unity-mcp/MCPForUnity/Editor/Clients/McpClientConfiguratorBase.cs

768 lines
33 KiB
C#
Raw Normal View History

using System;
using System.Collections.Generic;
using System.IO;
using System.Runtime.InteropServices;
using MCPForUnity.Editor.Constants;
using MCPForUnity.Editor.Helpers;
using MCPForUnity.Editor.Models;
using MCPForUnity.Editor.Services;
using Newtonsoft.Json;
using Newtonsoft.Json.Linq;
using UnityEditor;
using UnityEngine;
namespace MCPForUnity.Editor.Clients
{
/// <summary>Shared base class for MCP configurators.</summary>
public abstract class McpClientConfiguratorBase : IMcpClientConfigurator
{
protected readonly McpClient client;
protected McpClientConfiguratorBase(McpClient client)
{
this.client = client;
}
internal McpClient Client => client;
public string Id => client.name.Replace(" ", "").ToLowerInvariant();
public virtual string DisplayName => client.name;
public McpStatus Status => client.status;
public virtual bool SupportsAutoConfigure => true;
public virtual string GetConfigureActionLabel() => "Configure";
public abstract string GetConfigPath();
public abstract McpStatus CheckStatus(bool attemptAutoRewrite = true);
public abstract void Configure();
public abstract string GetManualSnippet();
public abstract IList<string> GetInstallationSteps();
protected string GetUvxPathOrError()
{
string uvx = MCPServiceLocator.Paths.GetUvxPath();
if (string.IsNullOrEmpty(uvx))
{
fix: resolve Claude Code HTTP Remote UV path override not being detected in System Requirements .#550 * fix: resolve UV path override not being detected in System Requirements Fixes #538 The System Requirements panel showed "UV Package Manager: Not Found" even when a valid UV path override was configured in Advanced Settings. Root cause: PlatformDetectorBase.DetectUv() only searched PATH with bare command names ("uvx", "uv") and never consulted PathResolverService which respects the user's override setting. Changes: - Refactor DetectUv() to use PathResolverService.GetUvxPath() which checks override path first, then system PATH, then falls back to "uvx" - Add TryValidateUvExecutable() to verify executables by running --version instead of just checking File.Exists - Prioritize PATH environment variable in EnumerateUvxCandidates() for better compatibility with official uv install scripts - Fix process output read order (ReadToEnd before WaitForExit) to prevent potential deadlocks Co-Authored-By: ChatGLM 4.7 <noreply@zhipuai.com> * fix: improve uv/uvx detection robustness on macOS and Linux - Read both stdout and stderr when validating uv/uvx executables - Respect WaitForExit timeout return value instead of ignoring it - Fix version parsing to handle extra tokens like "(Homebrew 2025-01-01)" - Resolve bare commands ("uv"/"uvx") to absolute paths after validation - Rename FindExecutableInPath to FindUvxExecutableInPath for clarity Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * refactor: unify process execution with ExecPath.TryRun and add Windows PATH augmentation Replace direct Process.Start calls with ExecPath.TryRun across all platform detectors. This change: - Fixes potential deadlocks by using async output reading - Adds proper timeout handling with process termination - Removes redundant fallback logic and simplifies version parsing - Adds Windows PATH augmentation with common uv, npm, and Python installation paths Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * fix: improve version parsing to handle both spaces and parentheses The version extraction logic now properly handles outputs like: - "uvx 0.9.18" -> "0.9.18" - "uvx 0.9.18 (hash date)" -> "0.9.18" - "uvx 0.9.18 extra info" -> "0.9.18" Uses Math.Min to find the first occurrence of either space or parenthesis. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * refactor: improve platform detectors with absolute path resolution - Add absolute path resolution in TryValidatePython and TryValidateUvWithPath for better UI display - Fix BuildAugmentedPath to avoid PATH duplication - Add comprehensive comments for version parsing logic - Ensure cross-platform consistency across all three detectors - Fix override path validation logic with clear state handling - Fix platform detector path resolution and Python version detection - Use UserProfile consistently in GetClaudeCliPath instead of Personal - All platforms now use protected BuildAugmentedPath method This change improves user experience by displaying full paths in the UI while maintaining robust fallback behavior if path resolution fails. Co-Authored-By: GLM4.7 <noreply@zhipuai.com> * fix: improve error handling in PathResolverService by logging exceptions * Remove .meta files added after fork and update .gitignore * Update .gitignore * save .meta * refactor: unify uv/uvx naming and path detection across platforms - Rename TryValidateUvExecutable -> TryValidateUvxExecutable for consistency - Add cross-platform FindInPath() helper in ExecPath.cs - Remove platform-specific where/which implementations in favor of unified helper - Add Windows-specific DetectUv() override with enhanced uv/uvx detection - Add WinGet shim path support for Windows uvx installation - Update UI labels: "UV Path" -> "UVX Path" - Only show uvx path status when override is configured Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * fix: improve validation light(uvxPathStatus) logic for UVX path overrides and system paths * refactor: streamline UV version validation and unify path detection methods across platform detectors * fix: add type handling for Claude Code client in config JSON builder * fix: correct command from 'uvx' to 'uv' for Python version listing in WindowsPlatformDetector * feat: add uvx path fallback with warning UI - When override path is invalid, automatically fall back to system path - Add HasUvxPathFallback flag to track fallback state - Show yellow warning indicator when using fallback - Display clear messages for invalid override paths - Updated all platform detectors (Windows, macOS, Linux) to support fallback logic * refactor: remove GetDetails method from PlatformDetectorBase * Update ExecPath.cs update Windows Path lookup --------- Co-authored-by: ChatGLM 4.7 <noreply@zhipuai.com> Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com> Co-authored-by: Shutong Wu <51266340+Scriptwonder@users.noreply.github.com>
2026-01-18 07:34:40 +08:00
throw new InvalidOperationException("uvx not found. Install uv/uvx or set the override in Advanced Settings.");
}
return uvx;
}
protected string CurrentOsPath()
{
if (RuntimeInformation.IsOSPlatform(OSPlatform.Windows))
return client.windowsConfigPath;
if (RuntimeInformation.IsOSPlatform(OSPlatform.OSX))
return client.macConfigPath;
return client.linuxConfigPath;
}
protected bool UrlsEqual(string a, string b)
{
if (string.IsNullOrWhiteSpace(a) || string.IsNullOrWhiteSpace(b))
{
return false;
}
if (Uri.TryCreate(a.Trim(), UriKind.Absolute, out var uriA) &&
Uri.TryCreate(b.Trim(), UriKind.Absolute, out var uriB))
{
return Uri.Compare(
uriA,
uriB,
UriComponents.HttpRequestUrl,
UriFormat.SafeUnescaped,
StringComparison.OrdinalIgnoreCase) == 0;
}
string Normalize(string value) => value.Trim().TrimEnd('/');
