fix: strip cache_control fields from content blocks (#189)

Claude Code CLI sends cache_control on text, thinking, tool_use, and
tool_result blocks for prompt caching. Cloud Code API rejects these
with "Extra inputs are not permitted".

- Add cleanCacheControl() to proactively strip cache_control at pipeline entry
- Add sanitizeTextBlock() and sanitizeToolUseBlock() for defense-in-depth
- Update reorderAssistantContent() to use block sanitizers
- Add test-cache-control.cjs with multi-model test coverage
- Update frontend dashboard tests to match current UI design
- Update strategy tests to match v2.4.0 fallback behavior
- Update CLAUDE.md and README.md with recent features

Inspired by Antigravity-Manager's clean_cache_control_from_messages() pattern.

Co-Authored-By: Claude <noreply@anthropic.com>

CLAUDE.md

@@ -55,6 +55,7 @@ npm run test:images        # Image processing
 npm run test:caching       # Prompt caching
 npm run test:crossmodel    # Cross-model thinking signatures
 npm run test:oauth         # OAuth no-browser mode
+npm run test:cache-control # Cache control field stripping
 
 # Run strategy unit tests (no server required)
 node tests/test-strategies.cjs
@@ -102,7 +103,8 @@ src/
 │       └── trackers/           # State trackers for hybrid strategy
 │           ├── index.js        # Re-exports trackers
 │           ├── health-tracker.js    # Account health scores
-│           └── token-bucket-tracker.js  # Client-side rate limiting
+│           ├── token-bucket-tracker.js  # Client-side rate limiting
+│           └── quota-tracker.js     # Quota-aware account selection
 │
 ├── auth/                       # Authentication
 │   ├── oauth.js                # Google OAuth with PKCE
@@ -211,11 +213,15 @@ public/
    - Maximizes concurrent request distribution
 
 3. **Hybrid Strategy** (default, smart distribution):
-   - Uses health scores, token buckets, and LRU for selection
-   - Scoring formula: `score = (Health × 2) + ((Tokens / MaxTokens × 100) × 5) + (LRU × 0.1)`
+   - Uses health scores, token buckets, quota awareness, and LRU for selection
+   - Scoring formula: `score = (Health × 2) + ((Tokens / MaxTokens × 100) × 5) + (Quota × 1) + (LRU × 0.1)`
    - Health scores: Track success/failure patterns with passive recovery
    - Token buckets: Client-side rate limiting (50 tokens, 6 per minute regeneration)
+   - Quota awareness: Accounts with critical quota (<5%) are deprioritized
    - LRU freshness: Prefer accounts that have rested longer
+   - **Emergency/Last Resort Fallback**: When all accounts are exhausted:
+     - Emergency fallback: Bypasses health check, adds 250ms throttle delay
+     - Last resort fallback: Bypasses both health and token checks, adds 500ms throttle delay
    - Configuration in `src/config.js` under `accountSelection`
 
 **Account Data Model:**
@@ -251,6 +257,14 @@ Each account object in `accounts.json` contains:
 - For Gemini targets: strict validation - drops unknown or mismatched signatures
 - For Claude targets: lenient - lets Claude validate its own signatures
 
+**Cache Control Handling (Issue #189):**
+- Claude Code CLI sends `cache_control` fields on content blocks for prompt caching
+- Cloud Code API rejects these with "Extra inputs are not permitted"
+- `cleanCacheControl(messages)` strips cache_control from ALL block types at pipeline entry
+- Called at the START of `convertAnthropicToGoogle()` before any other processing
+- Additional sanitizers (`sanitizeTextBlock`, `sanitizeToolUseBlock`) provide defense-in-depth
+- Pattern inspired by Antigravity-Manager's `clean_cache_control_from_messages()`
+
 **Native Module Auto-Rebuild:**
 - When Node.js is updated, native modules like `better-sqlite3` may become incompatible
 - The proxy automatically detects `NODE_MODULE_VERSION` mismatch errors
@@ -284,7 +298,9 @@ Each account object in `accounts.json` contains:
   - ARIA labels on search inputs and icon buttons
   - Keyboard navigation support (Escape to clear search)
 - **Security**: Optional password protection via `WEBUI_PASSWORD` env var
+- **Config Redaction**: Sensitive values (passwords, tokens) are redacted in API responses
 - **Smart Refresh**: Client-side polling with ±20% jitter and tab visibility detection (3x slower when hidden)
+- **i18n Support**: English, Chinese (中文), Indonesian (Bahasa), Portuguese (PT-BR)
 
