Plugin Architecture Overview: Understanding the Internal Design of OpenCode Antigravity Auth

What You'll Learn

This course will guide you through the architecture design and internal working principles of the Antigravity Auth plugin. After completing this course, you will be able to:

• Understand the plugin's module structure and the responsibilities of each module
• Master the complete processing flow of requests from OpenCode to Antigravity API
• Learn about the load balancing mechanism for multi-account management
• Understand the working principles of session recovery
• Master the technical details of request format transformation

Core Value of the Plugin

The core role of the Antigravity Auth plugin is to build a "translation bridge" between OpenCode and Antigravity API:

1. Protocol Transformation - Convert OpenCode request formats to formats required by Antigravity API
2. Authentication Management - Securely obtain and refresh access tokens through Google OAuth 2.0 PKCE
3. Load Balancing - Pool multiple accounts with automatic rotation to avoid rate limits
4. Error Recovery - Automatically detect and repair session interruption issues

Why Understand the Architecture?

Understanding the plugin architecture helps you:

• Quickly locate the module where a problem occurs
• Understand the scope of configuration options
• Optimize multi-account usage strategies
• Participate in plugin development and debugging

Module Structure Overview

The plugin adopts a layered architecture with clear responsibilities for each module:

Module Responsibilities

Module	File	Responsibilities
Main Entry	plugin.ts	Fetch interceptor, plugin initialization, event handling
Account Management	accounts.ts	Multi-account pool, load balancing strategy, rate limit tracking
Request Transformation	request.ts	Request format transformation, response streaming
Model Transformation	transform/	Claude/Gemini-specific transformation logic
OAuth Authentication	antigravity/oauth.ts	PKCE authentication flow, token exchange
Session Recovery	recovery/	Error detection, automatic repair of interrupted sessions
Configuration Management	config/	Configuration loading, Schema validation
Storage Layer	storage.ts	Account persistence (JSON file)
Cache Layer	cache.ts	Signature caching, reduce redundant computation
Debug Logging	debug.ts	Structured logging, troubleshooting

Request Processing Flow

When a model request is sent from OpenCode, the plugin goes through a complete processing chain:

Step Details

Step 1: Request Interception

• The plugin captures all requests to generativelanguage.googleapis.com through the fetch interceptor
• Uses isGenerativeLanguageRequest() to determine if it's an Antigravity-related request
• Non-target requests are passed through directly to avoid affecting other functionality

Step 2: Account Selection

• Selects an available account from the account pool (based on configured strategy: sticky/round-robin/hybrid)
• Checks if the token is expired, and automatically refreshes if so
• Applies PID offset (if enabled) to ensure even account distribution in multi-agent scenarios

Step 3: Request Transformation (request.ts)

• Model Detection: Identify whether it's a Claude or Gemini model
• Thinking Handling:
  • Claude: Strip all historical thinking blocks to avoid signature conflicts
  • Gemini: Add thinkingConfig configuration
• Tool Transformation: Convert OpenCode tools to functionDeclarations format
• Schema Cleaning: Remove unsupported JSON Schema fields (const, $ref, $defs, etc.)
• Request Wrapping: Encapsulate as { project, model, request: {...} } format

Step 4: Send Request

• Add Antigravity-specific headers (User-Agent, Client-Metadata)
• Use Bearer Token authentication
• Support endpoint fallback (daily → autopush → prod)

Step 5: Response Transformation

• SSE streaming reception, line-by-line transformation
• Convert thought: true to type: "reasoning"
• Cache thinking signature for subsequent requests
• Unwrap inner response object to maintain format consistency

Response Processing Flow

Response processing mainly involves streaming transformation and session recovery:

Key Processing Mechanisms

1. Thinking Block Transformation

The thinking block format returned by Antigravity API is:

{ "thought": true, "text": "thinking content" }

The plugin transforms it to OpenCode-compatible format:

{ "type": "reasoning", "reasoning": "thinking content" }

2. Signature Caching

Claude and Gemini 3 models require matching thinking block signatures in multi-turn conversations:

• Extract thoughtSignature from response
• Store to memory and disk using cacheSignature()
• Retrieve and inject via getCachedSignature() in next request

Why Need Signature Caching?