return string.Equals(Normalize(a), Normalize(b), StringComparison.OrdinalIgnoreCase);
}
}
/// <summary>JSON-file based configurator (Cursor, Windsurf, VS Code, etc.).</summary>
public abstract class JsonFileMcpConfigurator : McpClientConfiguratorBase
{
public JsonFileMcpConfigurator(McpClient client) : base(client) { }
public override string GetConfigPath() => CurrentOsPath();
public override McpStatus CheckStatus(bool attemptAutoRewrite = true)
{
try
{
string path = GetConfigPath();
if (!File.Exists(path))
{
client.SetStatus(McpStatus.NotConfigured);
return client.status;
}
string configJson = File.ReadAllText(path);
string[] args = null;
string configuredUrl = null;
bool configExists = false;
if (client.IsVsCodeLayout)
{
var vsConfig = JsonConvert.DeserializeObject<JToken>(configJson) as JObject;
if (vsConfig != null)
{
var unityToken =
vsConfig["servers"]?["unityMCP"]
?? vsConfig["mcp"]?["servers"]?["unityMCP"];
if (unityToken is JObject unityObj)
{
configExists = true;
var argsToken = unityObj["args"];
if (argsToken is JArray)
{
args = argsToken.ToObject<string[]>();
}
var urlToken = unityObj["url"] ?? unityObj["serverUrl"];
if (urlToken != null && urlToken.Type != JTokenType.Null)
{
configuredUrl = urlToken.ToString();
}
}
}
}
else
{
McpConfig standardConfig = JsonConvert.DeserializeObject<McpConfig>(configJson);
if (standardConfig?.mcpServers?.unityMCP != null)
{
args = standardConfig.mcpServers.unityMCP.args;
configExists = true;
}
}
if (!configExists)
{
client.SetStatus(McpStatus.MissingConfig);
return client.status;
}
bool matches = false;
if (args != null && args.Length > 0)
{
string expectedUvxUrl = AssetPathUtility.GetMcpServerPackageSource();
string configuredUvxUrl = McpConfigurationHelper.ExtractUvxUrl(args);
matches = !string.IsNullOrEmpty(configuredUvxUrl) &&
McpConfigurationHelper.PathsEqual(configuredUvxUrl, expectedUvxUrl);
}
else if (!string.IsNullOrEmpty(configuredUrl))
{
string expectedUrl = HttpEndpointUtility.GetMcpRpcUrl();
matches = UrlsEqual(configuredUrl, expectedUrl);
}
if (matches)
{
client.SetStatus(McpStatus.Configured);
return client.status;
}
if (attemptAutoRewrite)
{
var result = McpConfigurationHelper.WriteMcpConfiguration(path, client);
if (result == "Configured successfully")
{
client.SetStatus(McpStatus.Configured);
}
else
{
client.SetStatus(McpStatus.IncorrectPath);
}
}
else
{
client.SetStatus(McpStatus.IncorrectPath);
}
}
catch (Exception ex)
{
client.SetStatus(McpStatus.Error, ex.Message);
}
return client.status;
}
public override void Configure()
{
string path = GetConfigPath();
McpConfigurationHelper.EnsureConfigDirectoryExists(path);
string result = McpConfigurationHelper.WriteMcpConfiguration(path, client);
if (result == "Configured successfully")
{
client.SetStatus(McpStatus.Configured);
}
else
{
throw new InvalidOperationException(result);
}
}
public override string GetManualSnippet()
{
try
{
string uvx = GetUvxPathOrError();
return ConfigJsonBuilder.BuildManualConfigJson(uvx, client);
}
catch (Exception ex)
{
var errorObj = new { error = ex.Message };
return JsonConvert.SerializeObject(errorObj);
}
}
public override IList<string> GetInstallationSteps() => new List<string> { "Configuration steps not available for this client." };
}
/// <summary>Codex (TOML) configurator.</summary>
public abstract class CodexMcpConfigurator : McpClientConfiguratorBase
{
public CodexMcpConfigurator(McpClient client) : base(client) { }
public override string GetConfigPath() => CurrentOsPath();
public override McpStatus CheckStatus(bool attemptAutoRewrite = true)
{
try
{
string path = GetConfigPath();
if (!File.Exists(path))
{
client.SetStatus(McpStatus.NotConfigured);
return client.status;
}
string toml = File.ReadAllText(path);
if (CodexConfigHelper.TryParseCodexServer(toml, out _, out var args, out var url))
{
bool matches = false;
if (!string.IsNullOrEmpty(url))
{
matches = UrlsEqual(url, HttpEndpointUtility.GetMcpRpcUrl());
}
else if (args != null && args.Length > 0)
{
string expected = AssetPathUtility.GetMcpServerPackageSource();
string configured = McpConfigurationHelper.ExtractUvxUrl(args);
matches = !string.IsNullOrEmpty(configured) &&
McpConfigurationHelper.PathsEqual(configured, expected);
}
if (matches)
{
client.SetStatus(McpStatus.Configured);
return client.status;
}
}
if (attemptAutoRewrite)
{
string result = McpConfigurationHelper.ConfigureCodexClient(path, client);
if (result == "Configured successfully")
{
client.SetStatus(McpStatus.Configured);
}
else
{
client.SetStatus(McpStatus.IncorrectPath);
}
}
else
{
client.SetStatus(McpStatus.IncorrectPath);
}
}
catch (Exception ex)
{
client.SetStatus(McpStatus.Error, ex.Message);
}
return client.status;
}
public override void Configure()
{
string path = GetConfigPath();
McpConfigurationHelper.EnsureConfigDirectoryExists(path);
string result = McpConfigurationHelper.ConfigureCodexClient(path, client);
if (result == "Configured successfully")
{
client.SetStatus(McpStatus.Configured);
}
else
{
throw new InvalidOperationException(result);
}
}
public override string GetManualSnippet()
{
try
{
string uvx = GetUvxPathOrError();
return CodexConfigHelper.BuildCodexServerBlock(uvx);
}
catch (Exception ex)
{
return $"# error: {ex.Message}";
}
}
public override IList<string> GetInstallationSteps() => new List<string>
{
"Run 'codex config edit' or open the config path",
"Paste the TOML",
"Save and restart Codex"
};
}
/// <summary>CLI-based configurator (Claude Code).</summary>
public abstract class ClaudeCliMcpConfigurator : McpClientConfiguratorBase
{
public ClaudeCliMcpConfigurator(McpClient client) : base(client) { }
public override bool SupportsAutoConfigure => true;
public override string GetConfigureActionLabel() => client.status == McpStatus.Configured ? "Unregister" : "Register";
public override string GetConfigPath() => "Managed via Claude CLI";
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
/// <summary>
/// Checks the Claude CLI registration status.
/// MUST be called from the main Unity thread due to EditorPrefs and Application.dataPath access.
/// </summary>
public override McpStatus CheckStatus(bool attemptAutoRewrite = true)
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
{
// Capture main-thread-only values before delegating to thread-safe method
string projectDir = Path.GetDirectoryName(Application.dataPath);
bool useHttpTransport = EditorPrefs.GetBool(EditorPrefKeys.UseHttpTransport, true);
// Resolve claudePath on the main thread (EditorPrefs access)
string claudePath = MCPServiceLocator.Paths.GetClaudeCliPath();
return CheckStatusWithProjectDir(projectDir, useHttpTransport, claudePath, attemptAutoRewrite);
}
/// <summary>
/// Internal thread-safe version of CheckStatus.
/// Can be called from background threads because all main-thread-only values are passed as parameters.
/// projectDir, useHttpTransport, and claudePath are REQUIRED (non-nullable) to enforce thread safety at compile time.
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
/// NOTE: attemptAutoRewrite is NOT fully thread-safe because Configure() requires the main thread.
/// When called from a background thread, pass attemptAutoRewrite=false and handle re-registration
/// on the main thread based on the returned status.
