code it runs changes.

Reasoning models (LM Studio)

• A /check review on Qwen3.8 comes back two to three and a half times sooner. A question — /check included, with tools or without — now asks for a low reasoning effort: on Qwen3.8 27B, /check found the same issues in 38 seconds instead of 136 in English, and in 31 instead of 69 in French, and tool tasks pass as before. The level is sent only to a model LM Studio says takes it, so Gemma 4 and GLM 4.7 Flash are unchanged.
• gpt-oss is now always asked for low, the level LM Studio gives it by default: if you raised it in LM Studio, set it in Inferpal instead.
• Session titles, commit messages and summaries no longer wait for the model to think. They ask for no reasoning when LM Studio says the model can turn it off — Qwen3.8 27B then writes about 84 % fewer tokens for a summary.
• You choose the level. Settings → Server and models → Show advanced settings → Sampling and reasoning: one level for a question, one for an agent run (Server's, Low, Medium, High). Agent runs keep your server's setting by default: their time goes to the tools they run, not to reasoning.

/check

• /check reviews the whole change, with your checks on top. In a repository with checks, it reviewed the diff only against them and answered "all checks pass" on changes that broke something none of them named — the more checks you had written, the narrower the review. It now looks for any bug, regression or behaviour change, applies your checks as added criteria and says which of them pass. /check <name> still reviews against that one check only.

Security

• A repository's MCP server is asked about again when the code it runs changes, not only its definition. Your agreement to start a server your repository declares (.mcp.json, .vscode/mcp.json…) covered its command line: a script rewritten by a git pull — submodules included — was started without asking. The agreement now also covers the repository's files the server runs, checked again just before it starts and before it is restarted. In a git repository, what it writes beside itself is not part of it — but a file its definition names always is, even one git does not track, so a database named in its arguments asks again when it changes. A link that leads out of the repository is never followed. A server whose script you had agreed to is asked about once more after this update.
• MessagePack 2.5.303 in the VS Code host, for an advisory (GHSA-qhrr-8q5h-9q3h) on deserializing untrusted data into object. The host speaks JSON to VS Code and to ACP clients and never reads MessagePack from outside, so the affected path was not reachable. Visual Studio provides its own copy.

Inferpal 1.7.10

Nineteen fixes. The first closes a way around your rules on Windows: a file they protect could be written under another spelling of its name.

Your files and your rules

• Windows: another spelling of a file is that file, for your rules too. To Windows, .env::$DATA and memory.md. are .env and memory.md themselves: written that way, a file a deny rule protects — or one of the files the agent reads back as its instructions — was written without the rule or the forced approval applying. Paths are now read as the file they write, and a hidden stream of a file (Cart.cs:hidden) is refused.
• Linux: two files whose names differ only by case are no longer taken for one. The impact analysis could leave a real dependant out, /check filed one file's findings under the other's name, Fix with AI mixed their errors, and opening a solution in a folder whose name differed only by case did not move Inferpal to it.
• Linux: renaming a file to change only its case keeps it in the semantic index. Widget.cs renamed widget.cs disappeared from the codebase search until the next full indexing.
• The agent's build check builds your solution, not a side project's. With a .slnx at the root and an older .sln in a sub-folder, the build check — and /tdd's test run — used the sub-folder's and reported its result as yours.

Stopping and waiting

• VS Code: Stop works from the moment the button shows. Pressed while /explain or /review was still sizing the code, an @file was being read or a template expanded, the question went anyway. The turn now ends there, with nothing sent.
• LM Studio: stopping an answer no longer hides the model's real context from the next questions. A model loaded with less context than your setting then had its next questions refused instead of compacted to fit.
• Stopping a /task propose keeps what it proposed. Its changes are counted in the notice and listed by /task <id>; and /task apply on a task still running says it has not finished.
• Unloading a model no longer breaks an answer in your other editor. With Visual Studio and VS Code (or Zed) open, Unload in the settings pulled the model from under the other one's answer. It now waits for any answer on the machine to end.

Visual Studio

• "Edit with AI" shows the change before applying it, when you asked for that — as Fix, Refactor and Add docs already did.
• After a crash, the new window no longer inherits the old one's state. A crashed Visual Studio leaves small state files behind, and the next one could take its debugger position, its "answering" marker, a pending preview or its open solution for its own.

VS Code

• The files the model is told are open are your tabs — not every document VS Code had loaded, and no longer missing an open file too large to mirror.
• The @ picker no longer offers Inferpal's backup copies or your Python environment, and the files it lists first are your open tabs.

Zed and other ACP clients

• An /arena vote goes to the duel that thread showed, not to a duel started in another thread.
• A reloaded thread keeps the part of a failed answer you had read, marked as a notice, as in Visual Studio and VS Code.

Everywhere

• Saved conversations are listed newest first in the session pickers, /sessions and /history.
• Dates written to files are the same on every machine — on a Windows set to the Thai calendar, notes were dated 2569; what Inferpal shows you keeps your settings.
• /snippets copy and /snippets delete say when your snippet library cannot be read, instead of "no snippet #3".
• A repository's Continue MCP server connects with its headers and key, and a server declaring the older SSE transport says so when it fails.
• An MCP server declared with WebSocket says why it does not run, and /import leaves it out with the reason instead of adding a server that could only fail.