Claude and Gemini 3 APIs require thinking blocks before tool calls to have correct signatures. Caching signatures can:

• Avoid request failures caused by signature errors
• Reduce redundant generation of identical thinking content
• Improve continuity of multi-turn conversations

3. Session Recovery

When a session error is detected:

1. Capture error through session.error event
2. Call detectErrorType() to determine if recoverable
3. For tool_result_missing error:
   • Extract failed tool_use IDs
   • Inject synthetic tool_result messages
   • Automatically send "continue" (if enabled)
4. For thinking block order errors:
   • Analyze conversation state
   • Close damaged turn
   • Start new turn for model to regenerate

Core Mechanisms

OAuth Authentication Mechanism

The plugin uses OAuth 2.0 with PKCE (Proof Key for Code Exchange) flow:

PKCE Security:

• Generate random code_verifier and code_challenge
• Authorization URL contains code_challenge to prevent man-in-the-middle attacks
• Verify code_verifier during token exchange to ensure request comes from same client

Token Management:

• Access Token: Default 1-hour expiration, automatically refresh 30 minutes before expiration
• Refresh Token: Long-term valid, used to obtain new Access Token
• Persisted to ~/.config/opencode/antigravity-accounts.json

Multi-Account Management Mechanism

Multi-account management is one of the core features of the plugin, consisting of the following key components:

1. Account Pool Structure

interface AccountPool {
  version: 3;
  accounts: Account[];
  activeIndex: number;
  activeIndexByFamily: {
    claude: number;
    gemini: number;
  };
}

2. Account Selection Strategy

Strategy	Description	Use Case
sticky	Keep current account until rate limit	Single session use, preserve prompt cache
round-robin	Rotate to next account per request	Multi-session parallel, maximize throughput
hybrid	Health score + Token bucket + LRU combined decision	Default strategy, balance performance and reliability

3. Rate Limit Tracking

Each account tracks independently by model family:

• claude: Claude model quota
• gemini-antigravity: Gemini Antigravity quota
• gemini-cli: Gemini CLI quota

Rate limit status includes:

• rateLimitResetTimes: Reset times for each quota
• cooldownEndAt: Account cooldown end time
• consecutiveFailures: Number of consecutive failures

4. Dual Quota Pool (Gemini-Specific)

Gemini models support two independent quota pools:

• Antigravity Quota: Daily sandbox environment
• Gemini CLI Quota: Production environment

Through quota_fallback configuration:

• Prioritize primary quota (determined by model suffix)
• Try fallback quota when primary quota is exhausted
• Switch to next account when both pools are exhausted

5. PID Offset Optimization

When pid_offset_enabled is enabled:

• Agents from different processes (PIDs) start from different accounts
• Avoid all agents competing for the same account
• Suitable for multi-agent parallel scenarios

Request Transformation Mechanism

Request transformation is the most complex part of the plugin, handling protocol differences between different models:

Claude Model Transformation Key Points:

1. Thinking Block Handling

   // Strip all historical thinking blocks (avoid signature conflicts)
   const filteredContents = deepFilterThinkingBlocks(contents);
   
   // Build new thinkingConfig
   const thinkingConfig = {
     budgetTokens: variantThinkingConfig?.budget || DEFAULT_THINKING_BUDGET,
   };

2. Tool Schema Cleaning

   • Whitelist retain: type, properties, required, description, enum, items
   • Remove fields: const, $ref, $defs, default, examples, additionalProperties, $schema, title
   • Special transformation: const: "value" → enum: ["value"]

3. Tool Injection Parameter Signature

   injectParameterSignatures(tool, signature) {
     tool.description += `\n\nParameters: ${signature}`;
   }

Gemini Model Transformation Key Points:

1. Thinking Configuration

   // Gemini 3 Pro/Flash
   if (isGemini3Model(model)) {
     request.thinkingConfig = {
       thinkingLevel: "high" | "medium" | "low" | "minimal",
     };
   }
   
   // Gemini 2.5
   else if (isGemini2Model(model)) {
     request.thinkingConfig = {
       thinkingLevel: "high" | "medium" | "low",
     };
   }

2. Google Search Grounding

   if (webSearchConfig) {
     request.generationConfig = {
       ...request.generationConfig,
       dynamicRetrievalConfig: {
         mode: webSearchConfig.mode, // "auto" | "off"
         dynamicRetrievalConfig: {
           scoreThreshold: webSearchConfig.threshold, // 0.0 - 1.0
         },
       },
     };
   }

3. Image Generation Configuration

   if (isImageGenerationModel(model)) {
     request.generationConfig = {
       ...request.generationConfig,
       responseModalities: ["IMAGE", "TEXT"],
     };
   }

Session Recovery Mechanism

The session recovery mechanism ensures conversations can continue after unexpected interruptions:

1. Error Detection

function detectErrorType(error: unknown): RecoverableError | null {
  if (errorString.includes("tool_use ids were found without tool_result")) {
    return { type: "tool_result_missing" };
  }
  if (errorString.includes("Expected thinking but found text")) {
    return { type: "thinking_order_error" };
  }
  return null;
}

2. Turn Boundary Detection

// Turn boundary = first assistant message after user message
function analyzeConversationState(messages: Message[]): ConversationState {
  const lastUserMessage = findLastMessageByRole(messages, "user");
  const firstAssistantAfterUser = messages.find(m =>
    m.role === "assistant" && m.timestamp > lastUserMessage.timestamp
  );

  return {
    isTurnStart: true,
    turnAssistantMessage: firstAssistantAfterUser,
  };
}

3. Synthetic Message Injection

// Inject synthetic tool_result
function createSyntheticErrorResponse(toolUseIds: string[]): Message[] {
  return toolUseIds.map(id => ({
    type: "tool_result",
    tool_use_id: id,
    content: "Operation cancelled",
    isSynthetic: true,
  }));
}

// Close damaged turn
function closeToolLoopForThinking(sessionId: string): Message[] {
  return [
    {
      type: "text",
      text: "[Conversation turn closed due to error]",
      isSynthetic: true,
    },
  ];
}

4. Auto-Resume

When auto_resume is enabled, the recovery process automatically sends a continue instruction:

await client.session.prompt({
  path: { id: sessionID },
  body: { parts: [{ type: "text", text: config.resume_text }] },
  query: { directory },
});

Summary

This course has provided an in-depth analysis of the Antigravity Auth plugin's architecture design:

Core Architecture:

• Layered module design with clear responsibilities
• Plugin layer, authentication layer, transformation layer, and persistence layer each perform their duties
• Use plugin.ts as the unified entry point to coordinate all modules

Request Flow:

• Interception → Account selection → Token refresh → Request transformation → Send request → Response transformation
• Each step has detailed error handling and retry mechanisms

Key Mechanisms:

• OAuth 2.0 PKCE: Secure authentication flow
• Multi-Account Management: Load balancing, rate limit tracking, dual quota pools
• Request Transformation: Protocol compatibility, Schema cleaning, Thinking handling
• Session Recovery: Automatic error detection, synthetic message injection, Auto-Resume

Understanding these mechanisms will help you better use and optimize plugin configuration, as well as participate in plugin development and debugging.

Next Lesson

In the next lesson, we will learn Antigravity API Internal Specification.

You will learn:

• Request and response formats of Antigravity API
• Functions and parameters of each endpoint
• Error code and status code meanings
• Advanced features and experimental capabilities

---

Appendix: Source Code Reference

Click to expand source code locations

Last updated: 2026-01-23

Feature	File Path	Line Numbers
Plugin main entry, fetch interceptor	src/plugin.ts	654-1334
Account manager, load balancing	src/plugin/accounts.ts	1-715
Request transformation, response streaming	src/plugin/request.ts	1-1664
Claude model transformation	src/plugin/transform/claude.ts	Full file
Gemini model transformation	src/plugin/transform/gemini.ts	Full file
Session recovery	src/plugin/recovery/index.ts	Full file
Thinking block recovery	src/plugin/recovery/thinking-recovery.ts	Full file
OAuth authentication	src/antigravity/oauth.ts	1-271
Token management	src/plugin/token.ts	Full file
Configuration Schema	src/plugin/config/schema.ts	1-373
Account storage	src/plugin/storage.ts	Full file
Signature cache	src/plugin/cache.ts	Full file
Debug logging	src/plugin/debug.ts	Full file

Key Constants:

• MAX_OAUTH_ACCOUNTS = 10: Maximum number of accounts
• RATE_LIMIT_DEDUP_WINDOW_MS = 2000: Rate limit deduplication window (2 seconds)
• RATE_LIMIT_STATE_RESET_MS = 120000: Rate limit state reset time (2 minutes)
• FAILURE_COOLDOWN_MS = 30000: Account cooldown time (30 seconds)

Key Functions:

• createAntigravityPlugin(): Plugin main entry function
• prepareAntigravityRequest(): Request transformation main function
• transformAntigravityResponse(): Response streaming transformation function
• AccountManager.getCurrentOrNextForFamily(): Account selection function
• handleSessionRecovery(): Session recovery handling function
• refreshAccessToken(): Token refresh function