Added RAG implementation, setup wizard and other fixes

[lsoft]

Five things which grew out of one another: the MCP proxy was fixed first, then the natural language
search was taught to use an embedding index, the search machinery which came out of that turned out
to be the right home for a grep tool as well, and a setup wizard was added so a new install has a
guided way to replace the configuration this repository ships with.

Version bumped to 4.3.0. The index files of TestSubject are regenerated in the new format.

RAG for the natural language search, and a window to calibrate it

The search used to hand every text file of the scope to the model, one context window at a time.
With Use RAG checked it asks the embedding index which files are worth reading and gives the model
those only.

• Search\FreeAIr.Search.csproj (netstandard2.0) is extracted out of the VSIX: the index format, the
  vector codec, the outline tree and the ranking know nothing about the IDE, and the VSIX assembly
  cannot be loaded by a test runner. That is what makes any of this testable at all —
  Search.Tests\ is new too.
• New on-disk format: quantized vectors written one per line, no embedding for an outline which is
  nothing but the name of the member it describes, and an order which does not depend on the machine
  which wrote the file. Eighteen times less disk space on the sample solution, and a merge instead of
  a conflict when two branches rebuild the index. Not backward compatible — an index built by
  4.2.11 or earlier has to be rebuilt.
• The index remembers which model built it by storing the vectors of a few fixed sentences, so a
  search with another model is refused instead of returning noise. The vector length alone did not
  catch this, since unrelated models often share it.
• RagMinScore is gone. Every build measures what a query with no answer scores on this model and
  this solution, and Sensitivity says how far above that measured noise a file has to stand — one
  number which keeps its meaning when the embedding model changes.
• Extensions / FreeAIr / Open RAG search calibration window... asks the index the questions the
  user would ask the search, shows what the threshold lets through and what it cuts, and saves the
  result without rebuilding the vectors: a threshold is fixed in a minute instead of in an hour of
  embedding.
• The search is cancellable at every step, shows its shortlist before the first answer arrives, and
  writes what it is doing into a FreeAIr natural language search pane of the Output window instead
  of giving up in silence in six different places.

A grep tool for the built-in Visual Studio MCP server

Finding where a text occurs used to mean handing whole files to the model one by one.
VisualStudio.SearchFileContent answers that question directly: the matching lines, each with its
file and line number.

Nothing is installed and no process is started — the matching is Search\Grep\. Shipping a
grep.exe inside the VSIX would have bought speed nobody asked for, at the price of a third party
binary, its licence and the antivirus reports about it.

Parameter
pattern	the only required one; a literal by default, a .NET regular expression when is_regular_expression is set
case_sensitive, whole_word	the usual two
invert_match	the lines which do not match, the way grep -v does
search_scope	the files of the projects (default), or every file in the solution folder
subfolder, file_mask	*.cs;*.xaml; a mask which starts with ! subtracts, so *.cs;!*.Designer.cs keeps generated code from eating the budget
context_line_count, max_result_count	0–5 lines around a match; 100 matches by default, 500 at most

• The answer 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.
• Running inside devenv is what a grep.exe could not have done: a document which is open and not
  saved yet is searched as the user sees it. 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.
• Build output, version control and package folders are skipped, files which are not text are
  detected by extension and by content, and the regular expression — written by a model — is given a
  five second timeout, so one pathological file is reported instead of hanging the IDE.

MCP proxy lifecycle, options and concurrency

• A server which was reconfigured no longer gets a second wrapper, so its tools stopped being offered
  to the model twice, three times and so on.
• The JSON-RPC channel is re-attached to the streams of a restarted proxy, which used to leave every
  MCP call after the first crash lost, and the configuration is pushed into the fresh process again.
  A failed start is retried instead of ending the monitoring for the session.
• Proxy.zip decides "already unpacked" by a marker written after the last entry, so an interrupted
  unpacking is repaired instead of being trusted.
• Options: a deserialization which returned null against its own contract, a serialization which
  threw when no solution was open, and an MCP server installation which added a duplicate entry.
• Concurrency: ChatContainer handed out its live list of chats and read it without the lock,
  LLMReader could hand a cancelled token source to a read started right after a stop, and
  FileCacheEntry updated its value across an await unsynchronized.
• UI: the buttons inserted into the Git Changes pane and the Find in Files dialog are restored
  after Visual Studio rebuilds them, and windows are no longer subscribed to Closed on every
  Loaded.