 ## Testing Notes
 

README.md

@@ -84,6 +84,8 @@ Choose one of the following methods to authorize the proxy:
 2. Navigate to the **Accounts** tab and click **Add Account**.
 3. Complete the Google OAuth authorization in the popup window.
 
+> **Headless/Remote Servers**: If running on a server without a browser, the WebUI supports a "Manual Authorization" mode. After clicking "Add Account", you can copy the OAuth URL, complete authorization on your local machine, and paste the authorization code back.
+
 #### **Method B: CLI (Desktop or Headless)**
 
 If you prefer the terminal or are on a remote server:
@@ -280,7 +282,7 @@ Choose a strategy based on your needs:
 
 | Strategy | Best For | Description |
 | --- | --- | --- |
-| **Hybrid** (Default) | Most users | Smart selection combining health score, token bucket rate limiting, and LRU freshness |
+| **Hybrid** (Default) | Most users | Smart selection combining health score, token bucket rate limiting, quota awareness, and LRU freshness |
 | **Sticky** | Prompt caching | Stays on the same account to maximize cache hits, switches only when rate-limited |
 | **Round-Robin** | Even distribution | Cycles through accounts sequentially for balanced load |
 
@@ -298,6 +300,8 @@ antigravity-claude-proxy start --strategy=round-robin  # Load-balanced
 
 - **Health Score Tracking**: Accounts earn points for successful requests and lose points for failures/rate-limits
 - **Token Bucket Rate Limiting**: Client-side throttling with regenerating tokens (50 max, 6/minute)
+- **Quota Awareness**: Accounts with critical quota (<5%) are deprioritized; exhausted accounts trigger emergency fallback
+- **Emergency Fallback**: When all accounts appear exhausted, bypasses checks with throttle delays (250-500ms)
 - **Automatic Cooldown**: Rate-limited accounts recover automatically after reset time expires
 - **Invalid Account Detection**: Accounts needing re-authentication are marked and skipped
 - **Prompt Caching Support**: Session IDs derived from conversation enable cache hits across turns
@@ -340,13 +344,14 @@ The proxy includes a built-in, modern web interface for real-time monitoring and
 - **Real-time Dashboard**: Monitor request volume, active accounts, model health, and subscription tier distribution.
 - **Visual Model Quota**: Track per-model usage and next reset times with color-coded progress indicators.
 - **Account Management**: Add/remove Google accounts via OAuth, view subscription tiers (Free/Pro/Ultra) and quota status at a glance.
+- **Manual OAuth Mode**: Add accounts on headless servers by copying the OAuth URL and pasting the authorization code.
 - **Claude CLI Configuration**: Edit your `~/.claude/settings.json` directly from the browser.
 - **Persistent History**: Tracks request volume by model family for 30 days, persisting across server restarts.
 - **Time Range Filtering**: Analyze usage trends over 1H, 6H, 24H, 7D, or All Time periods.
 - **Smart Analysis**: Auto-select top 5 most used models or toggle between Family/Model views.
 - **Live Logs**: Stream server logs with level-based filtering and search.
 - **Advanced Tuning**: Configure retries, timeouts, and debug mode on the fly.
-- **Bilingual Interface**: Full support for English and Chinese (switch via Settings).
+- **Multi-language Interface**: Full support for English, Chinese (中文), Indonesian (Bahasa), and Portuguese (PT-BR).
 
