fix(core): Claude Opus & Sonnet 5.5 return 404/429 loop because upstream Antigravity uses 4.6 thinking models #35

Summary

In @cortexkit/opencode-antigravity-auth v3.0+, all requests to antigravity-claude-opus-5-5 and antigravity-claude-sonnet-5-5 fail or hang indefinitely in OpenCode. The plugin gets trapped in a 404 -> 429 fallback loop, penalizes every active account in the pool with rate-limit cooldowns, and eventually aborts with 0 tokens.

Meanwhile, official agy CLI works immediately and succeeds in ~6 seconds because the real upstream Google Antigravity backend hosts claude-opus-4-6-thinking and claude-sonnet-4-6, whereas 5-5 models do not exist on the server.

Environment

• OS: Linux (Ubuntu) / Cross-platform
• OpenCode: 1.19.0
• Node.js: v24.21.0
• Plugin version: @cortexkit/opencode-antigravity-auth 3.0.2 (@cortexkit/antigravity-auth-core 3.0.2)
• Accounts: 9 active, verified Google accounts with Google AI / Antigravity access

Root Cause Analysis

1. 5-5 Model IDs Return 404 on Upstream Backend

The plugin defines claude-opus-5-5-low/medium/high and vertexModelId: "claude-opus-5-5@default" in model-registry.ts.
When sent to Google Antigravity's primary endpoint (https://daily-cloudcode-pa.googleapis.com/v1internal:streamGenerateContent?alt=sse):

• claude-opus-5-5-low -> HTTP 404 (NOT_FOUND: "Requested entity was not found.")
• claude-opus-5-5@default -> HTTP 404 (NOT_FOUND)
• claude-sonnet-5-5-low -> HTTP 404 (NOT_FOUND)
• claude-opus-4-6-thinking -> HTTP 200 (SUCCESS)
• claude-sonnet-4-6 -> HTTP 200 (SUCCESS)

2. 404 Triggers Faulty Fallback to Prod (429 Loop)

In request-services.ts, ANTIGRAVITY_ENDPOINT_FALLBACKS retries on 404 to https://cloudcode-pa.googleapis.com (prod) with project: aicode-consumers:

timestamp=2026-10-13T13:44:44.550Z level=ERROR message="stream error" providerID=google modelID=antigravity-claude-sonnet-5-5 error.error="AI_RetryError: Failed after 3 attempts. Last error: Resource has been exhausted (e.g. check quota).
[Debug Info]
Requested Model: antigravity-claude-sonnet-5-5
Effective Model: claude-sonnet-5-5-low
Project: aicode-consumers
Endpoint: https://cloudcode-pa.googleapis.com/v1internal:streamGenerateContent?alt=sse
Status: 429
Request ID: N/A"

The rotation engine marks the account rate-limited for Claude, switches to the next account, hits 404->429 again, until all accounts in the pool are marked in rateLimitResetTimes and the request aborts.

3. isRetiredAgyModel Prevents Using the Actual Working Models

In model-registry.ts:

export function isRetiredAgyModel(model: string): boolean {
    const route = model.toLowerCase().replace(/^antigravity-/, '').replace(/^gemini-(?=claude-)/, '');
    return (/^claude-(opus|sonnet)-4-6(?:-thinking)?(?:-(?:minimal|low|medium|high))?$/.test(route) && getAgyNativeModel(route) === undefined);
}

If a user tries to configure antigravity-claude-opus-4-6-thinking, the plugin intercepts it with:
412 FAILED_PRECONDITION: This Antigravity model is no longer offered. Choose another model.

Control Test (Official agy CLI)

The official Google agy binary (agy models) lists:

• claude-opus-4-6-thinking
• claude-sonnet-4-6

Executing:

agy --model claude-opus-4-6-thinking -p "Say OK"
# Output: OK (completed in 6 seconds)

Inspection of /tmp/agy.log confirms agy hits https://daily-cloudcode-pa.googleapis.com/v1internal:streamGenerateContent?alt=sse with claude-opus-4-6-thinking and receives a 200 SSE stream with req_vrtx_... response IDs.

Steps to Reproduce in OpenCode

1. Install @cortexkit/opencode-antigravity-auth@latest (3.0.2).
2. Configure or run:

   opencode run -m google/antigravity-claude-opus-5-5 "Say OK"

3. OpenCode hangs for ~4 minutes, rotates through all accounts, logs 429 errors, and aborts with 0 tokens.

Proposed Fix

1. In packages/core/src/model-registry.ts:

• Remove or disable isRetiredAgyModel rejection:

  export function isRetiredAgyModel(model: string): boolean {
      return false;
  }

• Ensure claude-opus-4-6-thinking and claude-sonnet-4-6 are registered in AGY_NATIVE_MODEL_DEFINITIONS:

  'claude-opus-4-6-thinking': {
      displayName: 'Claude Opus 4.6 (Thinking)',
      maxTokens: 250000,
      maxOutputTokens: 64000,
      model: 'claude-opus-4-6-thinking',
      thinkingBudget: 1024,
      supportsThinking: true,
      supportsAdaptiveThinking: false,
      vertexModelId: 'claude-opus-4-6-thinking',
  },
  'claude-sonnet-4-6': {
      displayName: 'Claude Sonnet 4.6 (Thinking)',
      maxTokens: 250000,
      maxOutputTokens: 64000,
      model: 'claude-sonnet-4-6',
      thinkingBudget: 1024,
      supportsThinking: true,
      supportsAdaptiveThinking: false,
      vertexModelId: 'claude-sonnet-4-6',
  },

2. In packages/core/src/transform/model-resolver.ts:

Route claude-opus and claude-sonnet (including the 5-5 catalog IDs) to the real wire model names:

if (/^claude-opus/i.test(modelWithoutQuota)) {
    return {
        actualModel: 'claude-opus-4-6-thinking',
        thinkingBudget: 1024,
        isThinkingModel: true,
        tier: tier ?? 'high',
        quotaPreference,
        explicitQuota,
    };
}
if (/^claude-sonnet/i.test(modelWithoutQuota)) {
    return {
        actualModel: 'claude-sonnet-4-6',
        thinkingBudget: 1024,
        isThinkingModel: true,
        tier: tier ?? 'high',
        quotaPreference,
        explicitQuota,
    };
}

Verification of Fix

With this patch applied locally to the runtime:

$ opencode run -m google/antigravity-claude-opus-5-5 "Say OK"
> orchestrator · antigravity-claude-opus-5-5
OK  # Completed in 9s!

$ opencode run -m google/antigravity-claude-opus-5-5 "What is 2+2? Answer in one word."
> orchestrator · antigravity-claude-opus-5-5
Four.  # Completed in 9s!

$ opencode run -m google/antigravity-claude-sonnet-5-5 "Say OK"
> orchestrator · antigravity-claude-sonnet-5-5
OK  # Completed in 8s!