fix(gemini): add conversation history truncation to prevent O(N²) token cost growth
GeminiAgent sends the full conversation history with every API call, causing quadratic token growth per session. A 100-observation session sends ~30M cumulative input tokens. This ports the proven truncateHistory() sliding window from OpenRouterAgent to GeminiAgent. - Add CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES (default: 20) and CLAUDE_MEM_GEMINI_MAX_TOKENS (default: 100000) settings - Add truncateHistory() to GeminiAgent using shared estimateTokens() - Always preserve at least the newest message to avoid empty API requests - Add settings validation in SettingsRoutes (1-100 messages, 1K-1M tokens) - Add regression tests for truncation and oversized single-prompt edge case Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
vnz
@@ -18,6 +18,8 @@ import { logger } from '../../utils/logger.js';
|
||||
import { buildInitPrompt, buildObservationPrompt, buildSummaryPrompt, buildContinuationPrompt } from '../../sdk/prompts.js';
|
||||
import { SettingsDefaultsManager } from '../../shared/SettingsDefaultsManager.js';
|
||||
import { getCredential } from '../../shared/EnvManager.js';
|
||||
import { USER_SETTINGS_PATH } from '../../shared/paths.js';
|
||||
import { estimateTokens } from '../../shared/timeline-formatting.js';
|
||||
import type { ActiveSession, ConversationMessage } from '../worker-types.js';
|
||||
import { ModeManager } from '../domain/ModeManager.js';
|
||||
import {
|
||||
@@ -56,6 +58,10 @@ const GEMINI_RPM_LIMITS: Record<GeminiModel, number> = {
|
||||
// Track last request time for rate limiting
|
||||
let lastRequestTime = 0;
|
||||
|
||||
// Context window limits (prevents O(N²) token cost growth)
|
||||
const DEFAULT_MAX_CONTEXT_MESSAGES = 20; // Maximum messages to keep in conversation history
|
||||
const DEFAULT_MAX_ESTIMATED_TOKENS = 100000; // ~100k tokens max context (safety limit)
|
||||
|
||||
/**
|
||||
* Enforce RPM rate limit for Gemini free tier.
|
||||
* Waits the required time between requests based on model's RPM limit + 100ms safety buffer.
|
||||
@@ -342,6 +348,54 @@ export class GeminiAgent {
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Truncate conversation history to prevent runaway context costs.
|
||||
* Keeps most recent messages within both message count and token budget.
|
||||
* Returns a new array — never mutates the original history.
|
||||
*/
|
||||
private truncateHistory(history: ConversationMessage[]): ConversationMessage[] {
|
||||
const settings = SettingsDefaultsManager.loadFromFile(USER_SETTINGS_PATH);
|
||||
|
||||
const MAX_CONTEXT_MESSAGES = parseInt(settings.CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES) || DEFAULT_MAX_CONTEXT_MESSAGES;
|
||||
const MAX_ESTIMATED_TOKENS = parseInt(settings.CLAUDE_MEM_GEMINI_MAX_TOKENS) || DEFAULT_MAX_ESTIMATED_TOKENS;
|
||||
|
||||
if (history.length <= MAX_CONTEXT_MESSAGES) {
|
||||
// Check token count even if message count is ok
|
||||
const totalTokens = history.reduce((sum, m) => sum + estimateTokens(m.content), 0);
|
||||
if (totalTokens <= MAX_ESTIMATED_TOKENS) {
|
||||
return history;
|
||||
}
|
||||
}
|
||||
|
||||
// Sliding window: keep most recent messages within limits
|
||||
const truncated: ConversationMessage[] = [];
|
||||
let tokenCount = 0;
|
||||
|
||||
// Process messages in reverse (most recent first)
|
||||
for (let i = history.length - 1; i >= 0; i--) {
|
||||
const msg = history[i];
|
||||
const msgTokens = estimateTokens(msg.content);
|
||||
|
||||
// Always include at least the newest message — an empty contents array
|
||||
// would cause a hard Gemini API error, which is worse than an oversized request.
