File size: 16,644 Bytes
7a1ad33 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 | /**
* @license
* Copyright 2025 Google LLC
* SPDX-License-Identifier: Apache-2.0
*/
import type { Content } from '@google/genai';
import type { Config } from '../config/config.js';
import type { GeminiChat } from '../core/geminiChat.js';
import { type ChatCompressionInfo, CompressionStatus } from '../core/turn.js';
import { tokenLimit } from '../core/tokenLimits.js';
import { getCompressionPrompt } from '../core/prompts.js';
import { getResponseText } from '../utils/partUtils.js';
import { logChatCompression } from '../telemetry/loggers.js';
import { makeChatCompressionEvent, LlmRole } from '../telemetry/types.js';
import {
saveTruncatedToolOutput,
formatTruncatedToolOutput,
} from '../utils/fileUtils.js';
import { debugLogger } from '../utils/debugLogger.js';
import { getInitialChatHistory } from '../utils/environmentContext.js';
import {
calculateRequestTokenCount,
estimateTokenCountSync,
} from '../utils/tokenCalculation.js';
import {
DEFAULT_GEMINI_FLASH_LITE_MODEL,
DEFAULT_GEMINI_FLASH_MODEL,
DEFAULT_GEMINI_MODEL,
PREVIEW_GEMINI_MODEL,
PREVIEW_GEMINI_FLASH_MODEL,
PREVIEW_GEMINI_3_1_MODEL,
PREVIEW_GEMINI_FLASH_LITE_MODEL,
} from '../config/models.js';
import { PreCompressTrigger } from '../hooks/types.js';
/**
* Default threshold for compression token count as a fraction of the model's
* token limit. If the chat history exceeds this threshold, it will be compressed.
*/
const DEFAULT_COMPRESSION_TOKEN_THRESHOLD = 0.5;
/**
* The fraction of the latest chat history to keep. A value of 0.3
* means that only the last 30% of the chat history will be kept after compression.
*/
const COMPRESSION_PRESERVE_THRESHOLD = 0.3;
/**
* The budget for function response tokens in the preserved history.
*/
const COMPRESSION_FUNCTION_RESPONSE_TOKEN_BUDGET = 50_000;
/**
* Returns the index of the oldest item to keep when compressing. May return
* contents.length which indicates that everything should be compressed.
*
* Exported for testing purposes.
*/
export function findCompressSplitPoint(
contents: Content[],
fraction: number,
): number {
if (fraction <= 0 || fraction >= 1) {
throw new Error('Fraction must be between 0 and 1');
}
const charCounts = contents.map((content) => JSON.stringify(content).length);
const totalCharCount = charCounts.reduce((a, b) => a + b, 0);
const targetCharCount = totalCharCount * fraction;
let lastSplitPoint = 0; // 0 is always valid (compress nothing)
let cumulativeCharCount = 0;
for (let i = 0; i < contents.length; i++) {
const content = contents[i];
if (
content.role === 'user' &&
!content.parts?.some((part) => !!part.functionResponse)
) {
if (cumulativeCharCount >= targetCharCount) {
return i;
}
lastSplitPoint = i;
}
cumulativeCharCount += charCounts[i];
}
// We found no split points after targetCharCount.
// Check if it's safe to compress everything.
const lastContent = contents[contents.length - 1];
if (
lastContent?.role === 'model' &&
!lastContent?.parts?.some((part) => part.functionCall)
) {
return contents.length;
}
// Can't compress everything so just compress at last splitpoint.
return lastSplitPoint;
}
export function modelStringToModelConfigAlias(model: string): string {
switch (model) {
case PREVIEW_GEMINI_MODEL:
case PREVIEW_GEMINI_3_1_MODEL:
return 'chat-compression-3-pro';
case PREVIEW_GEMINI_FLASH_MODEL:
return 'chat-compression-3-flash';
case PREVIEW_GEMINI_FLASH_LITE_MODEL:
// fallthrough
case DEFAULT_GEMINI_FLASH_LITE_MODEL:
return 'chat-compression-3.1-flash-lite';
case 'gemini-2.5-flash-lite':
return 'chat-compression-2.5-flash-lite';
case DEFAULT_GEMINI_MODEL:
return 'chat-compression-2.5-pro';
case DEFAULT_GEMINI_FLASH_MODEL:
return 'chat-compression-2.5-flash';
default:
return 'chat-compression-default';
}
}
/**
* Processes the chat history to ensure function responses don't exceed a specific token budget.
*
* This function implements a "Reverse Token Budget" strategy:
* 1. It iterates through the history from the most recent turn to the oldest.
* 2. It keeps a running tally of tokens used by function responses.
* 3. Recent tool outputs are preserved in full to maintain high-fidelity context for the current turn.
* 4. Once the budget (COMPRESSION_FUNCTION_RESPONSE_TOKEN_BUDGET) is exceeded, any older large
* tool responses are truncated to their last 30 lines and saved to a temporary file.
*
* This ensures that compression effectively reduces context size even when recent turns
* contain massive tool outputs (like large grep results or logs).
