Skip to content

Inferpal 1.8.12

Choose a tag to compare

@github-actions github-actions released this 12 Nov 17:38
· 49 commits to master since this release

This release is about the moments around your work: switching servers, reading today's C#, and files that are busy or broken when Inferpal looks at them.

Switch servers without restarting

Visual Studio no longer asks you to restart when you change model server. Pick LM Studio instead of Ollama in the settings, with /setup, /import, or from VS Code, and your next question goes there. An answer already running finishes where it started.

On LM Studio, the context window Inferpal works with is now the one the server actually loaded for the model. It could count on 8K while LM Studio had loaded 32K, or the other way round, and a long run would hit the limit without being compacted.

And after LM Studio unloads a model you left idle, the next question no longer teaches Inferpal that the model reads slowly. The reload used to be counted as reading time, so the waiting card would promise about the time of a load for answers that take a few seconds. A time learned before the update can still show for a few dozen questions, until newer answers replace it.

Today's C#

The assistant knows C# up to version 15. Valid code such as a C# 14 compound operator or C# 15's union is no longer refused as a syntax error, and extension blocks are read as code by search and by renaming. Renaming reads your code at the language your projects compile.

Edits that stay where you asked

Some models, asked to change one part of a page, sent the whole page back as the replacement, and the top of the page ended up written twice. Such an edit is now refused before you are asked, and the model is told to send only the part it changes. Gemma 4 12B stopped doubling pages, and corrected its edit right away.

The agent also plans the question you just asked. After an earlier request, a small model could plan your new question as the old one and redo a file instead of answering.

When a file is busy or broken

A team's .inferpal/permissions.json that another program holds, or that holds invalid JSON, keeps its deny rules in force. If they were never read, every action asks you first, and you're told why. The repository's Smart Fix checks get the same care.

Your settings file read at the wrong moment no longer undoes a choice. Inline completion in Visual Studio could come back on right after you turned it off. Sign-ins to MCP servers added from your other editor could be erased. Neither happens now.

A package configuration you can't read no longer counts as "no private index", so a private package name isn't sent to the public registry. A failed build keeps its banner when its signal is busy for a moment.

Analyze with Inferpal, in the Debug menu

Stopped at a breakpoint or an exception in Visual Studio, Debug › Analyze with Inferpal now asks the chat why the program stopped, with the call stack, the variables and the code around them. It was announced in the Debug menu since 1.8.5 but never appeared there; only the call stack and variables windows offered it.

The file in front of you

In Visual Studio, going back to a file that was already open, without moving the cursor, could leave the previous file active for the assistant: it read that one as the file in front of you, and offered to insert at its cursor. The file you go back to is now the active one.

In Visual Studio 2022, a question could wait for minutes while the assistant asked which file was active. It now waits ten seconds at most, then says that Visual Studio did not answer and that you can click into the file or name it.

Safer cards

  • Plan mode no longer builds your project. It reads and plans; a build becomes a step.
  • A card that asks every time no longer offers "Always", which used to let the tool through everywhere else.
  • When you decline to let the assistant run a repository's code, that also covers dotnet test or pytest started through the shell.
  • A network path a model passes to a tool is refused at once, without contacting the other machine.

Long conversations and your settings window

A long conversation that was compacted twice, once in a small context window and once in a larger one, now comes back compacted when you reopen it. It used to come back whole, and its first question summarised everything again, from less than before.

In Visual Studio, the settings window shows what is saved each time you open it. Closed and reopened, it used to keep the values of its first opening, so a setting you changed in VS Code meanwhile looked unchanged. Cancel now puts back your lists too: MCP servers, pinned files, prompt templates and tools.

The bubbles that tell you the conversation was compacted, or recapped, now have a readable title in your language, "Context window" and "Session recap", instead of context_compact and ooda_recap. They are no longer steps of the agent's run either: in VS Code, a run that only read files showed a red cross, and offered "Fix with AI", when the chat had to drop old messages. And when you reopen a conversation saved from this version on that lost messages, that is said in plain text, as it was when it happened.

When a long conversation has its oldest messages dropped a second time (summaries off, or one that could not be made), the model now learns how many messages are really gone. It used to read "5 messages dropped" for 41, and could tell you something you said early on was never said.

In the settings window, turning a command or a custom tool on or off no longer removes one that VS Code or /import added while the window was open.

When a checkout or a pull changes the code of an MCP server your repository declares, Inferpal asks you about the new code before your next question and starts it once you agree. The old code no longer keeps answering until you restart.

Smaller changes

  • Screen readers announce the buttons of the Tools page with the tool's real name.
  • The status line says "Thinking…" instead of showing the model's reasoning cut mid-word.
  • The ACP agent comes as one executable with a README and the license, and starts even when your profile folder is read-only.

Update whenever you like. Your settings, conversations and indexes come along.