feat(desktop): Node producer + FixtureBackend + MCP tools + Windows bridge IPC

.gitignore

@@ -17,3 +17,20 @@ dist/
 
 # Tesseract.js downloads language data into cwd by default
 *.traineddata
+
+# .NET / MSBuild output (used by apps/agent-runner/bridges/windows)
+bin/
+obj/
+*.suo
+*.user
+*.userprefs
+
+# JetBrains
+.idea/
+*.iml
+
+# Visual Studio
+.vs/
+
+# Local dev junk
+*.swp

apps/agent-runner/bridges/windows/AgentMark.Bridge.Windows.csproj

@@ -0,0 +1,32 @@
+<Project Sdk="Microsoft.NET.Sdk">
+
+  <PropertyGroup>
+    <OutputType>Exe</OutputType>
+    <TargetFramework>net8.0-windows</TargetFramework>
+    <RootNamespace>AgentMark.Bridge.Windows</RootNamespace>
+    <AssemblyName>agentmark-bridge-windows</AssemblyName>
+    <Nullable>enable</Nullable>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <LangVersion>12.0</LangVersion>
+    <TreatWarningsAsErrors>true</TreatWarningsAsErrors>
+    <!-- Build to a predictable output path so the Node-side spawn knows
+         where to find the exe. -->
+    <BaseOutputPath>$(MSBuildThisFileDirectory)bin\</BaseOutputPath>
+    <BaseIntermediateOutputPath>$(MSBuildThisFileDirectory)obj\</BaseIntermediateOutputPath>
+  </PropertyGroup>
+
+  <PropertyGroup>
+    <!-- The bridge talks to UIA (Windows accessibility API). FlaUI is a
+         maintained MIT-licensed wrapper around UIA. -->
+    <Description>AgentMark Windows UIA bridge — accessibility-tree capture and action execution for kind:'desktop' snapshots.</Description>
+    <Authors>ThinkFleet</Authors>
+    <Company>ThinkFleet</Company>
+    <Product>AgentMark Desktop</Product>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <PackageReference Include="FlaUI.Core" Version="5.0.0" />
+    <PackageReference Include="FlaUI.UIA3" Version="5.0.0" />
+  </ItemGroup>
+
+</Project>