*/
async function truncateHistoryToBudget(
history: readonly Content[],
config: Config,
): Promise<Content[]> {
let functionResponseTokenCounter = 0;
const truncatedHistory: Content[] = [];
// Iterate backwards: newest messages first to prioritize their context.
for (let i = history.length - 1; i >= 0; i--) {
const content = history[i];
const newParts = [];
if (content.parts) {
// Process parts of the message backwards as well.
for (let j = content.parts.length - 1; j >= 0; j--) {
const part = content.parts[j];
if (part.functionResponse) {
const responseObj = part.functionResponse.response;
// Ensure we have a string representation to truncate.
// If the response is an object, we try to extract a primary string field (output or content).
let contentStr: string;
if (typeof responseObj === 'string') {
contentStr = responseObj;
} else if (responseObj && typeof responseObj === 'object') {
if (
'output' in responseObj &&
// eslint-disable-next-line no-restricted-syntax
typeof responseObj['output'] === 'string'
) {
contentStr = responseObj['output'];
} else if (
'content' in responseObj &&
// eslint-disable-next-line no-restricted-syntax
typeof responseObj['content'] === 'string'
) {
contentStr = responseObj['content'];
} else {
contentStr = JSON.stringify(responseObj, null, 2);
}
} else {
contentStr = JSON.stringify(responseObj, null, 2);
}
const tokens = estimateTokenCountSync([{ text: contentStr }]);
if (
functionResponseTokenCounter + tokens >
COMPRESSION_FUNCTION_RESPONSE_TOKEN_BUDGET
) {
try {
// Budget exceeded: Truncate this response.
const { outputFile } = await saveTruncatedToolOutput(
contentStr,
part.functionResponse.name ?? 'unknown_tool',
config.getNextCompressionTruncationId(),
config.storage.getProjectTempDir(),
);
const truncatedMessage = formatTruncatedToolOutput(
contentStr,
outputFile,
config.getTruncateToolOutputThreshold(),
);
newParts.unshift({
functionResponse: {
// eslint-disable-next-line @typescript-eslint/no-misused-spread
...part.functionResponse,
response: { output: truncatedMessage },
},
});
// Count the small truncated placeholder towards the budget.
functionResponseTokenCounter += estimateTokenCountSync([
{ text: truncatedMessage },
]);
} catch (error) {
// Fallback: if truncation fails, keep the original part to avoid data loss in the chat.
debugLogger.debug('Failed to truncate history to budget:', error);
newParts.unshift(part);
functionResponseTokenCounter += tokens;
}
} else {
// Within budget: keep the full response.
functionResponseTokenCounter += tokens;
newParts.unshift(part);
}
} else {
// Non-tool response part: always keep.
newParts.unshift(part);
}
}
}
// Reconstruct the message with processed (potentially truncated) parts.
truncatedHistory.unshift({ ...content, parts: newParts });
}
return truncatedHistory;
}
export class ChatCompressionService {
async compress(
chat: GeminiChat,
promptId: string,
force: boolean,
model: string,
config: Config,
hasFailedCompressionAttempt: boolean,
abortSignal?: AbortSignal,
): Promise<{ newHistory: Content[] | null; info: ChatCompressionInfo }> {
const curatedHistory = chat.getHistory(true);
// Regardless of `force`, don't do anything if the history is empty.
if (curatedHistory.length === 0) {
return {
newHistory: null,
info: {
originalTokenCount: 0,
newTokenCount: 0,
compressionStatus: CompressionStatus.NOOP,
},
};
}
// Fire PreCompress hook before compression
// This fires for both manual and auto compression attempts
const trigger = force ? PreCompressTrigger.Manual : PreCompressTrigger.Auto;
await config.getHookSystem()?.firePreCompressEvent(trigger);
const originalTokenCount = chat.getLastPromptTokenCount();
// Don't compress if not forced and we are under the limit.
if (!force) {
const threshold =
(await config.getCompressionThreshold()) ??
DEFAULT_COMPRESSION_TOKEN_THRESHOLD;
if (originalTokenCount < threshold * tokenLimit(model)) {
return {
newHistory: null,
info: {
originalTokenCount,
newTokenCount: originalTokenCount,
compressionStatus: CompressionStatus.NOOP,
},
};
}
}
// Apply token-based truncation to the entire history before splitting.
// This ensures that even the "to compress" portion is within safe limits for the summarization model.
const truncatedHistory = await truncateHistoryToBudget(
curatedHistory,
config,
);
// If summarization previously failed (and not forced), we only rely on truncation.