|
||||
if (truncated.length > 0 && (truncated.length >= MAX_CONTEXT_MESSAGES || tokenCount + msgTokens > MAX_ESTIMATED_TOKENS)) {
|
||||
logger.warn('SDK', 'Context window truncated to prevent runaway costs', {
|
||||
originalMessages: history.length,
|
||||
keptMessages: truncated.length,
|
||||
droppedMessages: i + 1,
|
||||
estimatedTokens: tokenCount,
|
||||
tokenLimit: MAX_ESTIMATED_TOKENS
|
||||
});
|
||||
break;
|
||||
}
|
||||
|
||||
truncated.unshift(msg); // Add to beginning
|
||||
tokenCount += msgTokens;
|
||||
}
|
||||
|
||||
return truncated;
|
||||
}
|
||||
|
||||
/**
|
||||
* Convert shared ConversationMessage array to Gemini's contents format
|
||||
* Maps 'assistant' role to 'model' for Gemini API compatibility
|
||||
@@ -354,8 +408,8 @@ export class GeminiAgent {
|
||||
}
|
||||
|
||||
/**
|
||||
* Query Gemini via REST API with full conversation history (multi-turn)
|
||||
* Sends the entire conversation context for coherent responses
|
||||
* Query Gemini via REST API with truncated conversation history (multi-turn)
|
||||
* Truncates history to prevent O(N²) token cost growth, then sends for coherent responses
|
||||
*/
|
||||
private async queryGeminiMultiTurn(
|
||||
history: ConversationMessage[],
|
||||
@@ -363,11 +417,13 @@ export class GeminiAgent {
|
||||
model: GeminiModel,
|
||||
rateLimitingEnabled: boolean
|
||||
): Promise<{ content: string; tokensUsed?: number }> {
|
||||
const contents = this.conversationToGeminiContents(history);
|
||||
const totalChars = history.reduce((sum, m) => sum + m.content.length, 0);
|
||||
const truncatedHistory = this.truncateHistory(history);
|
||||
const contents = this.conversationToGeminiContents(truncatedHistory);
|
||||
const totalChars = truncatedHistory.reduce((sum, m) => sum + m.content.length, 0);
|
||||
|
||||
logger.debug('SDK', `Querying Gemini multi-turn (${model})`, {
|
||||
turns: history.length,
|
||||
turns: truncatedHistory.length,
|
||||
totalTurns: history.length,
|
||||
totalChars
|
||||
});
|
||||
|
||||
|
||||
@@ -94,6 +94,8 @@ export class SettingsRoutes extends BaseRouteHandler {
|
||||
'CLAUDE_MEM_GEMINI_API_KEY',
|
||||
'CLAUDE_MEM_GEMINI_MODEL',
|
||||
'CLAUDE_MEM_GEMINI_RATE_LIMITING_ENABLED',
|
||||
'CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES',
|
||||
'CLAUDE_MEM_GEMINI_MAX_TOKENS',
|
||||
// OpenRouter Configuration
|
||||
'CLAUDE_MEM_OPENROUTER_API_KEY',
|
||||
'CLAUDE_MEM_OPENROUTER_MODEL',
|
||||
@@ -248,6 +250,22 @@ export class SettingsRoutes extends BaseRouteHandler {
|
||||
}
|
||||
}
|
||||
|
||||
// Validate CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES
|
||||
if (settings.CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES) {
|
||||
const count = parseInt(settings.CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES, 10);
|
||||
if (isNaN(count) || count < 1 || count > 100) {
|
||||
return { valid: false, error: 'CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES must be between 1 and 100' };
|
||||
}
|
||||
}
|
||||
|
||||
// Validate CLAUDE_MEM_GEMINI_MAX_TOKENS
|
||||
if (settings.CLAUDE_MEM_GEMINI_MAX_TOKENS) {
|
||||
const tokens = parseInt(settings.CLAUDE_MEM_GEMINI_MAX_TOKENS, 10);
|
||||
if (isNaN(tokens) || tokens < 1000 || tokens > 1000000) {
|
||||
return { valid: false, error: 'CLAUDE_MEM_GEMINI_MAX_TOKENS must be between 1000 and 1000000' };
|
||||
}
|
||||
}
|
||||
|
||||
// Validate CLAUDE_MEM_CONTEXT_OBSERVATIONS
|
||||
if (settings.CLAUDE_MEM_CONTEXT_OBSERVATIONS) {
|
||||
const obsCount = parseInt(settings.CLAUDE_MEM_CONTEXT_OBSERVATIONS, 10);
|
||||
|
||||
@@ -23,6 +23,8 @@ export interface SettingsDefaults {
|
||||
CLAUDE_MEM_GEMINI_API_KEY: string;
|
||||
CLAUDE_MEM_GEMINI_MODEL: string; // 'gemini-2.5-flash-lite' | 'gemini-2.5-flash' | 'gemini-3-flash-preview'
|
||||
CLAUDE_MEM_GEMINI_RATE_LIMITING_ENABLED: string; // 'true' | 'false' - enable rate limiting for free tier
|
||||
CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES: string; // Max messages in Gemini context window (prevents O(N²) cost growth)
|
||||
CLAUDE_MEM_GEMINI_MAX_TOKENS: string; // Max estimated tokens for Gemini context (~100k safety limit)
|
||||
CLAUDE_MEM_OPENROUTER_API_KEY: string;
|
||||
CLAUDE_MEM_OPENROUTER_MODEL: string;
|
||||
CLAUDE_MEM_OPENROUTER_SITE_URL: string;
|
||||
@@ -82,6 +84,8 @@ export class SettingsDefaultsManager {
|
||||
CLAUDE_MEM_GEMINI_API_KEY: '', // Empty by default, can be set via UI or env
|
||||
CLAUDE_MEM_GEMINI_MODEL: 'gemini-2.5-flash-lite', // Default Gemini model (highest free tier RPM)
|
||||
CLAUDE_MEM_GEMINI_RATE_LIMITING_ENABLED: 'true', // Rate limiting ON by default for free tier users
|
||||
CLAUDE_MEM_GEMINI_MAX_CONTEXT_MESSAGES: '20', // Max messages in Gemini context window
|
||||
CLAUDE_MEM_GEMINI_MAX_TOKENS: '100000', // Max estimated tokens (~100k safety limit)
|
||||
CLAUDE_MEM_OPENROUTER_API_KEY: '', // Empty by default, can be set via UI or env
|
||||
CLAUDE_MEM_OPENROUTER_MODEL: 'xiaomi/mimo-v2-flash:free', // Default OpenRouter model (free tier)
|
||||
CLAUDE_MEM_OPENROUTER_SITE_URL: '', // Optional: for OpenRouter analytics
|
||||
|
||||
Reference in New Issue
Block a user