 ---
 
@@ -360,9 +365,11 @@ While most users can use the default settings, you can tune the proxy behavior v
 - **WebUI Password**: Secure your dashboard with `WEBUI_PASSWORD` env var or in config.
 - **Custom Port**: Change the default `8080` port.
 - **Retry Logic**: Configure `maxRetries`, `retryBaseMs`, and `retryMaxMs`.
+- **Rate Limit Handling**: Comprehensive rate limit detection from headers and error messages with intelligent retry-after parsing.
 - **Load Balancing**: Adjust `defaultCooldownMs` and `maxWaitBeforeErrorMs`.
 - **Persistence**: Enable `persistTokenCache` to save OAuth sessions across restarts.
 - **Max Accounts**: Set `maxAccounts` (1-100) to limit the number of Google accounts. Default: 10.
+- **Endpoint Fallback**: Automatic 403/404 endpoint fallback for API compatibility.
 
 Refer to `config.example.json` for a complete list of fields and documentation.
 
@@ -421,6 +428,7 @@ npm run test:interleaved   # Interleaved thinking
 npm run test:images        # Image processing
 npm run test:caching       # Prompt caching
 npm run test:strategies    # Account selection strategies
+npm run test:cache-control # Cache control field stripping
 ```
 
 ---

package.json

@@ -35,7 +35,8 @@
     "test:oauth": "node tests/test-oauth-no-browser.cjs",
     "test:emptyretry": "node tests/test-empty-response-retry.cjs",
     "test:sanitizer": "node tests/test-schema-sanitizer.cjs",
-    "test:strategies": "node tests/test-strategies.cjs"
+    "test:strategies": "node tests/test-strategies.cjs",
+    "test:cache-control": "node tests/test-cache-control.cjs"
   },
   "keywords": [
     "claude",

src/format/request-converter.js

@@ -18,7 +18,8 @@ import {
     hasGeminiHistory,
     hasUnsignedThinkingBlocks,
     needsThinkingRecovery,
-    closeToolLoopForThinking
+    closeToolLoopForThinking,
+    cleanCacheControl
 } from './thinking-utils.js';
 import { logger } from '../utils/logger.js';
 
@@ -32,7 +33,13 @@ import { logger } from '../utils/logger.js';
  * @returns {Object} Request body for Cloud Code API
  */
 export function convertAnthropicToGoogle(anthropicRequest) {
-    const { messages, system, max_tokens, temperature, top_p, top_k, stop_sequences, tools, tool_choice, thinking } = anthropicRequest;
+    // [CRITICAL FIX] Pre-clean all cache_control fields from messages (Issue #189)
+    // Claude Code CLI sends cache_control on various content blocks, but Cloud Code API
+    // rejects them with "Extra inputs are not permitted". Clean them proactively here
+    // before any other processing, following the pattern from Antigravity-Manager.
+    const messages = cleanCacheControl(anthropicRequest.messages || []);
+
+    const { system, max_tokens, temperature, top_p, top_k, stop_sequences, tools, tool_choice, thinking } = anthropicRequest;
     const modelName = anthropicRequest.model || '';
     const modelFamily = getModelFamily(modelName);
     const isClaudeModel = modelFamily === 'claude';

src/format/thinking-utils.js

@@ -7,6 +7,62 @@ import { MIN_SIGNATURE_LENGTH } from '../constants.js';
 import { getCachedSignatureFamily } from './signature-cache.js';
 import { logger } from '../utils/logger.js';
 
+// ============================================================================
+// Cache Control Cleaning (Issue #189)
+// ============================================================================
+
+/**
+ * Remove cache_control fields from all content blocks in messages.
+ * This is a critical fix for Issue #189 where Claude Code CLI sends cache_control
+ * fields that the Cloud Code API rejects with "Extra inputs are not permitted".
+ *
+ * Inspired by Antigravity-Manager's clean_cache_control_from_messages() approach,
+ * this function proactively strips cache_control from ALL block types at the
+ * entry point of the conversion pipeline.
+ *
+ * @param {Array<Object>} messages - Array of messages in Anthropic format
+ * @returns {Array<Object>} Messages with cache_control fields removed
+ */
+export function cleanCacheControl(messages) {
+    if (!Array.isArray(messages)) return messages;
+
+    let removedCount = 0;
+
+    const cleaned = messages.map(message => {
+        if (!message || typeof message !== 'object') return message;
+
+        // Handle string content (no cache_control possible)
+        if (typeof message.content === 'string') return message;
+
+        // Handle array content
