[BUG] Desktop 1.6259.1 adds model name validation that rejects non-Anthropic model names — breaks third-party gateways

[yang-12113]

Summary

Claude Desktop 1.6259.1 adds client-side model name validation in Gateway mode that rejects any model name not containing Anthropic-specific keywords (claude, sonnet, opus, haiku, anthropic). This breaks all third-party gateway users with custom model names. The validation did not exist in 1.5354.0.

Root Cause

In app.asar → .vite/build/index.js, the gateway validation function q5r calls Jte, which checks the model name against:

// Byte offset 6215292 in index.js
const eRt = /^(sonnet|opus|haiku)(-[\d.]+)?$/;

// Byte offset 6215331
const Y5r = ["claude", "sonnet", "opus", "haiku", "anthropic"];

// Byte offset 6216424 — gateway validator
function q5r(e) {
  return Jte(e)
    ? { ok: true }
    : {
        ok: false,
        reason:
          "expected a gateway model route referencing an Anthropic model " +
          "(e.g. claude-sonnet-4-5, anthropic/claude-*). " +
          "Name routes to match the underlying model.",
      };
}

A model name like mimo-v2.5-pro matches neither the regex nor any keyword, so it is rejected.

Steps to Reproduce

1. Install Claude Desktop (auto-updates to 1.6259.1).
2. Create a gateway config at %LOCALAPPDATA%\Claude-3p\configLibrary\<uuid>.json:

   {
     "inferenceProvider": "gateway",
     "inferenceGatewayBaseUrl": "https://your-gateway.example.com/anthropic",
     "inferenceGatewayApiKey": "sk-xxx",
     "inferenceModels": [{ "name": "custom-model-v1" }]
   }

3. Launch Claude Desktop, select custom-model-v1, send any message.

Expected: Message sent through gateway.
Actual: Error: expected a gateway model route referencing an Anthropic model...

What Should Happen

Gateway mode should pass through whatever model name the user configured. The gateway is responsible for routing — client-side validation of model names in gateway mode is incorrect behavior.

Proposed Fix

Remove or bypass the q5r validation for gateway connections:

 function q5r(e) {
-  return Jte(e)
-    ? { ok: true }
-    : {
-        ok: false,
-        reason:
-          "expected a gateway model route referencing an Anthropic model " +
-          "(e.g. claude-sonnet-4-5, anthropic/claude-*). " +
-          "Name routes to match the underlying model.",
-      };
+  return { ok: true };
 }

Or gate it behind an environment variable:

 function q5r(e) {
+  if (process.env.CLAUDE_CODE_SKIP_MODEL_VALIDATION) return { ok: true };
   return Jte(e) ? { ok: true } : { ok: false, reason: "..." };
 }

Additional Issue: Squirrel Launcher Crash

During investigation, we also found that the Squirrel launcher (%LOCALAPPDATA%\AnthropicClaude\claude.exe, 363KB, 32-bit PE) crashes with exit code -1073740791 (STATUS_INVALID_CRUNTIME_PARAMETER) when launched without arguments. This is independent of the asar issue — it crashes even with the original unmodified asar.

The app exe itself (app-1.6259.1\claude.exe, 64-bit PE) works fine when launched directly. The Squirrel launcher only works when called with explicit arguments: claude.exe --processStartAndWait claude.exe.

This means the Start Menu / desktop shortcuts (which point to the Squirrel launcher) don't work. Users must create a shortcut directly to app-1.6259.1\claude.exe.

Workaround (Verified Working)

Step 1: Replace app.asar (fixes model name validation)

Copy-Item "$env:LOCALAPPDATA\AnthropicClaude\app-1.5354.0\resources\app.asar" `
          "$env:LOCALAPPDATA\AnthropicClaude\app-1.6259.1\resources\app.asar" -Force

Step 2: Create direct shortcut (bypasses broken Squirrel launcher)

$shell = New-Object -ComObject WScript.Shell
$lnk = $shell.CreateShortcut("$env:USERPROFILE\Desktop\Claude.lnk")
$lnk.TargetPath = "$env:LOCALAPPDATA\AnthropicClaude\app-1.6259.1\claude.exe"
$lnk.WorkingDirectory = "$env:LOCALAPPDATA\AnthropicClaude\app-1.6259.1"
$lnk.IconLocation = "$env:LOCALAPPDATA\AnthropicClaude\app-1.6259.1\claude.exe,0"
$lnk.Save()

Restart Claude Desktop via the new shortcut. Both steps must be re-done after each update.

Why direct patching of app.asar doesn't work: The new app.asar has an Electron integrity check (asar_util.cc:143) that rejects modified archives. The expected hash is not stored in any discoverable file — it appears to be computed or embedded in a way that cannot be updated externally.

Impact

• Affected: All third-party gateway users with custom model names (e.g. Xiaomi MiMo, corporate API gateways, any non-Anthropic inference provider).
• Not affected: Users of Anthropic API, AWS Bedrock, Google Vertex AI, or gateways that happen to use model names containing Anthropic keywords.
• CLI not affected: claude --model <any-name> works fine — this validation only exists in the Desktop app.

Environment

• OS: Windows 11 Home China (10.0.26200)
• Claude Desktop: 1.6259.1 (regression from 1.5354.0)
• Gateway: Xiaomi MiMo API (token-plan-cn.xiaomimimo.com)
• Model: mimo-v2.5-pro

Related Issues

• [FEATURE] Support third-party inference providers in Claude Code (parity with Cowork 3P) #52572 — Support third-party inference providers in CLI (feature request)
• [BUG] Cowork 3P Gateway — Model picker ignores version and display name #52526 — 3P Gateway model selector issues
• [BUG] Claude Desktop 3P custom credential helper: ConfigHealth stuck in provider_error after cold-path helper auth succeeds #56758 — ConfigHealth stuck in provider_error

None report this specific model name validation regression.

[re-gion]

那我是不是还是用不了claude desktop/(ㄒoㄒ)/~~

[2ndbrainhub]

lỗi này thực sự khó chịu, mình không thể dùng model bên thứ 3 để chạy những thứ không cấp bách và không quan trọng. hãy xem xét sửa sớm nhất, đã thử nhiều cách để fix hoặc tạm bỏ qua lỗi nhưng không thực sự hữu dụng!

[UnrelatedWorks]

When using openRouter, only Ant model are found:

Manually setting does not work.

Proposed Fix:

A. Remove checking code.
B. Change error message: "We do not support 3rd party models".

[DhanshreeA]

This still persists. Any idea if third party inference through non Anthropic models will ever work on the Desktop app?