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
/// </summary>
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
internal McpStatus CheckStatusWithProjectDir(string projectDir, bool useHttpTransport, string claudePath, bool attemptAutoRewrite = false)
{
try
{
if (string.IsNullOrEmpty(claudePath))
{
client.SetStatus(McpStatus.NotConfigured, "Claude CLI not found");
return client.status;
}
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
// projectDir is required - no fallback to Application.dataPath
if (string.IsNullOrEmpty(projectDir))
{
throw new ArgumentNullException(nameof(projectDir), "Project directory must be provided for thread-safe execution");
}
string pathPrepend = null;
if (Application.platform == RuntimePlatform.OSXEditor)
{
pathPrepend = "/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin";
}
else if (Application.platform == RuntimePlatform.LinuxEditor)
{
pathPrepend = "/usr/local/bin:/usr/bin:/bin";
}
try
{
string claudeDir = Path.GetDirectoryName(claudePath);
if (!string.IsNullOrEmpty(claudeDir))
{
pathPrepend = string.IsNullOrEmpty(pathPrepend)
? claudeDir
: $"{claudeDir}:{pathPrepend}";
}
}
catch { }
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
// Check if UnityMCP exists (handles both "UnityMCP" and legacy "unityMCP")
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
if (ExecPath.TryRun(claudePath, "mcp list", projectDir, out var listStdout, out var listStderr, 10000, pathPrepend))
{
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
if (!string.IsNullOrEmpty(listStdout) && listStdout.IndexOf("UnityMCP", StringComparison.OrdinalIgnoreCase) >= 0)
{
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
// UnityMCP is registered - now verify transport mode matches
// useHttpTransport parameter is required (non-nullable) to ensure thread safety
bool currentUseHttp = useHttpTransport;
// Get detailed info about the registration to check transport type
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
// Try both "UnityMCP" and "unityMCP" (legacy naming)
string getStdout = null, getStderr = null;
bool gotInfo = ExecPath.TryRun(claudePath, "mcp get UnityMCP", projectDir, out getStdout, out getStderr, 7000, pathPrepend)
|| ExecPath.TryRun(claudePath, "mcp get unityMCP", projectDir, out getStdout, out getStderr, 7000, pathPrepend);
if (gotInfo)
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
{
// Parse the output to determine registered transport mode
// The CLI output format contains "Type: http" or "Type: stdio"
bool registeredWithHttp = getStdout.Contains("Type: http", StringComparison.OrdinalIgnoreCase);
bool registeredWithStdio = getStdout.Contains("Type: stdio", StringComparison.OrdinalIgnoreCase);
// Check for transport mismatch
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
bool hasTransportMismatch = (currentUseHttp && registeredWithStdio) || (!currentUseHttp && registeredWithHttp);
// For stdio transport, also check package version
bool hasVersionMismatch = false;
string configuredPackageSource = null;
string expectedPackageSource = null;
if (registeredWithStdio)
{
expectedPackageSource = AssetPathUtility.GetMcpServerPackageSource();
configuredPackageSource = ExtractPackageSourceFromCliOutput(getStdout);
hasVersionMismatch = !string.IsNullOrEmpty(configuredPackageSource) &&
!string.Equals(configuredPackageSource, expectedPackageSource, StringComparison.OrdinalIgnoreCase);
}
// If there's any mismatch and auto-rewrite is enabled, re-register
if (hasTransportMismatch || hasVersionMismatch)
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
{
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
// Configure() requires main thread (accesses EditorPrefs, Application.dataPath)
// Only attempt auto-rewrite if we're on the main thread
bool isMainThread = System.Threading.Thread.CurrentThread.ManagedThreadId == 1;
if (attemptAutoRewrite && isMainThread)
{
string reason = hasTransportMismatch
? $"Transport mismatch (registered: {(registeredWithHttp ? "HTTP" : "stdio")}, expected: {(currentUseHttp ? "HTTP" : "stdio")})"
: $"Package version mismatch (registered: {configuredPackageSource}, expected: {expectedPackageSource})";
McpLog.Info($"{reason}. Re-registering...");
try
{
// Force re-register by ensuring status is not Configured (which would toggle to Unregister)
client.SetStatus(McpStatus.IncorrectPath);
Configure();
return client.status;
}
catch (Exception ex)
{
McpLog.Warn($"Auto-reregister failed: {ex.Message}");
client.SetStatus(McpStatus.IncorrectPath, $"Configuration mismatch. Click Configure to re-register.");
return client.status;
}
}
else
{
if (hasTransportMismatch)
{
string errorMsg = $"Transport mismatch: Claude Code is registered with {(registeredWithHttp ? "HTTP" : "stdio")} but current setting is {(currentUseHttp ? "HTTP" : "stdio")}. Click Configure to re-register.";
client.SetStatus(McpStatus.Error, errorMsg);
McpLog.Warn(errorMsg);
}
else
{
client.SetStatus(McpStatus.IncorrectPath, $"Package version mismatch: registered with '{configuredPackageSource}' but current version is '{expectedPackageSource}'.");
}
return client.status;
}
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
}
}
client.SetStatus(McpStatus.Configured);
return client.status;
}
}
client.SetStatus(McpStatus.NotConfigured);
}
catch (Exception ex)
{
client.SetStatus(McpStatus.Error, ex.Message);
}
return client.status;
}
public override void Configure()
{
if (client.status == McpStatus.Configured)
{
Unregister();
}
else
{
Register();
}
}
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
/// <summary>
/// Thread-safe version of Configure that uses pre-captured main-thread values.
/// All parameters must be captured on the main thread before calling this method.
/// </summary>
public void ConfigureWithCapturedValues(
string projectDir, string claudePath, string pathPrepend,
bool useHttpTransport, string httpUrl,
string uvxPath, string gitUrl, string packageName, bool shouldForceRefresh)
{
if (client.status == McpStatus.Configured)
{
UnregisterWithCapturedValues(projectDir, claudePath, pathPrepend);
}
else
{
RegisterWithCapturedValues(projectDir, claudePath, pathPrepend,
useHttpTransport, httpUrl, uvxPath, gitUrl, packageName, shouldForceRefresh);
}
}
/// <summary>
/// Thread-safe registration using pre-captured values.
/// </summary>
private void RegisterWithCapturedValues(
string projectDir, string claudePath, string pathPrepend,
bool useHttpTransport, string httpUrl,
string uvxPath, string gitUrl, string packageName, bool shouldForceRefresh)
{
if (string.IsNullOrEmpty(claudePath))
{
throw new InvalidOperationException("Claude CLI not found. Please install Claude Code first.");
}
string args;
if (useHttpTransport)
{
args = $"mcp add --transport http UnityMCP {httpUrl}";
}
else
{
// Note: --reinstall is not supported by uvx, use --no-cache --refresh instead
string devFlags = shouldForceRefresh ? "--no-cache --refresh " : string.Empty;
args = $"mcp add --transport stdio UnityMCP -- \"{uvxPath}\" {devFlags}--from \"{gitUrl}\" {packageName}";
}
// Remove any existing registrations - handle both "UnityMCP" and "unityMCP" (legacy)
McpLog.Info("Removing any existing UnityMCP registrations before adding...");
ExecPath.TryRun(claudePath, "mcp remove UnityMCP", projectDir, out _, out _, 7000, pathPrepend);
ExecPath.TryRun(claudePath, "mcp remove unityMCP", projectDir, out _, out _, 7000, pathPrepend);
// Now add the registration
if (!ExecPath.TryRun(claudePath, args, projectDir, out var stdout, out var stderr, 15000, pathPrepend))
{
throw new InvalidOperationException($"Failed to register with Claude Code:\n{stderr}\n{stdout}");
}
McpLog.Info($"Successfully registered with Claude Code using {(useHttpTransport ? "HTTP" : "stdio")} transport.");
client.SetStatus(McpStatus.Configured);
}
/// <summary>
/// Thread-safe unregistration using pre-captured values.
/// </summary>
private void UnregisterWithCapturedValues(string projectDir, string claudePath, string pathPrepend)
{
if (string.IsNullOrEmpty(claudePath))
{
throw new InvalidOperationException("Claude CLI not found. Please install Claude Code first.");
}
// Remove both "UnityMCP" and "unityMCP" (legacy naming)
McpLog.Info("Removing all UnityMCP registrations...");
ExecPath.TryRun(claudePath, "mcp remove UnityMCP", projectDir, out _, out _, 7000, pathPrepend);
ExecPath.TryRun(claudePath, "mcp remove unityMCP", projectDir, out _, out _, 7000, pathPrepend);
McpLog.Info("MCP server successfully unregistered from Claude Code.");
client.SetStatus(McpStatus.NotConfigured);
}
private void Register()
{
var pathService = MCPServiceLocator.Paths;
string claudePath = pathService.GetClaudeCliPath();
if (string.IsNullOrEmpty(claudePath))
{
throw new InvalidOperationException("Claude CLI not found. Please install Claude Code first.");
}
bool useHttpTransport = EditorPrefs.GetBool(EditorPrefKeys.UseHttpTransport, true);
string args;
if (useHttpTransport)
{
string httpUrl = HttpEndpointUtility.GetMcpRpcUrl();
args = $"mcp add --transport http UnityMCP {httpUrl}";
}
else
{
var (uvxPath, gitUrl, packageName) = AssetPathUtility.GetUvxCommandParts();