+        if (!Array.isArray(message.content)) return message;
+
+        const cleanedContent = message.content.map(block => {
+            if (!block || typeof block !== 'object') return block;
+
+            // Check if cache_control exists before destructuring
+            if (block.cache_control === undefined) return block;
+
+            // Create a shallow copy without cache_control
+            const { cache_control, ...cleanBlock } = block;
+            removedCount++;
+
+            return cleanBlock;
+        });
+
+        return {
+            ...message,
+            content: cleanedContent
+        };
+    });
+
+    if (removedCount > 0) {
+        logger.debug(`[ThinkingUtils] Removed cache_control from ${removedCount} block(s)`);
+    }
+
+    return cleaned;
+}
+
 /**
  * Check if a part is a thinking block
  * @param {Object} part - Content part to check
@@ -104,6 +160,38 @@ function sanitizeAnthropicThinkingBlock(block) {
     return block;
 }
 
+/**
+ * Sanitize a text block by removing extra fields like cache_control.
+ * Only keeps: type, text
+ * @param {Object} block - Text block to sanitize
+ * @returns {Object} Sanitized text block
+ */
+function sanitizeTextBlock(block) {
+    if (!block || block.type !== 'text') return block;
+
+    const sanitized = { type: 'text' };
+    if (block.text !== undefined) sanitized.text = block.text;
+    return sanitized;
+}
+
+/**
+ * Sanitize a tool_use block by removing extra fields like cache_control.
+ * Only keeps: type, id, name, input, thoughtSignature (for Gemini)
+ * @param {Object} block - Tool_use block to sanitize
+ * @returns {Object} Sanitized tool_use block
+ */
+function sanitizeToolUseBlock(block) {
+    if (!block || block.type !== 'tool_use') return block;
+
+    const sanitized = { type: 'tool_use' };
+    if (block.id !== undefined) sanitized.id = block.id;
+    if (block.name !== undefined) sanitized.name = block.name;
+    if (block.input !== undefined) sanitized.input = block.input;
+    // Preserve thoughtSignature for Gemini models
+    if (block.thoughtSignature !== undefined) sanitized.thoughtSignature = block.thoughtSignature;
+    return sanitized;
+}
+
 /**
  * Filter content array, keeping only thinking blocks with valid signatures.
  */
@@ -259,11 +347,13 @@ export function reorderAssistantContent(content) {
             // Sanitize thinking blocks to remove cache_control and other extra fields
             thinkingBlocks.push(sanitizeAnthropicThinkingBlock(block));
         } else if (block.type === 'tool_use') {
-            toolUseBlocks.push(block);
+            // Sanitize tool_use blocks to remove cache_control and other extra fields
+            toolUseBlocks.push(sanitizeToolUseBlock(block));
         } else if (block.type === 'text') {
             // Only keep text blocks with meaningful content
             if (block.text && block.text.trim().length > 0) {
-                textBlocks.push(block);
+                // Sanitize text blocks to remove cache_control and other extra fields
+                textBlocks.push(sanitizeTextBlock(block));
             } else {
                 droppedEmptyBlocks++;
             }

tests/frontend/test-frontend-dashboard.cjs

@@ -91,11 +91,12 @@ const tests = [
             const res = await request('/views/dashboard.html');
             const html = res.data;
 
+            // Dashboard uses dropdown-based filters for time range, display mode, and model selection
             const filterElements = [
-                'filters.account',    // Account filter
-                'filters.family',     // Model family filter
-                'filters.search',     // Search input
-                'computeQuotaRows'    // Filter action
+                'showTimeRangeDropdown',  // Time range dropdown toggle
+                'showDisplayModeDropdown', // Display mode dropdown toggle
+                'showModelFilter',         // Model/family filter dropdown toggle
+                'setTimeRange'             // Time range action
             ];
 
             const missing = filterElements.filter(el => !html.includes(el));
@@ -106,23 +107,24 @@ const tests = [
         }
     },
     {
-        name: 'Dashboard table has required columns',
+        name: 'Dashboard has chart and visualization elements',
         async run() {
             const res = await request('/views/dashboard.html');