apps/agent-runner/bridges/windows/Program.cs

@@ -0,0 +1,220 @@
+// AgentMark Windows UIA Bridge
+// ----------------------------
+// Stdio JSON-RPC 2.0 server. Parent process (typically the Node-side
+// `WindowsUiaBackend` running inside the AgentMark MCP server, or a
+// ThinkFleet SaaS agent) launches this exe and talks to it over
+// stdin/stdout. One line of JSON per message.
+//
+// Protocol:
+//   Request : { "jsonrpc": "2.0", "id": <any>, "method": "<name>", "params": {...} }
+//   Response: { "jsonrpc": "2.0", "id": <same>, "result": {...} }   on success
+//             { "jsonrpc": "2.0", "id": <same>, "error": { "code": N, "message": "..." } }  on failure
+//
+// All diagnostic output goes to stderr — never stdout — so the framing stays clean.
+//
+// Methods (current set; more land as Phase 0e progresses):
+//   ping        — returns { pong: true, version: "...", arch: "arm64|x64" }
+//   capabilities— returns { methods: [...], uia_version: "..." }
+//
+// Phase 0e2/0e3 will add:
+//   capture     — walk a window's UIA tree, return DesktopCapture JSON
+//   execute     — drive an action by element_id (click/type/select/etc.)
+//   list_windows— enumerate top-level windows visible to the bridge
+
+using System.Text;
+using System.Text.Json;
+using System.Text.Json.Serialization;
+
+namespace AgentMark.Bridge.Windows;
+
+internal static class Program
+{
+    // JSON serialization options — camelCase to match the AgentMark
+    // wire format on the Node side.
+    internal static readonly JsonSerializerOptions JsonOpts = new()
+    {
+        PropertyNamingPolicy = JsonNamingPolicy.CamelCase,
+        DefaultIgnoreCondition = JsonIgnoreCondition.WhenWritingNull,
+    };
+
+    public static int Main(string[] args)
+    {
+        // Force UTF-8 on both pipes -- Windows defaults can mangle non-ASCII.
+        // Use UTF8Encoding(false) to avoid emitting a BOM on stdout, which
+        // would break JSON-RPC framing on the parent side.
+        Console.InputEncoding = new UTF8Encoding(false);
+        Console.OutputEncoding = new UTF8Encoding(false);
+        Console.Error.WriteLine($"[bridge] agentmark-bridge-windows starting (pid={Environment.ProcessId}, arch={System.Runtime.InteropServices.RuntimeInformation.ProcessArchitecture})");
+
+        // Synchronous read loop. Async stdin reads on Windows pipes have a
+        // known issue where ReadLineAsync() doesn't always return null on
+        // EOF, leaving the process hung even after the parent closes the
+        // pipe. ReadLine() handles EOF correctly. UIA calls are themselves
+        // blocking COM-STA invocations so async doesn't buy us anything.
+        var dispatcher = new RpcDispatcher();
+
+        string? line;
+        while ((line = Console.In.ReadLine()) != null)
+        {
+            line = line.Trim();
+            if (line.Length == 0) continue;
+
+            JsonElement reqId = default;
+            try
+            {
+                using var doc = JsonDocument.Parse(line);
+                var root = doc.RootElement;
+                reqId = root.TryGetProperty("id", out var idEl) ? idEl.Clone() : default;
+
+                var method = root.GetProperty("method").GetString()
+                             ?? throw new RpcException(RpcError.InvalidRequest, "method is required");
+                var paramsEl = root.TryGetProperty("params", out var p) ? p : default;
+
+                var result = dispatcher.Dispatch(method, paramsEl);
+                Console.Out.WriteLine(EncodeSuccess(reqId, result));
+            }
+            catch (RpcException rex)
+            {
+                Console.Out.WriteLine(EncodeError(reqId, rex.Error.Code, rex.Message));
+            }
+            catch (JsonException jex)
+            {
+                Console.Out.WriteLine(EncodeError(reqId, RpcError.ParseError.Code, $"JSON parse error: {jex.Message}"));
+            }
+            catch (Exception ex)
+            {
+                Console.Error.WriteLine($"[bridge] unhandled: {ex}");
+                Console.Out.WriteLine(EncodeError(reqId, RpcError.InternalError.Code, ex.Message));
+            }
+
+            Console.Out.Flush();
+        }
+
+        Console.Error.WriteLine("[bridge] stdin closed; exiting");
+        return 0;
+    }
+
+    private static string EncodeSuccess(JsonElement id, object? result)
+    {
+        var env = new RpcResponseSuccess
+        {
+            Id = id.ValueKind == JsonValueKind.Undefined ? null : id,
+            Result = result,
+        };
+        return JsonSerializer.Serialize(env, JsonOpts);
+    }
+
+    private static string EncodeError(JsonElement id, int code, string message)
+    {
+        var env = new RpcResponseError
+        {
+            Id = id.ValueKind == JsonValueKind.Undefined ? null : id,
+            Error = new RpcErrorBody { Code = code, Message = message },
+        };
+        return JsonSerializer.Serialize(env, JsonOpts);
+    }
+}
+
+// ──────────────────────────────────────────────────────────────────────
+// Wire format
+// ──────────────────────────────────────────────────────────────────────
+
+internal sealed class RpcResponseSuccess
+{
+    [JsonPropertyName("jsonrpc")]
+    public string Jsonrpc => "2.0";
+
+    [JsonPropertyName("id")]
+    public JsonElement? Id { get; init; }
+
+    [JsonPropertyName("result")]
+    public object? Result { get; init; }
+}
+
+internal sealed class RpcResponseError
+{
+    [JsonPropertyName("jsonrpc")]
+    public string Jsonrpc => "2.0";
+
+    [JsonPropertyName("id")]
+    public JsonElement? Id { get; init; }
+
+    [JsonPropertyName("error")]
+    public RpcErrorBody Error { get; init; } = null!;
+}
+
+internal sealed class RpcErrorBody
+{
+    [JsonPropertyName("code")]
+    public int Code { get; init; }
+
+    [JsonPropertyName("message")]
+    public string Message { get; init; } = "";
+}
+
+// ──────────────────────────────────────────────────────────────────────
+// Error catalog
+// ──────────────────────────────────────────────────────────────────────
+
+internal readonly record struct RpcError(int Code)
+{
+    public static readonly RpcError ParseError = new(-32700);
+    public static readonly RpcError InvalidRequest = new(-32600);
+    public static readonly RpcError MethodNotFound = new(-32601);
+    public static readonly RpcError InvalidParams = new(-32602);
+    public static readonly RpcError InternalError = new(-32603);
+
+    // Bridge-specific (above -32000)
+    public static readonly RpcError WindowNotFound = new(-32010);
+    public static readonly RpcError ElementNotFound = new(-32011);
+    public static readonly RpcError UnsupportedPattern = new(-32012);
+    public static readonly RpcError ActionFailed = new(-32013);
+}
+
+internal sealed class RpcException(RpcError error, string message) : Exception(message)
+{
+    public RpcError Error { get; } = error;
+}
+
+// ──────────────────────────────────────────────────────────────────────
+// Dispatcher
+// ──────────────────────────────────────────────────────────────────────
+
+internal sealed class RpcDispatcher
+{
+    private static readonly string BridgeVersion = "0.4.0";
+
+    public object? Dispatch(string method, JsonElement @params)
+    {
+        return method switch
+        {
+            "ping"         => HandlePing(),
+            "capabilities" => HandleCapabilities(),
+            _ => throw new RpcException(
+                RpcError.MethodNotFound,
+                $"Unknown method: {method}. Supported: ping, capabilities."),
+        };
+    }
+
+    private static object HandlePing() => new
+    {
+        pong = true,
+        version = BridgeVersion,
+        arch = System.Runtime.InteropServices.RuntimeInformation.ProcessArchitecture.ToString().ToLowerInvariant(),
+        processId = Environment.ProcessId,
+    };
+
+    private static object HandleCapabilities() => new
+    {
+        bridge = "agentmark-bridge-windows",
+        version = BridgeVersion,
+        methods = new[]
+        {
+            "ping",
+            "capabilities",
+            // "capture" and "execute" land in Phase 0e2/0e3
+        },
+        uiaProvider = "FlaUI.UIA3",
+        platform = "windows",
+    };
+}