Inferpal 1.7.9

Fifty-two fixes, and the first of them are about what you approve: permission rules now do what they say, an approval no longer writes over what you saved meanwhile or lands in another file, and an "Always" can no longer pass to a server you never approved.

What you approve

• Permission rules do what their comments say. The example rules in the documentation end with a comment (deny * \.env$ # never touch secrets), and copied as shown they matched nothing — .env could be written, rm -rf was not blocked. A # after a space now starts a comment.
• A command chained after an allowed one is no longer approved with it. A rule that auto-approves git also approved git status; Remove-Item -Recurse src, past the rule that denied Remove-Item. A command line is now approved only when every command on it is.
• What you save while an approval waits is no longer written over. If you saved the file — or VS Code's auto-save did — while the agent's change waited for your answer, approving wrote the older text back over your work. The change is now refused, your file is kept, and the agent reads it again.
• Text inserted by the agent goes to the file you approved, not to whichever file is active when you click.
• An "Always" given to one of your MCP tools can no longer pass to a repository's server whose name differs from yours only by punctuation (my.server next to your my-server). Such a server is not started, and the MCP status says which of yours it collides with.
• A file your permission rules protect no longer makes the agent give up the tool: it is told this path or command is refused, and carries on with its other edits.
• Visual Studio: a project's own build check is asked about again in another project — the command a repository declares in .inferpal/validators.json no longer runs unasked in the next solution you open.
• VS Code: Enter no longer approves a question you did not read. An approval asked outside an answer is now a dialog, as in Visual Studio, and Enter or Escape answers a card only when it is the only one waiting.
• VS Code: a link in an answer opens only if it is a web address — a vscode://… or ms-settings: link written by the model no longer launches a program.

Fixed, in every editor

• A settings file broken by a hand edit no longer costs you your settings. One comma too many in config.json, and the next setting you changed wrote the factory values back — model server, MCP servers and permission rules gone. Your settings are kept.
• Edit with AI, /fix, /refactor and /doc keep the nesting of files indented with two spaces or with tabs — a YAML key no longer moves out of its parent, a Python body no longer leaves its block.
• A failed build is no longer reported as a clean one, and a build that could not start (no project in the folder) is no longer counted as a compilation error.
• pytest and cargo runs that failed after their tests passed are red (pytest-cov's minimum coverage, a test binary that crashes later), and a pytest run longer than a minute no longer reads "nothing was proven".
• "Tests passed" only when tests ran, and a failing test run is no longer forgotten because a later check matched no test.
• /tdd runs the tests of a solution that is not at the root of your folder — with the solution in src/, every round used to say that no test ran.
• Windows: pytest's messages keep their accents, and a symbol outside the Windows code page no longer makes the run fail.
• A long file the agent writes no longer ends the run as raw text: a write cut at the answer's length limit is recognised, and the agent splits the work.
• A tool call with a parameter left unclosed is no longer run without its arguments — the whole test suite instead of the one test asked for.
• One broken tool call no longer takes the others down with it: the files the agent correctly asked to read are read.
• A reasoning model's summary or inline edit no longer vanishes because its reasoning mentioned a tool.
• The agent's closing answer fits a long conversation instead of overflowing the model and ending on a bare list of tools.
• A background task no longer mixes into the answer you are waiting for, and a failing command run in the background reports its failure instead of "exit code 0".
• Changing the embedding model no longer mixes two models in one index, which made semantic search return noise under a green status.
• The codebase search says what the index leaves out — files too large, a folder that could not be read — and a file busy during a rebuild (a build writing it) keeps its place in the index.
• Renaming and impact analysis follow edits to folders kept out of search (indexExclude), and the impact analysis no longer answers about another file of the same name.
• A language server that stops answering no longer slows the indexing down, ten seconds a file, in silence: it is set aside and the indexing goes on.
• @code and @tree no longer attach "nothing found": they say why nothing was attached.
• What you type after /explain, /review, /fix, /refactor, /doc and /test reaches the model — /fix the off-by-one in the loop fixes that.
• /debug stop the crash when the list is empty investigates — it no longer stops your debugging session. And /template name more words no longer clears the conversation.
• A conversation you clear is saved before it leaves — closing the editor right after /clear no longer loses it.
• /undo-run can be tried again when a file could not be given back (open in another program, say), and tries those files only.
• Signing in to an MCP server through Microsoft Entra, Keycloak, Okta and similar providers works, and asks only for the permissions the server needs.
• You can sign in to an MCP server your repository declares, or that Zed passes: it has its own card with a Sign in button.
• An MCP sign-in survives a provider that is briefly down, and with Visual Studio and VS Code both open, one no longer asks you to sign in again after the other renewed it.
• Signing in to one MCP server no longer signs you out of the others when the file holding your sign-ins could not be read at that moment.
• When an MCP tool fails, the agent is told what to do about it — no longer advice that contradicts the cause.