             const html = res.data;
 
-            const columns = [
-                'modelIdentity',      // Model name column
-                'globalQuota',        // Quota column
-                'nextReset',          // Reset time column
-                'distribution'        // Account distribution column
+            // Dashboard now uses charts instead of tables
+            const visualElements = [
+                'quotaChart',           // Quota distribution pie chart
+                'usageTrendChart',      // Usage trend line chart
+                'usageStats.total',     // Total usage stat
+                'selectedFamilies'      // Family selection for chart
             ];
 
-            const missing = columns.filter(col => !html.includes(col));
+            const missing = visualElements.filter(col => !html.includes(col));
             if (missing.length > 0) {
-                throw new Error(`Missing table columns: ${missing.join(', ')}`);
+                throw new Error(`Missing visualization elements: ${missing.join(', ')}`);
             }
-            return 'All table columns present';
+            return 'All chart and visualization elements present';
         }
     }
 ];

tests/run-all.cjs

@@ -10,6 +10,7 @@ const path = require('path');
 
 const tests = [
     { name: 'Account Selection Strategies', file: 'test-strategies.cjs' },
+    { name: 'Cache Control Stripping', file: 'test-cache-control.cjs' },
     { name: 'Thinking Signatures', file: 'test-thinking-signatures.cjs' },
     { name: 'Multi-turn Tools (Non-Streaming)', file: 'test-multiturn-thinking-tools.cjs' },
     { name: 'Multi-turn Tools (Streaming)', file: 'test-multiturn-thinking-tools-streaming.cjs' },

tests/test-cache-control.cjs

@@ -0,0 +1,297 @@
+/**
+ * Cache Control Field Test (Issue #189)
+ *
+ * Tests that cache_control fields on content blocks are properly stripped
+ * before being sent to the Cloud Code API.
+ *
+ * Claude Code CLI sends cache_control on text, thinking, tool_use, tool_result,
+ * image, and document blocks for prompt caching optimization. The Cloud Code API
+ * rejects these with "Extra inputs are not permitted".
+ *
+ * This test verifies that:
+ * 1. Text blocks with cache_control work correctly
+ * 2. Multi-turn conversations with cache_control on assistant content work
+ * 3. Tool_result blocks with cache_control work correctly
+ *
+ * Runs for both Claude and Gemini model families.
+ */
+const { streamRequest, analyzeContent, commonTools } = require('./helpers/http-client.cjs');
+const { getTestModels, getModelConfig } = require('./helpers/test-models.cjs');
+
+const tools = [commonTools.getWeather];
+
+async function runTestsForModel(family, model) {
+    console.log('='.repeat(60));
+    console.log(`CACHE CONTROL TEST [${family.toUpperCase()}]`);
+    console.log(`Model: ${model}`);
+    console.log('Tests that cache_control fields are stripped from all block types');
+    console.log('='.repeat(60));
+    console.log('');
+
+    let allPassed = true;
+    const results = [];
+    const modelConfig = getModelConfig(family);
+
+    // ===== TEST 1: User text block with cache_control =====
+    console.log('TEST 1: User text block with cache_control');
+    console.log('-'.repeat(40));
+
+    try {
+        const test1Result = await streamRequest({
+            model,
+            max_tokens: modelConfig.max_tokens,
+            stream: true,
+            thinking: modelConfig.thinking,
+            messages: [
+                {
+                    role: 'user',
+                    content: [
+                        {
+                            type: 'text',
+                            text: 'What is the capital of France? Reply in one word.',
+                            cache_control: { type: 'ephemeral' }
+                        }
+                    ]
+                }
+            ]
+        });
+
+        const hasError1 = test1Result.events.some(e => e.type === 'error');
+        const errorMsg1 = hasError1
+            ? test1Result.events.find(e => e.type === 'error')?.data?.error?.message
+            : null;
+
+        console.log(`  Response received: ${test1Result.content.length > 0 ? 'YES' : 'NO'}`);
+        console.log(`  Has error: ${hasError1 ? 'YES' : 'NO'}`);
+        if (hasError1) {
+            console.log(`  Error message: ${errorMsg1}`);