Fix HTTP/Stdio Transport UX and Test Bug (#530) * refactor: Split ParseColorOrDefault into two overloads and change default to Color.white * Auto-format Python code * Remove unused Python module * Refactored VFX functionality into multiple files Tested everything, works like a charm * Rename ManageVfx folder to just Vfx We know what it's managing * Clean up whitespace on plugin tools and resources * Make ManageGameObject less of a monolith by splitting it out into different files * Remove obsolete FindObjectByInstruction method We also update the namespace for ManageVFX * Add local test harness for fast developer iteration Scripts for running the NL/T/GO test suites locally against a GUI Unity Editor, complementing the CI workflows in .github/workflows/. Benefits: - 10-100x faster than CI (no Docker startup) - Real-time Unity console debugging - Single test execution for rapid iteration - Auto-detects HTTP vs stdio transport Usage: ./scripts/local-test/setup.sh # One-time setup ./scripts/local-test/quick-test.sh NL-0 # Run single test ./scripts/local-test/run-nl-suite-local.sh # Full suite See scripts/local-test/README.md for details. Also updated .gitignore to: - Allow scripts/local-test/ to be tracked - Ignore generated artifacts (reports/*.xml, .claude/local/, .unity-mcp/) * Fix issue #525: Save dirty scenes for all test modes Move SaveDirtyScenesIfNeeded() call outside the PlayMode conditional so EditMode tests don't get blocked by Unity's "Save Scene" modal dialog. This prevents MCP from timing out when running EditMode tests with unsaved scene changes. * refactor: Consolidate editor state resources into single canonical implementation Merged EditorStateV2 into EditorState, making get_editor_state the canonical resource. Updated Unity C# to use EditorStateCache directly. Enhanced Python implementation with advice/staleness enrichment, external changes detection, and instance ID inference. Removed duplicate EditorStateV2 resource and legacy fallback mapping. * Validate editor state with Pydantic models in both C# and Python Added strongly-typed Pydantic models for EditorStateV2 schema in Python and corresponding C# classes with JsonProperty attributes. Updated C# to serialize using typed classes instead of anonymous objects. Python now validates the editor state payload before returning it, catching schema mismatches early. * Consolidate run_tests and run_tests_async into single async implementation Merged run_tests_async into run_tests, making async job-based execution the default behavior. Removed synchronous blocking test execution. Updated RunTests.cs to start test jobs immediately and return job_id for polling. Changed TestJobManager methods to internal visibility. Updated README to reflect single run_tests_async tool. Python implementation now uses async job pattern exclusively. * Validate test job responses with Pydantic models in Python * Change resources URI from unity:// to mcpforunity:// It should reduce conflicts with other Unity MCPs that users try, and to comply with Unity's requests regarding use of their company and product name * Update README with all tools + better listing for resources * Update other references to resources * Updated translated doc - unfortunately I cannot verify * Update the Chinese translation of the dev docks * Change menu item from Setup Window to Local Setup Window We now differentiate whether it's HTTP local or remote * Fix URIs for menu items and tests * Shouldn't have removed it * fix: add missing FAST_FAIL_TIMEOUT constant in PluginHub The FAST_FAIL_TIMEOUT class attribute was referenced on line 149 but never defined, causing AttributeError on every ping attempt. This error was silently caught by the broad 'except Exception' handler, causing all fast-fail commands (read_console, get_editor_state, ping) to fail after 6 seconds of retries with 'ping not answered' error. Added FAST_FAIL_TIMEOUT = 10 to define a 10-second timeout for fast-fail commands, matching the intent of the existing fast-fail infrastructure. * feat(ScriptableObject): enhance dry-run validation for AnimationCurve and Quaternion Dry-run validation now validates value formats, not just property existence: - AnimationCurve: Validates structure ({keys:[...]} or direct array), checks each keyframe is an object, validates numeric fields (time, value, inSlope, outSlope, inWeight, outWeight) and integer fields (weightedMode) - Quaternion: Validates array length (3 for Euler, 4 for raw) or object structure ({x,y,z,w} or {euler:[x,y,z]}), ensures all components are numeric Refactored shared validation helpers into appropriate locations: - ParamCoercion: IsNumericToken, ValidateNumericField, ValidateIntegerField - VectorParsing: ValidateAnimationCurveFormat, ValidateQuaternionFormat Added comprehensive XML documentation clarifying keyframe field defaults (all default to 0 except as noted). Added 5 new dry-run validation tests covering valid and invalid formats for both AnimationCurve and Quaternion properties. * test: fix integration tests after merge - test_refresh_unity_retry_recovery: Mock now handles both refresh_unity and get_editor_state commands (refresh_unity internally calls get_editor_state when wait_for_ready=True) - test_run_tests_async_forwards_params: Mock response now includes required 'mode' field for RunTestsStartResponse Pydantic validation - test_get_test_job_forwards_job_id: Updated to handle GetTestJobResponse as Pydantic model instead of dict (use model_dump() for assertions) * Update warning message to apply to all test modes Follow-up to PR #527: Since SaveDirtyScenesIfNeeded() now runs for all test modes, update the warning message to say 'tests' instead of 'PlayMode tests'. * feat(run_tests): add wait_timeout to get_test_job to avoid client loop detection When polling for test completion, MCP clients like Cursor can detect the repeated get_test_job calls as 'looping' and terminate the agent. Added wait_timeout parameter that makes the server wait internally for tests to complete (polling Unity every 2s) before returning. This dramatically reduces client-side tool calls from 10-20 down to 1-2, avoiding loop detection. Usage: get_test_job(job_id='xxx', wait_timeout=30) - Returns immediately if tests complete within timeout - Returns current status if timeout expires (client can call again) - Recommended: 30-60 seconds * fix: use Pydantic attribute access in test_run_tests_async for merge compatibility * revert: remove local test harness - will be submitted in separate PR * fix: stdio transport survives test runs without UI flicker Root cause: WriteToConfigTests.TearDown() was unconditionally deleting UseHttpTransport EditorPref even when tests were skipped on Windows (NUnit runs TearDown even after Assert.Ignore). Changes: - Fix WriteToConfigTests to save/restore prefs instead of deleting - Add centralized ShouldForceUvxRefresh() for local dev path detection - Clean stale Python build/ artifacts before client configuration - Improve reload handler flag management to prevent stuck Resuming state - Show Resuming status during stdio bridge restart - Initialize client config display on window open - Add DevModeForceServerRefresh to EditorPrefs window known types --------- Co-authored-by: Marcus Sanatan <msanatan@gmail.com> Co-authored-by: Scott Jennings <scott.jennings+CIGINT@cloudimperiumgames.com>
2026-01-08 11:33:22 +08:00
// Use central helper that checks both DevModeForceServerRefresh AND local path detection.
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
// Note: --reinstall is not supported by uvx, use --no-cache --refresh instead
Fix HTTP/Stdio Transport UX and Test Bug (#530) * refactor: Split ParseColorOrDefault into two overloads and change default to Color.white * Auto-format Python code * Remove unused Python module * Refactored VFX functionality into multiple files Tested everything, works like a charm * Rename ManageVfx folder to just Vfx We know what it's managing * Clean up whitespace on plugin tools and resources * Make ManageGameObject less of a monolith by splitting it out into different files * Remove obsolete FindObjectByInstruction method We also update the namespace for ManageVFX * Add local test harness for fast developer iteration Scripts for running the NL/T/GO test suites locally against a GUI Unity Editor, complementing the CI workflows in .github/workflows/. Benefits: - 10-100x faster than CI (no Docker startup) - Real-time Unity console debugging - Single test execution for rapid iteration - Auto-detects HTTP vs stdio transport Usage: ./scripts/local-test/setup.sh # One-time setup ./scripts/local-test/quick-test.sh NL-0 # Run single test ./scripts/local-test/run-nl-suite-local.sh # Full suite See scripts/local-test/README.md for details. Also updated .gitignore to: - Allow scripts/local-test/ to be tracked - Ignore generated artifacts (reports/*.xml, .claude/local/, .unity-mcp/) * Fix issue #525: Save dirty scenes for all test modes Move SaveDirtyScenesIfNeeded() call outside the PlayMode conditional so EditMode tests don't get blocked by Unity's "Save Scene" modal dialog. This prevents MCP from timing out when running EditMode tests with unsaved scene changes. * refactor: Consolidate editor state resources into single canonical implementation Merged EditorStateV2 into EditorState, making get_editor_state the canonical resource. Updated Unity C# to use EditorStateCache directly. Enhanced Python implementation with advice/staleness enrichment, external changes detection, and instance ID inference. Removed duplicate EditorStateV2 resource and legacy fallback mapping. * Validate editor state with Pydantic models in both C# and Python Added strongly-typed Pydantic models for EditorStateV2 schema in Python and corresponding C# classes with JsonProperty attributes. Updated C# to serialize using typed classes instead of anonymous objects. Python now validates the editor state payload before returning it, catching schema mismatches early. * Consolidate run_tests and run_tests_async into single async implementation Merged run_tests_async into run_tests, making async job-based execution the default behavior. Removed synchronous blocking test execution. Updated RunTests.cs to start test jobs immediately and return job_id for polling. Changed TestJobManager methods to internal visibility. Updated README to reflect single run_tests_async tool. Python implementation now uses async job pattern exclusively. * Validate test job responses with Pydantic models in Python * Change resources URI from unity:// to mcpforunity:// It should reduce conflicts with other Unity MCPs that users try, and to comply with Unity's requests regarding use of their company and product name * Update README with all tools + better listing for resources * Update other references to resources * Updated translated doc - unfortunately I cannot verify * Update the Chinese translation of the dev docks * Change menu item from Setup Window to Local Setup Window We now differentiate whether it's HTTP local or remote * Fix URIs for menu items and tests * Shouldn't have removed it * fix: add missing FAST_FAIL_TIMEOUT constant in PluginHub The FAST_FAIL_TIMEOUT class attribute was referenced on line 149 but never defined, causing AttributeError on every ping attempt. This error was silently caught by the broad 'except Exception' handler, causing all fast-fail commands (read_console, get_editor_state, ping) to fail after 6 seconds of retries with 'ping not answered' error. Added FAST_FAIL_TIMEOUT = 10 to define a 10-second timeout for fast-fail commands, matching the intent of the existing fast-fail infrastructure. * feat(ScriptableObject): enhance dry-run validation for AnimationCurve and Quaternion Dry-run validation now validates value formats, not just property existence: - AnimationCurve: Validates structure ({keys:[...]