Fixed, in Visual Studio

• /fix-build builds the solution you have open — not the first one in the folder, which then reported the build as fixed.
• Opening another solution no longer overwrites that solution's last conversation, and /clear empties the right one.
• A background task runs in the solution you gave it, even when another solution is open by the time it starts.
• After you close the file you were working in, the agent no longer edits an old copy of another one.
• Tables in answers follow the theme — readable again after switching between dark and light.
• @file attaches the file as you see it, unsaved changes included.
• An accepted inline suggestion keeps your file's line endings.

Fixed, in VS Code

• Signing in to an MCP server works from a remote window (SSH, WSL, a container): the page opens in your browser, and wherever no browser can be opened, the sign-in says so at once instead of waiting five minutes.
• A question sent while settings are being applied is told so — and the approval questions waiting are no longer answered "no" for you.
• Inferpal stays in your editor's language, instead of switching to Windows' language when the MCP servers start.

Fixed, in Zed and the other agent-protocol editors

• A thread coming back after a pause takes every request sent meanwhile — no more "Unknown session", and a thread closed meanwhile stays closed.
• A /task report finishing right after an answer stays in the saved thread.
• Inferpal stays in your editor's language in every thread.
• The settings window's @docs page follows a reset of the window.

Inferpal 1.7.8

Inferpal's documentation is now one click away from the chat — and forty-nine fixes, many of them for Zed, the JetBrains IDEs, Neovim and Emacs.

The documentation, from the chat

• /help ends with a link to the full documentation, in Visual Studio, VS Code and the agent-protocol editors, and the chat has a ? button that opens it — in Visual Studio's chat header, and beside the settings gear in VS Code (also Inferpal: Documentation in the command palette). It opens the documentation of the version you have installed.

Fixed, in every editor

• A result's Undo button undoes that result. After a /fix-build, /tdd or /task apply, the previous answer's Undo reverted their files and left its own; it now undoes its own changes, and says so when a later change must be undone first.
• When an answer fails halfway, what was already written stays — as a note. Visual Studio saved the half answer as the model's reply, which reopening the conversation handed back to the model; VS Code lost it at the next redraw.
• Signing in to an MCP server no longer restarts the others, so a tool call the agent was making to another server at that moment no longer fails.
• An MCP server's working directory ("cwd") is used, and editing a server in Visual Studio's list no longer removes it.
• The model lists no longer offer embedding models for the chat, in every settings page and model picker — picked by mistake, one answered no question at all. /model refuses one too.
• /test says why it could not write the tests — the request too large for the model's context, the server unreachable — instead of "check the model and try again".
• What /import applies takes effect at once: imported MCP servers start, an imported language applies, without restarting the editor.
• /import fixes: unnamed MCP servers in Continue's older config.json are all imported, each under its package's name; importing the same file twice no longer adds its rules again or offers a renamed command again; a relative path is found in the folder you have open; and an imported documentation site no longer replaces one of yours.

Fixed, in Visual Studio

• Links in the chat can be followed: underlined, their address shown on hover, opened in your browser on click (web addresses only).
• A large attached file no longer overflows the model. The active file, or a picked or @-mentioned one, was sent whole; it is now attached up to 60,000 characters, as in VS Code, and the chip says how much.
• The chat's model button follows the settings: a model chosen in the settings window or by /import no longer leaves the old name on the button.
• Regenerate says "restart" right after you switch model servers, instead of saying the old server cannot be reached.
• The first-launch message counts its models in your language.

Fixed, in VS Code

• Regenerate on an /explain or /review answer explains the same code again — it came back as an agent turn with tools, the whole code pasted into the question.
• A message sent while Inferpal starts or restarts is never lost: the chat says it waits, and a refused one goes back in the box. A button that sends for you no longer overwrites the draft in the box.
• When the assistant restarts during an answer, the chat says so in your language, instead of a raw English library message.
• Inferpal's own notices show when they come, not only after the chat is redrawn.
• With no folder open, the status badge keeps saying a folder is needed, instead of switching to "Backend unreachable".
• A file you closed is no longer taken for the one you are working on.
• A documentation crawl says how it ended — the pages indexed, or why there are none — when it ends. Also in Zed and the other agent-protocol editors.

Fixed, in Zed and the other agent-protocol editors

• Your other editors are left alone. Starting a new conversation in a thread no longer empties the one VS Code would restore on the same folder, and /model in a thread changes that thread only — not Visual Studio's, VS Code's and every other thread's model.
• In Agent mode, the answer no longer shows twice, and the model selector names the model that answers — a model you pick answers in every mode.
• Threads keep working through settings changes: a settings file broken by a hand edit no longer resets them to factory settings, switching to another kind of model server no longer breaks them, and a setting a thread saved no longer stops it following the settings window.
• A thread put aside and reopened comes back as you left it — its mode, its model, a file attached for the next question — and no longer answers "Unknown session" while it is being put aside. Stop works while it reopens, a thread at work (a /task, a background command, a crawl) is never the one put aside, and a closed thread stops indexing your folder.
• A damaged conversation no longer opens as an empty one that the next message would save over.
• A repository's MCP servers can be agreed to, and start: the agreement is asked at the thread's first message.
• A refused tool call is no longer hidden behind another one in a batch.
• A /task's report stays in the thread, reloaded or reopened.
• Models installed after a thread opened can be picked in it.
• An attached file is cut for the model that reads it, not for the code-actions model.
• Branches and Visual Studio's archived conversations are listed, and a branch saved again keeps its link to the conversation it came from.