// We do NOT attempt to invoke the LLM for summarization again to avoid repeated failures/costs.
if (hasFailedCompressionAttempt && !force) {
const truncatedTokenCount = estimateTokenCountSync(
truncatedHistory.flatMap((c) => c.parts || []),
);
// If truncation reduced the size, we consider it a successful "compression" (truncation only).
if (truncatedTokenCount < originalTokenCount) {
return {
newHistory: truncatedHistory,
info: {
originalTokenCount,
newTokenCount: truncatedTokenCount,
compressionStatus: CompressionStatus.CONTENT_TRUNCATED,
},
};
}
return {
newHistory: null,
info: {
originalTokenCount,
newTokenCount: originalTokenCount,
compressionStatus: CompressionStatus.NOOP,
},
};
}
const splitPoint = findCompressSplitPoint(
truncatedHistory,
1 - COMPRESSION_PRESERVE_THRESHOLD,
);
const historyToCompressTruncated = truncatedHistory.slice(0, splitPoint);
const historyToKeepTruncated = truncatedHistory.slice(splitPoint);
if (historyToCompressTruncated.length === 0) {
return {
newHistory: null,
info: {
originalTokenCount,
newTokenCount: originalTokenCount,
compressionStatus: CompressionStatus.NOOP,
},
};
}
// High Fidelity Decision: Should we send the original or truncated history to the summarizer?
const originalHistoryToCompress = curatedHistory.slice(0, splitPoint);
const originalToCompressTokenCount = estimateTokenCountSync(
originalHistoryToCompress.flatMap((c) => c.parts || []),
);
const historyForSummarizer =
originalToCompressTokenCount < tokenLimit(model)
? originalHistoryToCompress
: historyToCompressTruncated;
const hasPreviousSnapshot = historyForSummarizer.some((c) =>
c.parts?.some((p) => p.text?.includes('<state_snapshot>')),
);
const anchorInstruction = hasPreviousSnapshot
? 'A previous <state_snapshot> exists in the history. You MUST integrate all still-relevant information from that snapshot into the new one, updating it with the more recent events. Do not lose established constraints or critical knowledge.'
: 'Generate a new <state_snapshot> based on the provided history.';
const summaryResponse = await config.getBaseLlmClient().generateContent({
modelConfigKey: { model: modelStringToModelConfigAlias(model) },
contents: [
...historyForSummarizer,
{
role: 'user',
parts: [
{
text: `${anchorInstruction}\n\nFirst, reason in your scratchpad. Then, generate the updated <state_snapshot>.`,
},
],
},
],
systemInstruction: { text: getCompressionPrompt(config) },
promptId,
// TODO(joshualitt): wire up a sensible abort signal,
abortSignal: abortSignal ?? new AbortController().signal,
role: LlmRole.UTILITY_COMPRESSOR,
});
const summary = getResponseText(summaryResponse) ?? '';
// Phase 3: The "Probe" Verification (Self-Correction)
// We perform a second lightweight turn to ensure no critical information was lost.
const verificationResponse = await config
.getBaseLlmClient()
.generateContent({
modelConfigKey: { model: modelStringToModelConfigAlias(model) },
contents: [
...historyForSummarizer,
{
role: 'model',
parts: [{ text: summary }],
},
{
role: 'user',
parts: [
{
text: 'Critically evaluate the <state_snapshot> you just generated. Did you omit any specific technical details, file paths, tool results, or user constraints mentioned in the history? If anything is missing or could be more precise, generate a FINAL, improved <state_snapshot>. Otherwise, repeat the exact same <state_snapshot> again.',
},
],
},
],
systemInstruction: { text: getCompressionPrompt(config) },
promptId: `${promptId}-verify`,
role: LlmRole.UTILITY_COMPRESSOR,
abortSignal: abortSignal ?? new AbortController().signal,
});
const finalSummary = (
getResponseText(verificationResponse)?.trim() || summary
).trim();
if (!finalSummary) {
logChatCompression(
config,
makeChatCompressionEvent({
tokens_before: originalTokenCount,
tokens_after: originalTokenCount, // No change since it failed
}),
);
return {
newHistory: null,
info: {
originalTokenCount,
newTokenCount: originalTokenCount,
compressionStatus: CompressionStatus.COMPRESSION_FAILED_EMPTY_SUMMARY,
},
};
}
const extraHistory: Content[] = [
{
role: 'user',
parts: [{ text: finalSummary }],
},
{
role: 'model',
parts: [{ text: 'Got it. Thanks for the additional context!' }],
},
...historyToKeepTruncated,
];
// Use a shared utility to construct the initial history for an accurate token count.
const fullNewHistory = await getInitialChatHistory(config, extraHistory);
const newTokenCount = await calculateRequestTokenCount(
fullNewHistory.flatMap(
(c) => ('content' in c ? c.content.parts : c.parts) || [],
),
config.getContentGenerator(),
model,
);
logChatCompression(
config,
makeChatCompressionEvent({
tokens_before: originalTokenCount,
tokens_after: newTokenCount,
}),
);
if (newTokenCount > originalTokenCount) {
return {
newHistory: null,
info: {
originalTokenCount,
newTokenCount,
compressionStatus:
CompressionStatus.COMPRESSION_FAILED_INFLATED_TOKEN_COUNT,
},
};
} else {
return {
newHistory: extraHistory,
info: {
originalTokenCount,
newTokenCount,
compressionStatus: CompressionStatus.COMPRESSED,
},
};
}
}
}
|