} or direct array), checks each keyframe is an object, validates numeric fields (time, value, inSlope, outSlope, inWeight, outWeight) and integer fields (weightedMode) - Quaternion: Validates array length (3 for Euler, 4 for raw) or object structure ({x,y,z,w} or {euler:[x,y,z]}), ensures all components are numeric Refactored shared validation helpers into appropriate locations: - ParamCoercion: IsNumericToken, ValidateNumericField, ValidateIntegerField - VectorParsing: ValidateAnimationCurveFormat, ValidateQuaternionFormat Added comprehensive XML documentation clarifying keyframe field defaults (all default to 0 except as noted). Added 5 new dry-run validation tests covering valid and invalid formats for both AnimationCurve and Quaternion properties. * test: fix integration tests after merge - test_refresh_unity_retry_recovery: Mock now handles both refresh_unity and get_editor_state commands (refresh_unity internally calls get_editor_state when wait_for_ready=True) - test_run_tests_async_forwards_params: Mock response now includes required 'mode' field for RunTestsStartResponse Pydantic validation - test_get_test_job_forwards_job_id: Updated to handle GetTestJobResponse as Pydantic model instead of dict (use model_dump() for assertions) * Update warning message to apply to all test modes Follow-up to PR #527: Since SaveDirtyScenesIfNeeded() now runs for all test modes, update the warning message to say 'tests' instead of 'PlayMode tests'. * feat(run_tests): add wait_timeout to get_test_job to avoid client loop detection When polling for test completion, MCP clients like Cursor can detect the repeated get_test_job calls as 'looping' and terminate the agent. Added wait_timeout parameter that makes the server wait internally for tests to complete (polling Unity every 2s) before returning. This dramatically reduces client-side tool calls from 10-20 down to 1-2, avoiding loop detection. Usage: get_test_job(job_id='xxx', wait_timeout=30) - Returns immediately if tests complete within timeout - Returns current status if timeout expires (client can call again) - Recommended: 30-60 seconds * fix: use Pydantic attribute access in test_run_tests_async for merge compatibility * revert: remove local test harness - will be submitted in separate PR * fix: stdio transport survives test runs without UI flicker Root cause: WriteToConfigTests.TearDown() was unconditionally deleting UseHttpTransport EditorPref even when tests were skipped on Windows (NUnit runs TearDown even after Assert.Ignore). Changes: - Fix WriteToConfigTests to save/restore prefs instead of deleting - Add centralized ShouldForceUvxRefresh() for local dev path detection - Clean stale Python build/ artifacts before client configuration - Improve reload handler flag management to prevent stuck Resuming state - Show Resuming status during stdio bridge restart - Initialize client config display on window open - Add DevModeForceServerRefresh to EditorPrefs window known types --------- Co-authored-by: Marcus Sanatan <msanatan@gmail.com> Co-authored-by: Scott Jennings <scott.jennings+CIGINT@cloudimperiumgames.com>
2026-01-08 11:33:22 +08:00
string devFlags = AssetPathUtility.ShouldForceUvxRefresh() ? "--no-cache --refresh " : string.Empty;
Payload-safe paging for hierarchy/components + safer asset search + docs (#490) * Fix test teardown to avoid dropping MCP bridge CodexConfigHelperTests was calling MCPServiceLocator.Reset() in TearDown, which disposes the active bridge/transport during MCP-driven test runs. Replace with restoring only the mutated service (IPlatformService). * Avoid leaking PlatformService in CodexConfigHelperTests Capture the original IPlatformService before this fixture runs and restore it in TearDown. This preserves the MCP connection safety fix (no MCPServiceLocator.Reset()) while avoiding global state leakage to subsequent tests. * Fix SO MCP tooling: validate folder roots, normalize paths, expand tests; remove vestigial SO tools * Remove UnityMCPTests stress artifacts and ignore Assets/Temp * Ignore UnityMCPTests Assets/Temp only * Clarify array_resize fallback logic comments * Refactor: simplify action set and reuse slash sanitization * Enhance: preserve GUID on overwrite & support Vector/Color types in ScriptableObject tools * Fix: ensure asset name matches filename to suppress Unity warnings * Fix: resolve Unity warnings by ensuring asset name match and removing redundant import * Refactor: Validate assetName, strict object parsing for vectors, remove broken SO logic from ManageAsset * Hardening: reject Windows drive paths; clarify supported asset types * Delete FixscriptableobjecPlan.md * Paginate get_hierarchy and get_components to prevent large payload crashes * dev: add uvx dev-mode refresh + safer HTTP stop; fix server typing eval * Payload-safe paging defaults + docs; harden asset search; stabilize Codex tests * chore: align uvx args + coercion helpers; tighten safety guidance * chore: minor cleanup + stabilize EditMode SO tests
2025-12-29 12:57:57 +08:00
args = $"mcp add --transport stdio UnityMCP -- \"{uvxPath}\" {devFlags}--from \"{gitUrl}\" {packageName}";
}
string projectDir = Path.GetDirectoryName(Application.dataPath);
string pathPrepend = null;
if (Application.platform == RuntimePlatform.OSXEditor)
{
pathPrepend = "/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin";
}
else if (Application.platform == RuntimePlatform.LinuxEditor)
{
pathPrepend = "/usr/local/bin:/usr/bin:/bin";
}
try
{
string claudeDir = Path.GetDirectoryName(claudePath);
if (!string.IsNullOrEmpty(claudeDir))
{
pathPrepend = string.IsNullOrEmpty(pathPrepend)
? claudeDir
: $"{claudeDir}:{pathPrepend}";
}
}
catch { }
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
// Remove any existing registrations - handle both "UnityMCP" and "unityMCP" (legacy)
McpLog.Info("Removing any existing UnityMCP registrations before adding...");
ExecPath.TryRun(claudePath, "mcp remove UnityMCP", projectDir, out _, out _, 7000, pathPrepend);
ExecPath.TryRun(claudePath, "mcp remove unityMCP", projectDir, out _, out _, 7000, pathPrepend);
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
// Now add the registration with the current transport mode
if (!ExecPath.TryRun(claudePath, args, projectDir, out var stdout, out var stderr, 15000, pathPrepend))
{
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
throw new InvalidOperationException($"Failed to register with Claude Code:\n{stderr}\n{stdout}");
}
HTTP setup overhaul: transport selection (HTTP local/remote vs stdio), safer lifecycle, cleaner UI, better Claude Code integration (#499) * Avoid blocking Claude CLI status checks on focus * Fix Claude Code registration to remove existing server before re-registering When registering with Claude Code, if a UnityMCP server already exists, remove it first before adding the new registration. This ensures the transport mode (HTTP vs stdio) is always updated to match the current UseHttpTransport EditorPref setting. Previously, if a stdio registration existed and the user tried to register with HTTP, the command would fail with 'already exists' and the old stdio configuration would remain unchanged. * Fix Claude Code transport validation to parse CLI output format correctly The validation code was incorrectly parsing the output of 'claude mcp get UnityMCP' by looking for JSON format ("transport": "http"), but the CLI actually returns human-readable text format ("Type: http"). This caused the transport mismatch detection to never trigger, allowing stdio to be selected in the UI while HTTP was registered with Claude Code. Changes: - Fix parsing logic to check for "Type: http" or "Type: stdio" in CLI output - Add OnTransportChanged event to refresh client status when transport changes - Wire up event handler to trigger client status refresh on transport dropdown change This ensures that when the transport mode in Unity doesn't match what's registered with Claude Code, the UI will correctly show an error status with instructions to re-register. * Fix Claude Code registration UI blocking and thread safety issues This commit resolves three issues with Claude Code registration: 1. UI blocking: Removed synchronous CheckStatus() call after registration that was blocking the editor. Status is now set immediately with async verification happening in the background. 2. Thread safety: Fixed "can only be called from the main thread" errors by capturing Application.dataPath and EditorPrefs.GetBool() on the main thread before spawning async status check tasks. 3. Transport mismatch detection: Transport mode changes now trigger immediate status checks to detect HTTP/stdio mismatches, instead of waiting for the 45-second refresh interval. The registration button now turns green immediately after successful registration without blocking, and properly detects transport mismatches when switching between HTTP and stdio modes. * Enforce thread safety for Claude Code status checks at compile time Address code review feedback by making CheckStatusWithProjectDir thread-safe by design rather than by convention: 1. Made projectDir and useHttpTransport parameters non-nullable to prevent accidental background thread calls without captured values 2. Removed nullable fallback to EditorPrefs.GetBool() which would cause thread safety violations if called from background threads 3. Added ArgumentNullException for null projectDir instead of falling back to Application.dataPath (which is main-thread only) 4. Added XML documentation clearly stating threading contracts: - CheckStatus() must be called from main thread - CheckStatusWithProjectDir() is safe for background threads 5. Removed unreachable else branch in async status check code These changes make it impossible to misuse the API from background threads, with compile-time enforcement instead of runtime errors. * Consolidate local HTTP Start/Stop and auto-start session * HTTP improvements: Unity-owned server lifecycle + UI polish * Deterministic HTTP stop via pidfile+token; spawn server in terminal * Fix review feedback: token validation, host normalization, safer casts * Fix stop heuristics edge cases; remove dead pid capture * Fix unity substring guard in stop heuristics * Fix local server cleanup and connection checks * Fix read_console default limits; cleanup Unity-managed server vestiges * Fix unfocused reconnect stalls; fast-fail retryable Unity commands * Simplify PluginHub reload handling; honor run_tests timeout * Fix Windows Claude CLI status check threading