+        }
+
+        const content1 = analyzeContent(test1Result.content);
+        if (content1.hasText) {
+            console.log(`  Response preview: "${content1.text[0].text.substring(0, 50)}..."`);
+        }
+
+        const test1Pass = !hasError1 && test1Result.content.length > 0;
+        results.push({ name: 'User text block with cache_control', passed: test1Pass });
+        console.log(`  Result: ${test1Pass ? 'PASS' : 'FAIL'}`);
+        if (!test1Pass) allPassed = false;
+    } catch (err) {
+        console.log(`  ERROR: ${err.message}`);
+        results.push({ name: 'User text block with cache_control', passed: false });
+        allPassed = false;
+    }
+
+    // ===== TEST 2: Multi-turn with cache_control on assistant content =====
+    console.log('\nTEST 2: Multi-turn with cache_control on assistant content');
+    console.log('-'.repeat(40));
+
+    try {
+        // First turn - get a response
+        const turn1 = await streamRequest({
+            model,
+            max_tokens: modelConfig.max_tokens,
+            stream: true,
+            thinking: modelConfig.thinking,
+            messages: [
+                { role: 'user', content: 'Say hello.' }
+            ]
+        });
+
+        if (turn1.content.length === 0) {
+            console.log('  SKIPPED - Turn 1 returned empty response');
+            results.push({ name: 'Multi-turn with cache_control', passed: false, skipped: true });
+        } else {
+            // Add cache_control to ALL blocks in assistant response (simulating Claude Code)
+            const modifiedContent = turn1.content.map(block => ({
+                ...block,
+                cache_control: { type: 'ephemeral' }
+            }));
+
+            // Second turn - use modified content with cache_control
+            const turn2 = await streamRequest({
+                model,
+                max_tokens: modelConfig.max_tokens,
+                stream: true,
+                thinking: modelConfig.thinking,
+                messages: [
+                    { role: 'user', content: 'Say hello.' },
+                    { role: 'assistant', content: modifiedContent },
+                    {
+                        role: 'user',
+                        content: [
+                            {
+                                type: 'text',
+                                text: 'Now say goodbye.',
+                                cache_control: { type: 'ephemeral' }
+                            }
+                        ]
+                    }
+                ]
+            });
+
+            const hasError2 = turn2.events.some(e => e.type === 'error');
+            const errorMsg2 = hasError2
+                ? turn2.events.find(e => e.type === 'error')?.data?.error?.message
+                : null;
+
+            console.log(`  Turn 1 blocks: ${turn1.content.length}`);
+            console.log(`  Turn 2 response received: ${turn2.content.length > 0 ? 'YES' : 'NO'}`);
+            console.log(`  Has error: ${hasError2 ? 'YES' : 'NO'}`);
+            if (hasError2) {
+                console.log(`  Error message: ${errorMsg2}`);
+                // Check specifically for cache_control error
+                if (errorMsg2 && errorMsg2.includes('cache_control')) {
+                    console.log('  >>> cache_control field NOT stripped properly! <<<');
+                }
+            }
+
+            const content2 = analyzeContent(turn2.content);
+            if (content2.hasText) {
+                console.log(`  Response preview: "${content2.text[0].text.substring(0, 50)}..."`);
+            }
+
+            const test2Pass = !hasError2 && turn2.content.length > 0;
+            results.push({ name: 'Multi-turn with cache_control', passed: test2Pass });
+            console.log(`  Result: ${test2Pass ? 'PASS' : 'FAIL'}`);
+            if (!test2Pass) allPassed = false;
+        }
+    } catch (err) {
+        console.log(`  ERROR: ${err.message}`);
+        results.push({ name: 'Multi-turn with cache_control', passed: false });
+        allPassed = false;
+    }
+
+    // ===== TEST 3: Tool loop with cache_control on tool_result =====
+    console.log('\nTEST 3: Tool loop with cache_control on tool_result');
+    console.log('-'.repeat(40));
+
+    try {
+        // First turn - request tool use
+        const toolTurn1 = await streamRequest({