A first-run setup wizard

A new install starts with the configuration this repository ships with — placeholder tokens, the
author's own endpoints — and the only way to replace it was to reverse-engineer the settings file or
to find the control center's per-entity editors one at a time. The wizard walks through agents, MCP
servers, actions and the remaining settings, says what each step is for, and writes nothing until
Finish. An info bar offers it the first time FreeAIr runs (alongside the release notes, once), and
it stays available from Extensions / FreeAIr / Open setup wizard....

• SetupWizard\FreeAIr.SetupWizard.csproj (netstandard2.0) holds what needs neither Visual Studio
  nor WPF: step navigation and skipping, the agent-field and action-binding validators, the known
  endpoint and known MCP server catalogs. SetupWizard.Tests\ covers it with 45 tests. Same split,
  and the same reason, as FreeAIr.Search.
• That project carries no resources, so its validators report enum codes and the view model renders
  them into localized sentences. The window itself is fully localized — 79 new keys in the en,
  ru and zh-Hans resx — and takes its colours from the Visual Studio theme rather than from
  literals, so its text is readable under a dark theme too.
• On a first run the shipped sample agents are wiped and the user is guided through creating one
  of their own field by field, each field with its own explanation, plus a test connection check of
  the endpoint (the same GetModelsAsync call the model picker makes) and a choice between pasting a
  token and naming an environment variable which holds it. A later run gets the ordinary multi-agent
  list editor instead.
• The actions page explains what an action is, gives examples, and warns in red about every
  action naming an agent which does not exist — which is what renaming or deleting an agent silently
  produces. A Use default values button restores the shipped prompt library bound to the first
  agent.
• The run whole line completion as you type switch sits on that same page, because it is what makes
  the whole line completion action live or dormant: while it is off that 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 on the MCP servers page — a public
  HTTP endpoint, so nothing is downloaded and no token is needed. Its name and endpoint moved into
  KnownMcpServerCatalog, which the control center's install command reads too, so the two cannot
  drift apart and a server installed by one is recognized by the other.
• The GitHub MCP server is opt-in: nothing asks for a GitHub token until the box is ticked.
• Nothing here may throw into WPF's dispatcher — an unhandled exception from a modal dialog inside
  devenv ends Visual Studio rather than the dialog — so every command body, the window loading, the
  info bar and the package hook each catch, log to the activity log, and report.
• Existing editors are reused rather than reimplemented: the MCP servers and actions steps open
  McpServerConfigureWindow / ActionConfigureWindow against a clone, so a cancelled sub-dialog
  leaves the wizard untouched. DirectOrEnvStringHelper moved into the new project instead of being
  copied into it, and the shipped defaults stayed the single source they already were.

Separately, whole line completion no longer opens an error dialog on every pause in typing when its
action names no existing agent. That is how the shipped agent_name_must_be_set placeholder used to
interrupt a brand new install, the wizard included. The complaint goes to the activity log now, and
is shown as a dialog only when a suggestion was asked for explicitly with Alt+A.

Documentation

ARCHITECTURE.md describes the projects, the flow of a chat and the MCP, search and setup wizard
subsystems; CLAUDE.md holds what an automated agent needs to build this repository — full MSBuild
rather than the .NET SDK, why dotnet build breaks the next build, how to run the tests, and the
fact that a new file in the legacy-format FreeAIr.csproj has to be added to it by hand or it is
silently not compiled.

Checks

FreeAIr.sln builds with 0 errors in Release (the ~800 warnings are pre-existing), and
run-tests.bat is green: 112 passing in FreeAIr.Search.Tests with 6 integration tests skipped
without an embedding server, and 45 passing in FreeAIr.SetupWizard.Tests.

🤖 Generated with Claude Code