apps/agent-runner/bridges/windows/README.md

@@ -0,0 +1,85 @@
+# AgentMark Windows UIA Bridge
+
+The Windows-side sidecar process for AgentMark Desktop. Walks the OS
+accessibility tree via UIA (using [FlaUI](https://github.com/FlaUI/FlaUI),
+MIT) and exposes capture/execute operations the Node-side
+`WindowsUiaBackend` consumes through stdio JSON-RPC.
+
+This is one implementation of the `DesktopCaptureBackend` interface
+defined in `@thinkfleet/agentmark` v0.4. macOS (AXAPI) and Linux (AT-SPI)
+bridges follow the same protocol.
+
+## Build
+
+Requires the .NET 8 SDK. From this directory:
+
+```powershell
+dotnet build
+```
+
+Output exe: `bin\Debug\net8.0-windows\agentmark-bridge-windows.exe`.
+
+## Run + smoke test
+
+The bridge speaks stdio JSON-RPC 2.0. One JSON message per line.
+
+```powershell
+.\bin\Debug\net8.0-windows\agentmark-bridge-windows.exe
+```
+
+It blocks waiting on stdin. Paste a `ping`:
+
+```
+{"jsonrpc":"2.0","id":1,"method":"ping"}
+```
+
+Expected response (single line, here pretty-printed):
+
+```json
+{
+  "jsonrpc": "2.0",
+  "id": 1,
+  "result": {
+    "pong": true,
+    "version": "0.4.0",
+    "arch": "arm64",
+    "processId": 12345
+  }
+}
+```
+
+Ctrl-Z then Enter closes stdin and the bridge exits cleanly.
+
+## Protocol
+
+All responses are JSON-RPC 2.0. Diagnostic output goes to stderr only —
+stdout is reserved for framed JSON messages.
+
+| Method | Status | Description |
+|---|---|---|
+| `ping` | Phase 0e1 ✅ | Liveness check, returns pong + bridge version |
+| `capabilities` | Phase 0e1 ✅ | Lists supported methods and UIA provider info |
+| `list_windows` | Phase 0e2 (planned) | Enumerate top-level windows visible to UIA |
+| `capture` | Phase 0e2 (planned) | Walk a window's UIA tree → `DesktopCapture` JSON |
+| `execute` | Phase 0e3 (planned) | Drive an action by element_id (click/type/select/etc.) |
+
+## Architecture notes
+
+- **Single-threaded for now.** UIA is COM-STA; once `capture` / `execute`
+  land we marshal those calls onto a dedicated STA thread.
+- **UTF-8 forced on both pipes.** Windows defaults will mangle non-ASCII
+  in window titles / cell values otherwise.
+- **All stderr goes to the parent process's stderr.** Useful for log
+  aggregation in the AgentMark MCP server when this bridge is spawned
+  as a subprocess.
+
+## Security posture
+
+This bridge runs as the logged-in user (never SYSTEM/admin). It can
+only see/drive windows that user can already see/drive — no privilege
+escalation. Stdio mode has zero network surface; only the parent
+process that spawned this exe can write to its stdin.
+
+A future `--transport=ws` mode (Phase 0f) will add a localhost-only
+WebSocket transport with auth-token gating, origin-header rejection,
+and process-identity verification for multi-consumer scenarios.