Fixed, in the browser settings window

• It no longer stops under a tab still open, and it no longer indexes the folder it was opened in — it shows the saved index, says when it is reading it, and builds one only on Rebuild.
• It opens only the files its form lists.
• It follows a new kind of model server saved in it — its model lists and loaded-models card no longer come back empty.

Inferpal 1.7.7

Visual Studio 2022 joins Visual Studio 2026, moving from Continue or Roo Code takes one command, and Zed, the JetBrains IDEs, Neovim and Emacs get Inferpal's settings window — in your browser. Plus sixteen fixes.

Visual Studio 2022

• The same extension now installs and runs in Visual Studio 2022 (17.14 and later), not only in Visual Studio 2026: the chat and the agent with its approvals, inline suggestions, the change preview, /debug and the build banner.

Coming from Continue or Roo Code: /import

• /import continue reads your Continue configuration, /import roo the file Roo Code's settings export. A card lists, item by item, what Inferpal can take — the model server and the models by role, the context window, rules and system messages, prompts and custom modes as commands, MCP servers, documentation sites, allowed and denied commands as approval rules — and, folded underneath, everything else in the file with the reason it stays. A hosted provider is never imported, the commands Roo let run without asking are offered unticked, and nothing is written until you apply; your settings file is copied aside first. In Visual Studio, in VS Code, and in the agent-protocol editors.
• Commands for every workspace. A prompts folder beside Inferpal's settings holds commands available in every workspace — it is where /import puts them. A project's own command of the same name still wins.

Settings in Zed, the JetBrains IDEs, Neovim and Emacs

• The settings window, in your browser. The same seven pages as in Visual Studio and VS Code, with the Test button, the models loaded right now and the model suggestions. Your editor offers Open Inferpal's settings the first time; after that, type /settings in any thread. What you save applies from your next question, in the threads already open too; a model you picked in a thread stays that thread's. The page is served on your machine only, under an address nobody else can guess, and reaches the settings and nothing else.
• /settings also opens the settings from the chat in Visual Studio and VS Code.
• The sound at the end of a long run can be turned off in Visual Studio: Settings › Language and appearance. It stays on by default.

Fixed, in every editor

• Approving a change to several files names every file. The card said "Change a file" and named only the first one; it now counts them and names each. A rename is asked in your language, and says when some files could not be scanned.
• The approval card colours the change again when the agent edits several places at once, renames a symbol, deletes or restores a file — removed lines in red, added lines in green.
• /undo-run can still undo an agent's edit after many more questions. Fifteen questions after an edit, it said there was nothing to undo. And with Visual Studio and VS Code open on the same folder, one editor no longer deletes the backup the other needs.
• Visual Studio and VS Code on the same folder each get their own last conversation back. Closing Visual Studio and asking one question in VS Code lost Visual Studio's conversation; starting a new one in VS Code emptied Visual Studio's.
• A documentation source added in one window no longer replaces one added in another — it used to vanish from the settings, its pages from the search, without a word.
• Background tasks no longer load a second model. Naming a session or summarizing older turns used the chat model even in Agent mode — loading it next to the agent's model, often too slowly to finish. They now use the model already loaded.
• A context window left at 0 is no longer reported as "0". /hardware and the support bundle now say the server chooses, and name the window it loaded.
• Continue's MCP server files with multi-line values or YAML anchors are read, instead of being refused as unreadable.

Fixed, in Visual Studio

• The "last build failed" banner comes up again, with its button to fix the errors.
• A window no longer brings back another window's conversation — a solution opened for the first time could restore the conversation of a window that had no solution open, and the other way round.

Fixed, in VS Code

• A long Visual Studio conversation is no longer cut when VS Code saves it. VS Code saved only the last 2,000 entries it shows, and the oldest messages were lost — in the file Visual Studio had kept whole.
• A command's reply no longer looks like the model's answer: it shows as a notice at once, not only after a reload.
• The send button is no longer cut off in a narrow side bar.

Fixed, in Zed and the other agent-protocol editors

• Two threads no longer share their model. A model picked in one thread was the one the other threads asked, while their model selector still showed their own.
• Stop works from the first instant. A Stop pressed while Inferpal was still reading attached files was lost: the whole turn ran afterwards, file changes included, under a reply that said it was cancelled.
• An older thread no longer answers "Unknown session". With more than four threads open, the oldest idle one was closed and the next message in it failed; it now reopens with its conversation.

Inferpal 1.7.6

Inferpal now runs in Zed, the JetBrains IDEs, Neovim and Emacs, through the Agent Client Protocol — the same agent, against your own model server. Plus two fixes for every editor.