+            model,
+            max_tokens: modelConfig.max_tokens,
+            stream: true,
+            tools,
+            thinking: modelConfig.thinking,
+            messages: [
+                { role: 'user', content: 'What is the weather in Tokyo? Use the get_weather tool.' }
+            ]
+        });
+
+        const content3a = analyzeContent(toolTurn1.content);
+
+        if (!content3a.hasToolUse) {
+            console.log('  SKIPPED - Model did not use tool in turn 1');
+            results.push({ name: 'Tool_result with cache_control', passed: true, skipped: true });
+        } else {
+            const toolUseId = content3a.toolUse[0].id;
+            console.log(`  Tool use ID: ${toolUseId}`);
+
+            // Second turn - provide tool result with cache_control
+            const toolTurn2 = await streamRequest({
+                model,
+                max_tokens: modelConfig.max_tokens,
+                stream: true,
+                tools,
+                thinking: modelConfig.thinking,
+                messages: [
+                    { role: 'user', content: 'What is the weather in Tokyo? Use the get_weather tool.' },
+                    { role: 'assistant', content: toolTurn1.content },
+                    {
+                        role: 'user',
+                        content: [
+                            {
+                                type: 'tool_result',
+                                tool_use_id: toolUseId,
+                                content: 'The weather in Tokyo is 22°C and partly cloudy.',
+                                cache_control: { type: 'ephemeral' }
+                            }
+                        ]
+                    }
+                ]
+            });
+
+            const hasError3 = toolTurn2.events.some(e => e.type === 'error');
+            const errorMsg3 = hasError3
+                ? toolTurn2.events.find(e => e.type === 'error')?.data?.error?.message
+                : null;
+
+            console.log(`  Turn 2 response received: ${toolTurn2.content.length > 0 ? 'YES' : 'NO'}`);
+            console.log(`  Has error: ${hasError3 ? 'YES' : 'NO'}`);
+            if (hasError3) {
+                console.log(`  Error message: ${errorMsg3}`);
+                if (errorMsg3 && errorMsg3.includes('cache_control')) {
+                    console.log('  >>> cache_control field NOT stripped properly! <<<');
+                }
+            }
+
+            const content3b = analyzeContent(toolTurn2.content);
+            if (content3b.hasText) {
+                console.log(`  Response preview: "${content3b.text[0].text.substring(0, 50)}..."`);
+            }
+
+            const test3Pass = !hasError3 && toolTurn2.content.length > 0;
+            results.push({ name: 'Tool_result with cache_control', passed: test3Pass });
+            console.log(`  Result: ${test3Pass ? 'PASS' : 'FAIL'}`);
+            if (!test3Pass) allPassed = false;
+        }
+    } catch (err) {
+        console.log(`  ERROR: ${err.message}`);
+        results.push({ name: 'Tool_result with cache_control', passed: false });
+        allPassed = false;
+    }
+
+    // ===== Summary =====
+    console.log('\n' + '='.repeat(60));
+    console.log(`SUMMARY [${family.toUpperCase()}]`);
+    console.log('='.repeat(60));
+
+    for (const result of results) {
+        const status = result.skipped ? 'SKIP' : (result.passed ? 'PASS' : 'FAIL');
+        console.log(`  [${status}] ${result.name}`);
+    }
+
+    const passedCount = results.filter(r => r.passed && !r.skipped).length;
+    const skippedCount = results.filter(r => r.skipped).length;
+    const totalTests = results.length - skippedCount;
+
+    console.log('\n' + '='.repeat(60));
+    console.log(`[${family.toUpperCase()}] ${allPassed ? 'ALL TESTS PASSED' : 'SOME TESTS FAILED'} (${passedCount}/${totalTests})`);
+    console.log('='.repeat(60));
+
+    return allPassed;
+}
+
+async function runTests() {
+    console.log('');
+    console.log('='.repeat(60));
+    console.log('CACHE CONTROL FIELD STRIPPING TEST (Issue #189)');
+    console.log('='.repeat(60));
+    console.log('');
+    console.log('This test verifies that cache_control fields are properly');
+    console.log('stripped from all content blocks before sending to Cloud Code API.');
+    console.log('');
+
+    const models = await getTestModels();