2026-01-02 09:08:51 +08:00
McpLog.Info($"Successfully registered with Claude Code using {(useHttpTransport ? "HTTP" : "stdio")} transport.");
// Set status to Configured immediately after successful registration
// The UI will trigger an async verification check separately to avoid blocking
client.SetStatus(McpStatus.Configured);
}
private void Unregister()
{
var pathService = MCPServiceLocator.Paths;
string claudePath = pathService.GetClaudeCliPath();
if (string.IsNullOrEmpty(claudePath))
{
throw new InvalidOperationException("Claude CLI not found. Please install Claude Code first.");
}
string projectDir = Path.GetDirectoryName(Application.dataPath);
string pathPrepend = null;
if (Application.platform == RuntimePlatform.OSXEditor)
{
pathPrepend = "/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin";
}
else if (Application.platform == RuntimePlatform.LinuxEditor)
{
pathPrepend = "/usr/local/bin:/usr/bin:/bin";
}
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
// Remove both "UnityMCP" and "unityMCP" (legacy naming)
McpLog.Info("Removing all UnityMCP registrations...");
ExecPath.TryRun(claudePath, "mcp remove UnityMCP", projectDir, out _, out _, 7000, pathPrepend);
ExecPath.TryRun(claudePath, "mcp remove unityMCP", projectDir, out _, out _, 7000, pathPrepend);
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
McpLog.Info("MCP server successfully unregistered from Claude Code.");
client.SetStatus(McpStatus.NotConfigured);
}
public override string GetManualSnippet()
{
string uvxPath = MCPServiceLocator.Paths.GetUvxPath();
bool useHttpTransport = EditorPrefs.GetBool(EditorPrefKeys.UseHttpTransport, true);
if (useHttpTransport)
{
string httpUrl = HttpEndpointUtility.GetMcpRpcUrl();
return "# Register the MCP server with Claude Code:\n" +
$"claude mcp add --transport http UnityMCP {httpUrl}\n\n" +
"# Unregister the MCP server:\n" +
"claude mcp remove UnityMCP\n\n" +
"# List registered servers:\n" +
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
"claude mcp list";
}
if (string.IsNullOrEmpty(uvxPath))
{
return "# Error: Configuration not available - check paths in Advanced Settings";
}
string packageSource = AssetPathUtility.GetMcpServerPackageSource();
Fix HTTP/Stdio Transport UX and Test Bug (#530) * refactor: Split ParseColorOrDefault into two overloads and change default to Color.white * Auto-format Python code * Remove unused Python module * Refactored VFX functionality into multiple files Tested everything, works like a charm * Rename ManageVfx folder to just Vfx We know what it's managing * Clean up whitespace on plugin tools and resources * Make ManageGameObject less of a monolith by splitting it out into different files * Remove obsolete FindObjectByInstruction method We also update the namespace for ManageVFX * Add local test harness for fast developer iteration Scripts for running the NL/T/GO test suites locally against a GUI Unity Editor, complementing the CI workflows in .github/workflows/. Benefits: - 10-100x faster than CI (no Docker startup) - Real-time Unity console debugging - Single test execution for rapid iteration - Auto-detects HTTP vs stdio transport Usage: ./scripts/local-test/setup.sh # One-time setup ./scripts/local-test/quick-test.sh NL-0 # Run single test ./scripts/local-test/run-nl-suite-local.sh # Full suite See scripts/local-test/README.md for details. Also updated .gitignore to: - Allow scripts/local-test/ to be tracked - Ignore generated artifacts (reports/*.xml, .claude/local/, .unity-mcp/) * Fix issue #525: Save dirty scenes for all test modes Move SaveDirtyScenesIfNeeded() call outside the PlayMode conditional so EditMode tests don't get blocked by Unity's "Save Scene" modal dialog. This prevents MCP from timing out when running EditMode tests with unsaved scene changes. * refactor: Consolidate editor state resources into single canonical implementation Merged EditorStateV2 into EditorState, making get_editor_state the canonical resource. Updated Unity C# to use EditorStateCache directly. Enhanced Python implementation with advice/staleness enrichment, external changes detection, and instance ID inference. Removed duplicate EditorStateV2 resource and legacy fallback mapping. * Validate editor state with Pydantic models in both C# and Python Added strongly-typed Pydantic models for EditorStateV2 schema in Python and corresponding C# classes with JsonProperty attributes. Updated C# to serialize using typed classes instead of anonymous objects. Python now validates the editor state payload before returning it, catching schema mismatches early. * Consolidate run_tests and run_tests_async into single async implementation Merged run_tests_async into run_tests, making async job-based execution the default behavior. Removed synchronous blocking test execution. Updated RunTests.cs to start test jobs immediately and return job_id for polling. Changed TestJobManager methods to internal visibility. Updated README to reflect single run_tests_async tool. Python implementation now uses async job pattern exclusively. * Validate test job responses with Pydantic models in Python * Change resources URI from unity:// to mcpforunity:// It should reduce conflicts with other Unity MCPs that users try, and to comply with Unity's requests regarding use of their company and product name * Update README with all tools + better listing for resources * Update other references to resources * Updated translated doc - unfortunately I cannot verify * Update the Chinese translation of the dev docks * Change menu item from Setup Window to Local Setup Window We now differentiate whether it's HTTP local or remote * Fix URIs for menu items and tests * Shouldn't have removed it * fix: add missing FAST_FAIL_TIMEOUT constant in PluginHub The FAST_FAIL_TIMEOUT class attribute was referenced on line 149 but never defined, causing AttributeError on every ping attempt. This error was silently caught by the broad 'except Exception' handler, causing all fast-fail commands (read_console, get_editor_state, ping) to fail after 6 seconds of retries with 'ping not answered' error. Added FAST_FAIL_TIMEOUT = 10 to define a 10-second timeout for fast-fail commands, matching the intent of the existing fast-fail infrastructure. * feat(ScriptableObject): enhance dry-run validation for AnimationCurve and Quaternion Dry-run validation now validates value formats, not just property existence: - AnimationCurve: Validates structure ({keys:[...]} or direct array), checks each keyframe is an object, validates numeric fields (time, value, inSlope, outSlope, inWeight, outWeight) and integer fields (weightedMode) - Quaternion: Validates array length (3 for Euler, 4 for raw) or object structure ({x,y,z,w} or {euler:[x,y,z]}), ensures all components are numeric Refactored shared validation helpers into appropriate locations: - ParamCoercion: IsNumericToken, ValidateNumericField, ValidateIntegerField - VectorParsing: ValidateAnimationCurveFormat, ValidateQuaternionFormat Added comprehensive XML documentation clarifying keyframe field defaults (all default to 0 except as noted). Added 5 new dry-run validation tests covering valid and invalid formats for both AnimationCurve and Quaternion properties. * test: fix integration tests after merge - test_refresh_unity_retry_recovery: Mock now handles both refresh_unity and get_editor_state commands (refresh_unity internally calls get_editor_state when wait_for_ready=True) - test_run_tests_async_forwards_params: Mock response now includes required 'mode' field for RunTestsStartResponse Pydantic validation - test_get_test_job_forwards_job_id: Updated to handle GetTestJobResponse as Pydantic model instead of dict (use model_dump() for assertions) * Update warning message to apply to all test modes Follow-up to PR #527: Since SaveDirtyScenesIfNeeded() now runs for all test modes, update the warning message to say 'tests' instead of 'PlayMode tests'. * feat(run_tests): add wait_timeout to get_test_job to avoid client loop detection When polling for test completion, MCP clients like Cursor can detect the repeated get_test_job calls as 'looping' and terminate the agent. Added wait_timeout parameter that makes the server wait internally for tests to complete (polling Unity every 2s) before returning. This dramatically reduces client-side tool calls from 10-20 down to 1-2, avoiding loop detection. Usage: get_test_job(job_id='xxx', wait_timeout=30) - Returns immediately if tests complete within timeout - Returns current status if timeout expires (client can call again) - Recommended: 30-60 seconds * fix: use Pydantic attribute access in test_run_tests_async for merge compatibility * revert: remove local test harness - will be submitted in separate PR * fix: stdio transport survives test runs without UI flicker Root cause: WriteToConfigTests.TearDown() was unconditionally deleting UseHttpTransport EditorPref even when tests were skipped on Windows (NUnit runs TearDown even after Assert.Ignore). Changes: - Fix WriteToConfigTests to save/restore prefs instead of deleting - Add centralized ShouldForceUvxRefresh() for local dev path detection - Clean stale Python build/ artifacts before client configuration - Improve reload handler flag management to prevent stuck Resuming state - Show Resuming status during stdio bridge restart - Initialize client config display on window open - Add DevModeForceServerRefresh to EditorPrefs window known types --------- Co-authored-by: Marcus Sanatan <msanatan@gmail.com> Co-authored-by: Scott Jennings <scott.jennings+CIGINT@cloudimperiumgames.com>
2026-01-08 11:33:22 +08:00
// Use central helper that checks both DevModeForceServerRefresh AND local path detection.
