RELEASE_NOTES.md

Overview

Please report any bugs to the github repo.

Feedback

Visual Studio extension authors suffers of lack of feedback. Please share your feelings and gratitude. Choose one or few available options:

1. Please gift a ★★★★★ rating for this VSIX in the VS Marketplace.
2. Gift a ★ to the github repo.
3. If you are enjoying FreeAIr to the enough level to donate, there are many small cancer patients that need your help. Please provide your help them!

Other my VSIXes may interest you

• Fix incorrect namespaces for a single file, folder, project or a whole solution and rules the resulting regressions in the code (including XAML), e.g. fixes the broken references. This extension works like Resharper Adjust namespaces function.
• Visual Studio extension for quick watch window which allows to debug and edit multilines expressions.
• If you are using plain SQL inside you code base you may want to validate these queries against your DB schema right inside Visual Studio. ReSequel does exactly that.
• This extension puts Roslyn metadata of your project into the in-memory sqlite database and allows to you to execute queries to the database.
• The faster way to add strings to your multilanguage resx files. Just install the extension, select the text and press Alt+J.
• A Visual Studio extension which creates codelenses allows you to go to sync sibling method for async methods and vice-versa even if sibling method is in different file or code generated.

My others extensions lives here.

FreeAIr Release Notes

4.3.1

• Fixed few bugs.
• User chats are now persistent (.freeair/chats) if the chat is started with a solution open.