Inferpal in Zed, the JetBrains IDEs, Neovim and Emacs

• A coding agent for any Agent Client Protocol client. Download the archive for your system below (inferpal-acp-win32-x64, -linux-x64 or -darwin-arm64), extract it anywhere — it carries its own runtime — and point your editor at the executable with --acp: Zed, Rider and the other JetBrains IDEs, Neovim (CodeCompanion, avante.nvim) and Emacs (agent-shell). Setup for each editor: docs/acp.md.
• The same safety as in Visual Studio and VS Code. Every change is shown as its real diff before it is written, and asked for; a file you are editing with unsaved changes is never written over; your permission rules apply.
• Modes, model and commands. The three modes (default, agent, plan) and the model are options of the session, the slash commands are listed, and the MCP servers configured in the editor are started for the session.
• Sessions come back. A conversation is saved after every turn, listed for its folder, and reloaded whole when the editor reopens it — those VS Code saves for the same folder too.
• A setup in the terminal. --acp --setup asks for your model server, checks that it answers and saves the model you pick. With Visual Studio or VS Code already set up, or an Ollama running on this machine, there is nothing to do.
• What stays in Visual Studio and VS Code: inline suggestions, the debugger tools, the X-Ray and the settings pages — the first answer says so.

Fixed, in every editor

• An answer you stop is no longer answered again at your next question. After Stop, the next answer started with the one you had just stopped — a long essay written again before the one-line answer you asked for. The part you saw now stays with its question, marked as stopped.
• The agent's plan no longer shows an edit you refused as done. When you declined a change and the run ended, every remaining step was ticked off, right above the notice that no file was changed. A step that was to write a file and did not now shows as skipped.

Inferpal 1.7.5

Inferpal now reads the setup your repository already wrote for other coding agents — its instructions, its commands, its skills and its MCP servers — and connects to MCP servers of the latest protocol revision. Plus eighteen fixes and two security changes.

Your repository's agent setup, read as it is

• The instructions your repository gives coding agents go with your questions. AGENTS.md, Copilot's instruction files, CLAUDE.md and the rule folders of Claude Code, Cursor, Cline, Roo and Continue — each read the way its own tool reads it: a rule written for some files only when you are on one of them, AGENTS.md from the repository's root down to your file. Inferpal's own .inferpal/ files still win a conflict, and a file too long for the model's window is shortened and says so.
• You can see what each question carries. The Context settings page and the new /instructions command list every instruction file found, what the next question sends of it, and why when it sends nothing. A new setting turns off the files of a tool your team does not want applied.
• The commands written for Copilot, Claude Code or Continue are slash commands in Inferpal (.github/prompts, .claude/commands, .continue/prompts), and the words you type after one go where it asks for them.
• Skills. /skill lists the SKILL.md skills of your repository and your own, and /skill <name> <request> asks with one. A new setting, off by default, lets the model pick the skill a request needs. A skill's scripts only ever run through the usual command approval.
• The MCP servers your repository declares start beside yours — once you agree, even with approvals switched off. The question shows everything that changes what runs — the command, its environment, its folder, its headers — and your answer is kept until the definition changes. A repository file Inferpal cannot read is named, and your own servers keep running.

MCP

• Servers written for the 2026-07-28 revision of the protocol connect, on stdio and over HTTP. They refused Inferpal at once; servers of the earlier revisions keep receiving exactly what they did.
• A tool's structured argument reaches the model with its shape. With Ollama, an argument described as a nested object arrived with no type at all.
• A tool that answers with structured content only no longer reads "(no output)", and a tool that asks for something Inferpal cannot give says so instead of answering nothing.

Safety

• Signing in to an MCP server checks which authorization server answered before the sign-in goes on, and a client registered with one authorization server is no longer offered to another.
• The agent always asks before editing the instructions a repository gives coding agents (AGENTS.md, CLAUDE.md, Copilot's and the other tools' rule files), even with an allow rule, "Always" or security alerts switched off.
• A permission rule removed in one window stops approving in the others. It went on approving commands and edits without asking until those windows were restarted.

Nothing lost

• What you type is no longer lost when the agent edits the same file. In Visual Studio, reloading a file the agent had changed erased your unsaved typing; the agent now refuses to write over unsaved changes, and checks again after you approve.
• Nothing Inferpal started keeps running after a crash — background commands, their shells, the processes an MCP server started — on Windows, Linux and macOS.
• The first index of a large solution no longer starts over when the editor closes: it saves its progress as it goes and picks up where it stopped.
• Two solutions open side by side each get their last conversation back, instead of one overwriting the other's.
• A conversation reopened in the other editor comes back whole, including the messages and the plan card only the other editor shows.

Figures that tell the truth

• A count agrees with what it counts, in every language — no more "1 files"; French, Russian and the other languages each count their own way.
• /task, the run line, /history, Smart Fix and the code added to each question count what they say they count: every step a task took, the files a run read, the messages of a conversation, compilation errors only where the compiler gives them, and each snippet left out for its own reason.
• The project map sees positional records and primary constructors, and so do /doc and the impact analysis — in a modern C# codebase, most records were missing.