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
// Note: --reinstall is not supported by uvx, use --no-cache --refresh instead
Fix HTTP/Stdio Transport UX and Test Bug (#530) * refactor: Split ParseColorOrDefault into two overloads and change default to Color.white * Auto-format Python code * Remove unused Python module * Refactored VFX functionality into multiple files Tested everything, works like a charm * Rename ManageVfx folder to just Vfx We know what it's managing * Clean up whitespace on plugin tools and resources * Make ManageGameObject less of a monolith by splitting it out into different files * Remove obsolete FindObjectByInstruction method We also update the namespace for ManageVFX * Add local test harness for fast developer iteration Scripts for running the NL/T/GO test suites locally against a GUI Unity Editor, complementing the CI workflows in .github/workflows/. Benefits: - 10-100x faster than CI (no Docker startup) - Real-time Unity console debugging - Single test execution for rapid iteration - Auto-detects HTTP vs stdio transport Usage: ./scripts/local-test/setup.sh # One-time setup ./scripts/local-test/quick-test.sh NL-0 # Run single test ./scripts/local-test/run-nl-suite-local.sh # Full suite See scripts/local-test/README.md for details. Also updated .gitignore to: - Allow scripts/local-test/ to be tracked - Ignore generated artifacts (reports/*.xml, .claude/local/, .unity-mcp/) * Fix issue #525: Save dirty scenes for all test modes Move SaveDirtyScenesIfNeeded() call outside the PlayMode conditional so EditMode tests don't get blocked by Unity's "Save Scene" modal dialog. This prevents MCP from timing out when running EditMode tests with unsaved scene changes. * refactor: Consolidate editor state resources into single canonical implementation Merged EditorStateV2 into EditorState, making get_editor_state the canonical resource. Updated Unity C# to use EditorStateCache directly. Enhanced Python implementation with advice/staleness enrichment, external changes detection, and instance ID inference. Removed duplicate EditorStateV2 resource and legacy fallback mapping. * Validate editor state with Pydantic models in both C# and Python Added strongly-typed Pydantic models for EditorStateV2 schema in Python and corresponding C# classes with JsonProperty attributes. Updated C# to serialize using typed classes instead of anonymous objects. Python now validates the editor state payload before returning it, catching schema mismatches early. * Consolidate run_tests and run_tests_async into single async implementation Merged run_tests_async into run_tests, making async job-based execution the default behavior. Removed synchronous blocking test execution. Updated RunTests.cs to start test jobs immediately and return job_id for polling. Changed TestJobManager methods to internal visibility. Updated README to reflect single run_tests_async tool. Python implementation now uses async job pattern exclusively. * Validate test job responses with Pydantic models in Python * Change resources URI from unity:// to mcpforunity:// It should reduce conflicts with other Unity MCPs that users try, and to comply with Unity's requests regarding use of their company and product name * Update README with all tools + better listing for resources * Update other references to resources * Updated translated doc - unfortunately I cannot verify * Update the Chinese translation of the dev docks * Change menu item from Setup Window to Local Setup Window We now differentiate whether it's HTTP local or remote * Fix URIs for menu items and tests * Shouldn't have removed it * fix: add missing FAST_FAIL_TIMEOUT constant in PluginHub The FAST_FAIL_TIMEOUT class attribute was referenced on line 149 but never defined, causing AttributeError on every ping attempt. This error was silently caught by the broad 'except Exception' handler, causing all fast-fail commands (read_console, get_editor_state, ping) to fail after 6 seconds of retries with 'ping not answered' error. Added FAST_FAIL_TIMEOUT = 10 to define a 10-second timeout for fast-fail commands, matching the intent of the existing fast-fail infrastructure. * feat(ScriptableObject): enhance dry-run validation for AnimationCurve and Quaternion Dry-run validation now validates value formats, not just property existence: - AnimationCurve: Validates structure ({keys:[...]} or direct array), checks each keyframe is an object, validates numeric fields (time, value, inSlope, outSlope, inWeight, outWeight) and integer fields (weightedMode) - Quaternion: Validates array length (3 for Euler, 4 for raw) or object structure ({x,y,z,w} or {euler:[x,y,z]}), ensures all components are numeric Refactored shared validation helpers into appropriate locations: - ParamCoercion: IsNumericToken, ValidateNumericField, ValidateIntegerField - VectorParsing: ValidateAnimationCurveFormat, ValidateQuaternionFormat Added comprehensive XML documentation clarifying keyframe field defaults (all default to 0 except as noted). Added 5 new dry-run validation tests covering valid and invalid formats for both AnimationCurve and Quaternion properties. * test: fix integration tests after merge - test_refresh_unity_retry_recovery: Mock now handles both refresh_unity and get_editor_state commands (refresh_unity internally calls get_editor_state when wait_for_ready=True) - test_run_tests_async_forwards_params: Mock response now includes required 'mode' field for RunTestsStartResponse Pydantic validation - test_get_test_job_forwards_job_id: Updated to handle GetTestJobResponse as Pydantic model instead of dict (use model_dump() for assertions) * Update warning message to apply to all test modes Follow-up to PR #527: Since SaveDirtyScenesIfNeeded() now runs for all test modes, update the warning message to say 'tests' instead of 'PlayMode tests'. * feat(run_tests): add wait_timeout to get_test_job to avoid client loop detection When polling for test completion, MCP clients like Cursor can detect the repeated get_test_job calls as 'looping' and terminate the agent. Added wait_timeout parameter that makes the server wait internally for tests to complete (polling Unity every 2s) before returning. This dramatically reduces client-side tool calls from 10-20 down to 1-2, avoiding loop detection. Usage: get_test_job(job_id='xxx', wait_timeout=30) - Returns immediately if tests complete within timeout - Returns current status if timeout expires (client can call again) - Recommended: 30-60 seconds * fix: use Pydantic attribute access in test_run_tests_async for merge compatibility * revert: remove local test harness - will be submitted in separate PR * fix: stdio transport survives test runs without UI flicker Root cause: WriteToConfigTests.TearDown() was unconditionally deleting UseHttpTransport EditorPref even when tests were skipped on Windows (NUnit runs TearDown even after Assert.Ignore). Changes: - Fix WriteToConfigTests to save/restore prefs instead of deleting - Add centralized ShouldForceUvxRefresh() for local dev path detection - Clean stale Python build/ artifacts before client configuration - Improve reload handler flag management to prevent stuck Resuming state - Show Resuming status during stdio bridge restart - Initialize client config display on window open - Add DevModeForceServerRefresh to EditorPrefs window known types --------- Co-authored-by: Marcus Sanatan <msanatan@gmail.com> Co-authored-by: Scott Jennings <scott.jennings+CIGINT@cloudimperiumgames.com>
2026-01-08 11:33:22 +08:00
string devFlags = AssetPathUtility.ShouldForceUvxRefresh() ? "--no-cache --refresh " : string.Empty;
Payload-safe paging for hierarchy/components + safer asset search + docs (#490) * Fix test teardown to avoid dropping MCP bridge CodexConfigHelperTests was calling MCPServiceLocator.Reset() in TearDown, which disposes the active bridge/transport during MCP-driven test runs. Replace with restoring only the mutated service (IPlatformService). * Avoid leaking PlatformService in CodexConfigHelperTests Capture the original IPlatformService before this fixture runs and restore it in TearDown. This preserves the MCP connection safety fix (no MCPServiceLocator.Reset()) while avoiding global state leakage to subsequent tests. * Fix SO MCP tooling: validate folder roots, normalize paths, expand tests; remove vestigial SO tools * Remove UnityMCPTests stress artifacts and ignore Assets/Temp * Ignore UnityMCPTests Assets/Temp only * Clarify array_resize fallback logic comments * Refactor: simplify action set and reuse slash sanitization * Enhance: preserve GUID on overwrite & support Vector/Color types in ScriptableObject tools * Fix: ensure asset name matches filename to suppress Unity warnings * Fix: resolve Unity warnings by ensuring asset name match and removing redundant import * Refactor: Validate assetName, strict object parsing for vectors, remove broken SO logic from ManageAsset * Hardening: reject Windows drive paths; clarify supported asset types * Delete FixscriptableobjecPlan.md * Paginate get_hierarchy and get_components to prevent large payload crashes * dev: add uvx dev-mode refresh + safer HTTP stop; fix server typing eval * Payload-safe paging defaults + docs; harden asset search; stabilize Codex tests * chore: align uvx args + coercion helpers; tighten safety guidance * chore: minor cleanup + stabilize EditMode SO tests