4.3.0

• User-started chats are persisted next to the solution in .freeair\chats and restored when the solution opens. Automatic chats are not. A chat created with no solution open stays in memory only. Closing a chat deletes its file; shutting Visual Studio down does not.
• A first-run setup wizard replaces the trial-and-error of hand-editing the settings file: it walks through agents, MCP servers, actions and the remaining settings, explains what each of them is for, and writes nothing until the last page. Open it any time from Extensions > FreeAIr > Open setup wizard.... On a brand new install, an info bar offers it alongside the release notes the first time FreeAIr runs. It is fully localized, like the rest of the product.
  • On a first run it removes the shipped placeholder agents and guides you through creating one of your own field by field, with an explanation under each, a "test connection" check for the endpoint, and a choice between pasting a token and pointing at an environment variable that holds it.
  • The actions page explains what an action is, gives examples, and warns in red about any action bound to an agent that does not exist - which is what happens after renaming or deleting one. A Use default values button restores the shipped prompt library bound to your first agent.
  • The run whole line completion as you type switch sits on that same page, because it is what decides whether the whole line completion action is live or dormant: while it is off the action is neither bound by the defaults button nor reported as a problem, and ticking it does both.
  • The Microsoft Learn documentation MCP server is one checkbox away on the MCP servers page - it is a public HTTP endpoint, so nothing is downloaded and no token is needed.
  • The GitHub MCP server is opt-in: nothing asks for a GitHub token until you tick the box.
• The Use RAG checkbox of the natural language search is implemented. The search now takes the files whose outlines are the closest to the query and asks the LLM about those only, instead of reading through the whole solution.
• The embedding json files have a new, much smaller format: vectors are quantized and stored one per line, and outlines which are nothing but the name of the member they describe are not embedded at all. On the test solution that is 18 times less disk space. The files are also written in an order which does not depend on the machine which built them, so two people regenerating them in two branches no longer get a git conflict out of it.
• There are three index files instead of four: the tree of the outlines is not stored any more, it is derived from the outlines themselves. One file less to merge.
• The natural language search can now be cancelled at every step, including while the index is being read.
• The GetAllSolutionFiles MCP tool no longer reads the embeddings on every call, and the index is shared between the tool and the search instead of being loaded twice.
• Fixed an agent with no token being unusable for embeddings, which is how a local embedding server is normally configured.
• Whole line completion no longer opens an error dialog on every pause in typing when its action is not wired to an existing agent. The complaint now goes to the Activity Log, and is shown as a dialog only when you asked for a suggestion yourself with the keyboard shortcut. The broken binding is still reported in red by the setup wizard, where it can be fixed.
• The Use RAG search no longer has a similarity threshold to guess. Every index build now measures what a query with no answer scores on your model and your solution, and the new Sensitivity setting says how far above that measured level a file has to stand — one value which keeps its meaning when you change the embedding model. You can add your own questions, including ones whose answer you know, in the Calibration node; a query which fails to find its own file is reported at the end of the build, because that means the model does not understand your code.
• There is a window for that measurement: Extensions / FreeAIr / Open RAG search calibration window.... Ask it the questions you would ask the search, mark the file which answers each one — or say that nothing here does — and it shows what the threshold lets through and what it cuts. Saving stores the questions in the settings and the numbers in the index without rebuilding the vectors, so a threshold can be fixed in a minute instead of in an hour of embedding. The window names the agent it is asking and lets you change it, and it describes itself in a block at the top. The Создание NLO Json файла window now carries its own description in that same block.
• Fixed an agent without a token being invisible to everything which looks an agent up by name. An embedding server run locally wants no token, so the agent which had built the index was never found and a cloud chat agent was silently used to vectorize the query instead — which answers a request for embeddings with HTTP 400. The natural language search was affected as well.
• The natural language search no longer gives up in silence. It used to walk through the scope, the action and the agent pickers, and a return at any of them left the button looking broken: no window, no message, nothing in the log. Every one of those steps now says why it stopped — the most likely reason being that no agent has a token, which is how the agents of a local server are normally configured. An empty file mask in Find in Files means every file now, instead of killing the search with an exception nobody saw.
• Every step of the natural language search writes itself into a FreeAIr natural language search pane of the Output window: which agent it asks, how many files the mask matched, what the RAG shortlist kept, how big the prompt and the answer were, and how many matches of the answer were usable. An answer which carries no matches is reported as such instead of being counted as Found 0 items.
• Fixed the search results window failing to open at all: its progress bar was bound two way onto a read only property, which throws while the window is being built.
• The results window names the agents behind the answers: the one which was asked, and — with Use RAG — the one which turned the query into a vector. The second one is normally chosen for you, from the index, and until now there was no way to see which one it was.
• The Confidence column reads High (85) instead of 85. The number is a guess of the model, not a measurement, and the word says how much of the list is worth opening. The boundaries are 70 and 40.
• Fixed the Edit actions window: clicking an action did nothing. Its SelectedAction was writing into the backing field of the property while the whole rest of the window kept reading a separate field, which therefore stayed empty — the editing panel was hidden and every command disabled.
• An embedding server which refuses a request is now quoted. HTTP 400 on its own says nothing about which model, which endpoint or which parameter it did not like, and the server had already explained all of it in an answer that was being thrown away.
• The index remembers which embedding model built it, by storing the vectors of a few fixed sentences. A search with a different model is now refused instead of silently returning nonsense — the vector length alone did not catch it, since different models often share it. The model name the server itself reports is recorded too, which is the only name that changes when a local server is given a different model to load.
• The RagTopOutlineCount, RagMaxFileCount and RagMinScore options moved from the unsorted settings into a Rag node of their own, and RagMinScore is gone: it is replaced by Sensitivity above.
• The MCP proxy is restarted properly now. ProcessMonitor did restart a proxy which had died, but the JSON-RPC channel stayed bound to the streams of the dead process, so every MCP call after the first crash was lost. The channel is re-attached to the streams of the new process, and the configuration is pushed into it again — a freshly started proxy hosts no MCP servers until it is told about them.
• A server which is reconfigured while it runs no longer has its tools offered to the model twice, three times and so on: the wrapper of the running server is reused instead of a second one being appended, and the wrapper of a server which failed to start is dropped along with its tools.
• A proxy which fails to start is retried after a pause instead of ending the monitoring for the rest of the session, and is given up on only after several failures in a row.
• Proxy.zip decides that it is already unpacked by a marker file written after the last entry. The folder used to be created before the first one, so an unpacking which was interrupted left behind a truncated folder which was never repaired.
• Installing the MSDN MCP server overwrites its entry instead of adding a second one under a name which may be taken already.
• Asking the options for a place explicitly and finding nothing there no longer returns the null its callers dereference, and saving solution related options with no solution opened reports what has happened instead of throwing ArgumentNullException.
• A set of concurrency defects is fixed: the list of chats used to be handed out live and read without its lock, while chats are added from the UI thread and their statuses change on the threads which stream the answers; a read started right after a stop could be given the cancelled token source of the previous one; and the file cache updated its value and its signature around an await without any synchronization.
• The built-in Visual Studio MCP server has a new tool, VisualStudio.SearchFileContent: the model can ask where a text occurs and gets the matching lines back — each with its file and its line number — instead of reading whole files one by one until it finds them. Nothing is installed and no process is started, the matching is done by the extension itself.
• The pattern is literal by default and a .NET regular expression on request, both through the same code path. case_sensitive and whole_word are the other two switches, and invert_match reports the lines which do not match, the way grep -v does.
• search_scope chooses between the files of the projects — the default — and every file in the solution folder. The walk skips bin, obj, .git, node_modules and the like either way, so a search does not answer with build output.
• file_mask is a list like *.cs;*.xaml, and a mask which starts with ! subtracts: *.cs;!*.Designer.cs is what keeps generated code from eating the context budget.
• The answer is capped — 100 matching lines by default, 500 at most — and says whether the cap was reached. A truncated list which does not admit to being one is read by the model as the whole truth.
• A document which is open and edited but not saved yet is searched as you see it on the screen rather than as it is on the disk. Those buffers are collected on the main thread and the scan itself runs off it, so a solution of thousands of files does not freeze the IDE.
• The pattern is written by a model, so a regular expression is given a five second timeout: the file which makes it misbehave is reported and the rest of the search survives.
• Fixed a chat against LM Studio dying with Service request failed. Status: 400 (Bad Request) before the model was ever asked anything. Four of the built-in Visual Studio tools take no arguments and said so with an empty {} schema, which LM Studio validates and rejects — and it rejects the whole request, so a single such tool took every other tool and the conversation down with it. Every tool schema is now brought to the shape the endpoints ask for on its way to the model, which also covers the tools of third party MCP servers: their schemas reach FreeAIr unseen and are just as free to declare no arguments that way.
• Fixed the same 400 in the natural language search and in the generation of the outlines, from the other end: they asked for the answer in the json_object response format, which LM Studio does not implement at all. The json is asked for in the prompt now, as it already was, and the answer goes through the same cleanup that strips the markdown fences and the reasoning blocks before it is parsed.
• Tool calls are reassembled from the whole stream rather than from its first chunk. An endpoint may send the name of a tool in one chunk and its arguments in the chunks that follow, and only the first was being read: the tool then ran with no arguments at all, and the half-read call was sent back to the server with the next request, which ended the conversation with an Internal Server Error.
• The new format is not backward compatible: an index built by FreeAIr 4.2.11 or earlier has to be rebuilt.