Smaller

• A multi-line value in the front matter of a rule, a check or a prompt file is read whole, and the words you type after a custom command are no longer lost.

Full changelog: v1.7.4...v1.7.5

Inferpal 1.7.4

Eighty-two fixes, and most of them are the same story: something failed, or was lost, and nothing said so — a copy that never reached the clipboard, a conversation that came back after you cleared it, a setting one window wrote over another's, a test run that never started read as a failing one.

Safety first

• Visual Studio: an "Always" approval no longer carries over to another solution. Allowing a tool "always" in one solution kept it allowed after opening another in the same Visual Studio, so the commands of a freshly cloned repository ran without asking. Such an approval now ends when the workspace changes.
• Visual Studio: "Open diff" on an approval card says when it cannot open, and that the change is still waiting for your answer — the click did nothing, and the change could be approved without anyone seeing it in full.

Nothing lost without a word

• A conversation you clear stays cleared. When the automatic save could not be emptied (a locked file), the conversation you had just left came back at the next start; both editors now say so. In VS Code, the conversation a /template replaced no longer comes back either.
• A saved conversation that cannot be read is kept, and the editor says so — the next answer's auto-save used to write over it, gone for good.
• Indexed documentation that cannot be loaded says so, in the chat at start and above the /docs list, with /docs reindex to rebuild it — instead of every source showing 0 pages as if never indexed.
• Visual Studio: a copy that failed says so. While another program held the clipboard, the copy buttons did nothing without a sign, and /diagnostics export still said the support bundle was copied — what you pasted was whatever was there before.
• Stopping a question in its first seconds no longer damages the conversation, and in VS Code a question asked while Inferpal starts no longer replaces your previous conversation — it waits for the start to finish.
• Visual Studio: a turn you stopped, or one that failed, is still there after a restart.
• A conversation started with /template keeps its mode when it is reloaded, instead of greeting you in "Code Review mode" and answering as usual.
• VS Code: a restart or a crash of Inferpal says what it stopped — a background task, a command running in the background, a documentation source being indexed — and the breakpoints the assistant set no longer outlive it to stop your next debugging session.
• VS Code: "New conversation" no longer loses the conversation while Inferpal is stopped.
• Visual Studio: saving a snippet into a full library no longer deletes your oldest one; it refuses and says how to make room.

Two windows, one set of settings

• A list changed in one window no longer erases what another window added to it. Visual Studio, VS Code and their windows share one settings file: pinning a file, adding a documentation source, changing an MCP server or an approval rule removed the item the other window had just added — for an MCP server, with its sign-in.
• A value Inferpal fills in by itself no longer overwrites one you set elsewhere — the graphics memory budget detected by /hardware, the model picked at first start.
• Signing in to an MCP server in one editor is no longer undone by the other.
• VS Code: a utility model picked elsewhere is no longer set back at start, and changing inferpal.model takes effect at once instead of after a restart.
• Visual Studio: two open windows no longer erase each other's prompt history.
• /arena: a vote goes to the battle you saw, never to one started in the other editor.

The agent's edits

• An edit that could not be written no longer counts as a change. On a read-only file, the result bar said "edited 1" and offered Undo, and the plan marked the step done.
• The agent fixes the build it broke instead of stopping, even after reading files to understand the errors — and no longer stops in the middle of a change across files, where the build is broken until the last edit.
• The agent is faster at the default 8K context window: it no longer asks the model for a summary it cannot use at almost every task.
• The agent's closing summary no longer forgets the start of a long run ("no file was modified" after a run that wrote two).
• Typing /resume during a step pause resumes the agent — in Visual Studio it stopped the run.
• A rename done by hand in a large repository is no longer reported as complete by silence, and a C# rename that names its file with a leading slash is accepted.

Tests and builds read right

• No more "the last test run failed" when the tests never ran — the runner was not installed, npm could not start, the path did not exist — and /tdd stops at once instead of spending five rounds changing sound code.
• A test run that fails before printing its summary is reported as failed (a Rust crate that does not compile, an npm test whose tsc step fails), and a .NET run where one test project never ran is no longer reported as passed.
• A test run piped through tail no longer counts as a passing check.
• The result bar and the end-of-turn notice agree on the last check, and in VS Code errors in the Problems panel count as a failed build.
• Checking the build says what it did not cover when the workspace has several solutions or projects, and "Fix with AI" and /fix-build say which files they left out.
• Linux and macOS: the same project is built and the same files are picked on every machine — builds, Smart Fix, tests, /test's target project, the project map.

Tool calls from more models

• Qwen and GLM calls written as text keep their values as written — creating a package.json, or changing false to true, no longer fails.
• Gemma 4 calls in unexpected forms now run, and arguments written beside the tool name are read (a whole test suite no longer runs instead of the tests asked for; background commands can be checked and stopped).
• A search or a test filter written as a number now works.

Inline completions

• Suggestions stop where the code you already have begins, instead of copying the lines below the cursor and going on; an HTML, XAML or Razor element is no longer closed twice.
• Visual Studio suggests after a space or a dot (return , = , user.), as VS Code does.
• Suggestions in the middle of a line arrive much sooner with models that have no fill-in-the-middle support.