2025-12-29 12:57:57 +08:00
return "# Register the MCP server with Claude Code:\n" +
$"claude mcp add --transport stdio UnityMCP -- \"{uvxPath}\" {devFlags}--from \"{packageSource}\" mcp-for-unity\n\n" +
"# Unregister the MCP server:\n" +
"claude mcp remove UnityMCP\n\n" +
"# List registered servers:\n" +
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
"claude mcp list";
}
public override IList<string> GetInstallationSteps() => new List<string>
{
"Ensure Claude CLI is installed",
"Use Register to add UnityMCP (or run claude mcp add UnityMCP)",
"Restart Claude Code"
};
fix: comprehensive performance optimizations, claude code config, and stability improvements (issue #577) (#595) * fix: reduce per-frame GC allocations causing editor hitches (issue #577) Eliminate memory allocations that occurred every frame, which triggered garbage collection spikes (~28ms) approximately every second. Changes: - EditorStateCache: Skip BuildSnapshot() entirely when state unchanged (check BEFORE building). Increased poll interval from 0.25s to 1.0s. Cache DeepClone() results to avoid allocations on GetSnapshot(). - TransportCommandDispatcher: Early exit before lock/list allocation when Pending.Count == 0, eliminating per-frame allocations when idle. - StdioBridgeHost: Same early exit pattern for commandQueue. - MCPForUnityEditorWindow: Throttle OnEditorUpdate to 2-second intervals instead of every frame, preventing expensive socket checks 60+/sec. Fixes GitHub issue #577: High performance impact even when MCP server is off * fix: prevent multiple domain reloads when calling refresh_unity (issue #577) Root Cause: - send_command() had a hardcoded retry loop (min 6 attempts) on connection errors - Each retry resent the refresh_unity command, causing Unity to reload 6 times - retry_on_reload=False only controlled reload-state retries, not connection retries The Fix: 1. Unity C# (MCPForUnity/Editor): - Added --reinstall flag to uvx commands in dev mode - Ensures local development changes are picked up by uvx/Claude Code - Applies to all client configurators (Claude Code, Codex, etc.) 2. Python Server (Server/src): - Added max_attempts parameter to send_command() - Pass max_attempts=0 when retry_on_reload=False - Fixed type handling in refresh_unity.py (handle MCPResponse objects) - Added timeout to connection error recovery conditions - Recovery logic now returns success instead of error to prevent client retries Changes: - MCPForUnity/Editor: Added --reinstall to dev mode uvx commands - Server/refresh_unity.py: Fixed type handling, improved error recovery - Server/unity_connection.py: Added max_attempts param, disable retries when retry_on_reload=False Result: refresh_unity with compile=request now triggers only 1 domain reload instead of 6 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: UI and server stability improvements Unity Editor (C#): - Fix "Resuming..." stuck state when manually clicking End Session Clear ResumeStdioAfterReload and ResumeHttpAfterReload flags in OnConnectionToggleClicked and EndOrphanedSessionAsync to prevent UI from getting stuck showing "Resuming..." with disabled button - Remove unsupported --reinstall flag from all uvx command builders uvx does not support --reinstall and shows warning when used Use --no-cache --refresh instead for dev mode cache busting Python Server: - Add "aborted" to connection error patterns in refresh_unity Handle WinError 10053 (connection aborted) gracefully during Unity domain reload, treating it as expected behavior - Add WindowsSafeRotatingFileHandler to suppress log rotation errors Windows file locking prevents log rotation when file is open by another process; catch PermissionError to avoid noisy stack traces - Fix packaging: add py-modules = ["main"] to pyproject.toml setuptools.packages.find only discovers packages (directories with __init__.py), must explicitly list standalone module files Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * docs: improve refresh_unity connection loss handling documentation Add detailed comments and logging to clarify why connection loss during compile is treated as success (expected domain reload behavior, not failure). This addresses PR feedback about potentially masking real connection errors. The logic is intentional and correct: - Connection loss only treated as success when compile='request' - Domain reload causing disconnect is expected Unity behavior - Subsequent wait_for_ready loop validates Unity becomes ready - Prevents multiple domain reload loops (issue #577) Added logging for observability: - Info log when expected disconnect detected - Warning log for non-recoverable errors Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: add missing logger import in refresh_unity Missing logger import causes NameError at runtime when connection loss handling paths are triggered (lines 82 and 91). Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and semantics Addresses four issues raised in code review: 1. EditorStateCache.GetSnapshot() - Remove shared cached clone - Revert to always returning fresh DeepClone() to prevent mutation bugs - Main GC optimization remains: state-change detection prevents unnecessary _cached rebuilds (the expensive operation) - Removed _cachedClone and _cachedCloneSequence fields 2. refresh_unity.py - Fix blocking reason terminology mismatch - Changed "asset_refresh" to "asset_import" to match activityPhase values from EditorStateCache.cs - Ensures asset import is correctly detected as blocking state 3. TransportCommandDispatcher - Fix unsynchronized Count access - Moved Pending.Count check inside PendingLock - Prevents data races and InvalidOperationException from concurrent dictionary access 4. StdioBridgeHost - Fix unsynchronized Count access - Moved commandQueue.Count check inside lockObj - Ensures all collection access is properly serialized All changes maintain the GC allocation optimizations while fixing thread safety violations and semantic contract changes. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: address code review feedback on thread safety and timeout handling - refresh_unity.py: Track readiness explicitly and return failure on timeout instead of silently returning success when wait loop exits without confirming ready_for_tools=true - McpClientConfiguratorBase.cs: Add thread safety guard for Configure() call in CheckStatusWithProjectDir(). Changed default attemptAutoRewrite to false and added runtime check to prevent calling Configure() from background threads. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
2026-01-21 10:11:25 +08:00
/// <summary>
/// Extracts the package source (--from argument value) from claude mcp get output.
/// The output format includes args like: --from "mcpforunityserver==9.0.1"
/// </summary>
private static string ExtractPackageSourceFromCliOutput(string cliOutput)
{
if (string.IsNullOrEmpty(cliOutput))
return null;
// Look for --from followed by the package source
// The CLI output may have it quoted or unquoted
int fromIndex = cliOutput.IndexOf("--from", StringComparison.OrdinalIgnoreCase);
if (fromIndex < 0)
return null;
// Move past "--from" and any whitespace
int startIndex = fromIndex + 6;
while (startIndex < cliOutput.Length && char.IsWhiteSpace(cliOutput[startIndex]))
startIndex++;
if (startIndex >= cliOutput.Length)
return null;
// Check if value is quoted
char quoteChar = cliOutput[startIndex];
if (quoteChar == '"' || quoteChar == '\'')
{
startIndex++;
int endIndex = cliOutput.IndexOf(quoteChar, startIndex);
if (endIndex > startIndex)
return cliOutput.Substring(startIndex, endIndex - startIndex);
}
else
{
// Unquoted - read until whitespace or end of line
int endIndex = startIndex;
while (endIndex < cliOutput.Length && !char.IsWhiteSpace(cliOutput[endIndex]))
endIndex++;
if (endIndex > startIndex)
return cliOutput.Substring(startIndex, endIndex - startIndex);
}
return null;
}
}
}