+    let allPassed = true;
+
+    for (const { family, model } of models) {
+        console.log('\n');
+        const passed = await runTestsForModel(family, model);
+        if (!passed) allPassed = false;
+    }
+
+    console.log('\n' + '='.repeat(60));
+    console.log('FINAL RESULT');
+    console.log('='.repeat(60));
+    console.log(`Overall: ${allPassed ? 'ALL MODEL FAMILIES PASSED' : 'SOME MODEL FAMILIES FAILED'}`);
+    console.log('='.repeat(60));
+
+    process.exit(allPassed ? 0 : 1);
+}
+
+runTests().catch(err => {
+    console.error('Test failed with error:', err);
+    process.exit(1);
+});

tests/test-strategies.cjs

@@ -77,6 +77,12 @@ async function runTests() {
         }
     }
 
+    function assertNotNull(value, message = '') {
+        if (value === null || value === undefined) {
+            throw new Error(`${message}\nExpected non-null value but got: ${value}`);
+        }
+    }
+
     function assertWithin(actual, min, max, message = '') {
         if (actual < min || actual > max) {
             throw new Error(`${message}\nExpected value between ${min} and ${max}, got: ${actual}`);
@@ -691,7 +697,7 @@ async function runTests() {
         assertEqual(result.account.email, 'account3@example.com', 'Oldest account should be selected');
     });
 
-    test('HybridStrategy: filters out unhealthy accounts', () => {
+    test('HybridStrategy: uses emergency fallback for unhealthy accounts', () => {
         const strategy = new HybridStrategy({
             healthScore: { initial: 40, minUsable: 50 },
             tokenBucket: { initialTokens: 50, maxTokens: 50 }
@@ -699,19 +705,25 @@ async function runTests() {
         const accounts = createMockAccounts(3);
 
         // All accounts start with health 40, which is below minUsable 50
+        // But emergency fallback should still return an account
         const result = strategy.selectAccount(accounts, 'model');
-        assertNull(result.account, 'Should filter all accounts with low health');
+        assertNotNull(result.account, 'Emergency fallback should return an account');
+        // waitMs indicates fallback was used (250ms for emergency)
+        assertTrue(result.waitMs >= 250, 'Emergency fallback should add throttle delay');
     });
 
-    test('HybridStrategy: filters out accounts without tokens', () => {
+    test('HybridStrategy: uses last resort fallback for accounts without tokens', () => {
         const strategy = new HybridStrategy({
             healthScore: { initial: 70 },
             tokenBucket: { initialTokens: 0, maxTokens: 50 }
         });
         const accounts = createMockAccounts(3);
 
+        // No tokens, but last resort fallback should still return an account
         const result = strategy.selectAccount(accounts, 'model');
-        assertNull(result.account, 'Should filter all accounts without tokens');
+        assertNotNull(result.account, 'Last resort fallback should return an account');
+        // waitMs indicates fallback was used (500ms for lastResort)
+        assertTrue(result.waitMs >= 500, 'Last resort fallback should add throttle delay');
     });
 
     test('HybridStrategy: consumes token on selection', () => {
@@ -993,7 +1005,7 @@ async function runTests() {
         assertEqual(result.account.email, 'account2@example.com');
     });
 
-    test('Integration: Token consumption limits requests', () => {
+    test('Integration: Token exhaustion triggers last resort fallback', () => {
         const strategy = new HybridStrategy({
             tokenBucket: { initialTokens: 2, maxTokens: 10 }
         });
@@ -1003,9 +1015,11 @@ async function runTests() {
         strategy.selectAccount(accounts, 'model'); // 2 -> 1
         strategy.selectAccount(accounts, 'model'); // 1 -> 0
 
-        // Third request should fail (no tokens)
+        // Third request should use last resort fallback (not null)
         const result = strategy.selectAccount(accounts, 'model');
-        assertNull(result.account, 'Should return null when tokens exhausted');
+        assertNotNull(result.account, 'Last resort fallback should return an account');
+        // waitMs indicates fallback was used (500ms for lastResort)
+        assertTrue(result.waitMs >= 500, 'Last resort fallback should add throttle delay');
     });
 
     test('Integration: Multi-model rate limiting is independent', () => {