Code analysis and search

• Call graphs see your methods named Add, Get, Count…, mark a call ambiguous when several classes define it, and no longer miss a caller that makes many other calls first.
• The impact report counts every indirect dependant, not the first 60.
• The agent can read every line of a very large file (a large log, CSV or JSON).
• @-mentions offer the file named exactly as you type it first, in large repositories too.
• @diff and Visual Studio's @problems say when there is nothing to attach, and a cut diff says how much was kept.
• Project rules and index exclusions are no longer skipped on a busy machine.

Figures that are right

• The size shown for a pinned file is what each question carries of it, and its whole size when it is cut.
• The code search's file count follows the index after files are deleted or moved.
• @Docs says when a site was only partly indexed, and no longer blames a fully indexed source for another's missing embeddings.
• /history, /onboard and /onboard context say what they did not show — more matches in a session, indexExclude entries not applied — and the project brief is the same on every machine.

Visual Studio

• A question asked with tools off waits as long as any other, instead of timing out on a slow or loading model.
• Files attached in the chat box go with a template, /explain, /review or /debug.
• Regenerate sends the question again with its code, files and model.
• After you open another solution, the assistant is told about it.

VS Code

• The settings page fills in once Inferpal has started, instead of staying empty.
• Regenerate resends the question's files, and is offered only when there is a question to answer again.
• Explain, Fix, Refactor and Doc from the editor menu say why they wait while a turn is running, and Stop works while a change waits in the Refactor Preview.
• Deleting an approval rule deletes that rule, never the one below it.
• With no folder open, the chat no longer blames your model server; it says a folder is needed.
• Your next question is no longer merged into a turn where the agent only used tools, and after a stopped answer the next question is checked against the context window again.

Install

• Visual Studio 2026 — Inferpal-vs2026-1.7.4.vsix below, or the Marketplace. Close Visual Studio before installing a downloaded VSIX.
• VS Code — the VSIX for your platform below (win32-x64, linux-x64, darwin-arm64), or the Marketplace.
• Checksums: SHA256SUMS.txt.

Full changelog: v1.7.3...v1.7.4

Inferpal 1.7.3

Forty-five fixes, and most of them are the same story: the assistant was told something that was not true — a debugger that never answered read as "nothing is paused", a half-built index read as the whole project, an edit cut at a quote read as applied — or something was lost without a word.

Safety first

• VS Code: a key typed for something else can no longer approve a tool. When an approval card appeared it took the keyboard focus, so an Enter or a space meant for the message box or your code could click "Allow once" on a command you had not read. The card no longer takes focus; Enter and Esc answer it only from an empty message box, as in Visual Studio.

Edits and renames that hold

• No more minutes of "Thinking…" when the assistant garbles an edit, and an edit is no longer applied halfway when the model mis-writes a quote (typically the $" of a C# string): such an edit is stopped within seconds, refused before anything changes, and the assistant gets another try instead of being stopped for "repeating itself".
• The agent no longer stops before making the change it planned — reading files no longer counts as having done the edit step.
• An edit refused because a line did not match now says which line, and where it differs.
• Renaming code works far more often, in one step after you approve its changes; a rename limited to one folder no longer breaks the callers outside it, and calls in modern .NET projects (implicit usings) are no longer missed.

Tests, builds and commands read right

• /tdd no longer edits sound code when dotnet cannot tell which project to test, and a folder path ending with a backslash reaches dotnet intact.
• Failing Rust tests are named, with what their assertion compared.
• A failing test run is reported even when the assistant ran it as a shell command, and a command stopped at its time limit keeps its error output.
• Git status shows what was staged before the first commit, and Smart Fix says when it could not check a change.
• A command time limit of 0 or less no longer breaks every command; the VS Code settings refuse values outside their range and name the field.

Debugging

• An unresponsive debugger is no longer taken for an answer. When Visual Studio's debugger was held by a dialog, or a request failed in VS Code, the assistant was told its breakpoint was refused, that nothing was paused, or that the program had ended. It is now told the debugger did not answer, and why; /debug says so too.
• The breakpoints the assistant set are removed when your program finishes after it continues or steps — the most common way a session ends — instead of staying in your workspace and stopping your next run.
• A program still running under the debugger is no longer reported as "not started".

Models, servers and MCP

• /hardware, /models running and the settings page no longer say "nothing is loaded" when they cannot tell — LM Studio behind a proxy that forwards only /v1.
• /bench and /arena no longer pick an embedding model.
• Inline completions a model refuses really pause, instead of being retried every 30 seconds in VS Code.
• An MCP server that stops accepting your sign-in asks you to sign in again; removing an MCP server removes its stored sign-in; a server that refuses your API key says so; a server list written as an array is reported instead of silently ignored.
• Visual Studio: switching backend in Settings says it needs a restart, instead of "cannot reach".
• The support bundle no longer contains the folder of a model loaded from a file, and no longer says inline completion works after it stopped.