4.2.12

• Fixed bug, thanks to nrmncr for reporting it.

4.2.11

• Thanks to ish-1313 for his\her first contribution to FreeAIr project:
  • Added chat name.
  • Added ability to rename chat (by clicking on its name).
  • Added ability to filter models when asking a model provider about its available models.
• Added ability to resize in situ chat window by dragging a special control.

4.2.10

• Fixed bug with GetAllSolutionFiles MCP tool.

4.2.9

• Improved usability in case of no agent with non-empty token exists.

4.2.8

• Fixed bug with parsing markdown with leading \n\n.

4.2.7

• Fixed #51 (improved agent <-> action pairing).

4.2.6

• Fixed few bugs.

4.2.5

• Fixed few bugs.

4.2.2

• Fixed few bugs.

4.2.1

• Fixed stupid bug.

4.2.0

• Added initial support for Visual Studio 2026.
• Added voice prompting with 4 voice providers and LLM-based post-processor.
• Simplified IPC data flow with FreeAIr child process Proxy.exe.

4.0.2

• Added in situ chat mode.
• BREAKING CHANGE: changed MCP servers json format in FreeAIr options json file.
• Added easy way to enable Microsoft MDSN MCP server.
• Added asking permission from the user to call any MCP tools LLM requested.
• Fixed various bugs.

3.7.1

• Fixed minor bug.

3.7.0

• Added a command to add selected files (in Solution Explorer window) to the current chat context.
• Added ability to add items to chat context with drag-and-drop.
• Improved VS theme support.

3.6.0

• Improved markdown parser (added bold text, added support for markdown tables).
• Fixed bugs in markdown parser.

3.5.0

• Improved markdown parser.
• Added ability to collapse XML nodes in LLM answers.

3.4.0

• Added NLOs into solution item metadata for GetAllSolutionItemsTool MCP tool. This is helpful if you are using LLM to make decisions about your solution structure (where to add new entity, etc.).

3.3.0

• Added Russian and simplified Chinese localization.

3.2.0

• Added ability to set whole line completion prompts. The suggested prompt is:

- In the document {CONTEXT_ITEM_NAME} suggest whole local code completion at the place where {WHOLE_LINE_COMPLETION_ANCHOR} anchor is set.
- Follow the style, the formatting and the actual indent of the provided code.
- Do not post whole modified document. Do not post anything except the code snipped you suggest to add to that place. Skip any preamble, I need only suggested code part!
- If you see that nothing can be suggested, return answer `//Nothing to suggest`.

3.1.0

• Added ability to edit available MCP servers visually.
• Added ability to install MCP servers from Docker registry.
• Added ability to hide automatically created chats.

3.0.0

• Major revision has changed because of breaking changes in FreeAIr properties. Now you can control FreeAIr via its Control Center. Unfortunate, previous FreeAIr properties has been deleted, please, resetup it. For additional information please refer to readme.
• Added multiagent support.
• Added chat chosen agent.
• Added a solution-related or VS-related configuration file of FreeAIr.
• Added user-defined prompts.
• Fixed various bugs.

2.5.0

• Switched to different rendering of LLM answers.

2.4.0

• Added a FreeAIr Control Center window.
• Added a global MCP tools status, and chat-scoped MCP tools status.

2.3.0

• Added a natural language search (across whole solution or current project).
• Implemented UI improvements.
• Added a github MCP server and Visual Studio MCP server and its tool.
• And a lot more.