Memory and search

• The assistant's memory keeps its newest notes when it grows large — the notes just written were the ones the next session never saw.
• Re-indexing a documentation site that throttles the crawler no longer shrinks its @docs index.
• Code search says when the project is still being indexed, instead of answering from the first few files as if they were the whole project.
• /onboard context no longer describes a folder it could not read as empty.

Visual Studio

• Ghost text follows its settings: turning inline completion off, choosing a completion model or a mode now applies.
• "Accept all" in the diff preview no longer overrides a change you rejected.
• Inline completion stays light next to very long lines, and "Automatic" language can be chosen again in Settings.

VS Code

• "Regenerate" asks the model again, instead of running your last slash command again.
• Each answer is shown once, and the assistant's passing remarks ("Let me read the file first.") no longer stay above it.
• A file whose name contains a space is attached, a question that mentions <think> is shown as you typed it, and the exported conversation counts its turns right.
• Everything in a window uses that window's model, the utility model applied by /onboard apply stays applied, and the context ring keeps the model's real window after you save the settings.

Install

• Visual Studio 2026 — Inferpal-vs2026-1.7.3.vsix below, or the Marketplace. Close Visual Studio before installing a downloaded VSIX.
• VS Code — the VSIX for your platform below (win32-x64, linux-x64, darwin-arm64), or the Marketplace.
• Checksums: SHA256SUMS.txt.

Full changelog: v1.7.2...v1.7.3

Inferpal 1.7.2

Forty-one fixes, and most of them are the same story: something reported as done — or said nothing — when it had not happened. An edit the editor refused, a rename that broke the build, a restore that destroyed its own backup, a review that never read a new file.

Your work stays safe

• /test no longer loses tests you have not saved yet. Visual Studio now works from the test file open in the editor; VS Code asks you to save it first.
• Restoring a backup can no longer destroy it. Restoring the oldest backup of a file, or undoing a run that wrote the same file many times, deleted the very version being restored. And undoing a run leaves alone a file it could not back up first.
• Undoing a run twice no longer throws away your fixes, and the Undo button of a finished run no longer stops the answer in progress (Visual Studio).
• A conversation you cleared stays cleared, and /clear no longer loses the conversation when the model writes a long title.
• VS Code: saving a session under a name with : or / can no longer silently replace another one.
• A long file is no longer rewritten from its first page: the assistant must have read every line before replacing a whole file.
• Editing code that contains a tool-call example no longer deletes it.

Edits do what they say

• Renaming a symbol no longer breaks the build. A class is renamed with its constructors, an interface method with its implementations, a virtual method with its overrides.
• Visual Studio: an AI edit the editor refused now says so — when you kept typing in the document while the model was working, /refactor, Edit with AI and /test reported success, or nothing, on an unchanged file.
• Visual Studio: the agent works on the file you are working in, not the one the chat was opened from; a closed file is no longer treated as open.
• Inline completions keep their own closing brace or parenthesis, so an accepted block before a method's end compiles.
• "Replace every match" works when the code is indented differently, /test keeps an existing test file's line endings, and a change proposed by /task through a multi-edit can be applied.

Search and review see the whole project

• Code search finds properties, interface methods and enum values written on one line.
• analyze_code sees who uses a React component: .tsx and .jsx files were left out, and a component used by every page was called safe to refactor.
• @folder and @file read C, Rust, Ruby, PHP, Kotlin, Swift, SQL, shell and more, and say how many files of other types they left out.
• /check reviews new files anywhere in the repository, not only under the folder you opened.
• Renaming or moving a folder no longer leaves the assistant looking at the old one.
• @Docs crawls minified documentation sites, and fetch_url names a PDF or an image instead of handing it to the model as text.

The agent

• No longer stopped "for repeating itself" while polling a build, checking again after a fix or stepping through the debugger — and that notice no longer claims a summary that was never written.
• A turn or a /task that runs out of steps says so, with the real limit.
• A large attachment no longer pushes a long conversation past the model's window.
• /commit-exec waits for your git hooks instead of stopping them after 15 seconds.
• Commands show the folder they run in whenever it is not your project's root; custom agent tools put their arguments where you wrote {args} and run in your project; with PowerShell 7, command output no longer carries colour codes.

Visual Studio

• The first inline suggestion after a question is no longer slow.

VS Code

• The model menu lists the models your server has now, new prompt files appear in the / suggestions at once, and the pinned-file chips follow the settings.
• A slash command added in the settings without its / now works, and agent tool names are written the way the assistant reads them.
• A long answer full of email addresses no longer freezes the chat.

Also

• /branch keeps building on the session you loaded after a restart; /onboard apply and /prompts say what they leave aside; an empty project-memory note no longer asks for approval.

Install

• Visual Studio 2026 — Inferpal-vs2026-1.7.2.vsix below, or the Marketplace. Close Visual Studio before installing a downloaded VSIX.
• VS Code — the VSIX for your platform below (win32-x64, linux-x64, darwin-arm64), or the Marketplace.
• Checksums: SHA256SUMS.txt.

Full changelog: v1.7.1...v1.7.2