feat: v2.2.92 → v2.2.158 — god-file 분해 + Stocks feature + 대화 연속성

R56–R59: agent.ts 2731→1529줄 god-file 분해 (25 modules)
  · attrParsers + LLM 메서드 8개 (callNonStreaming, streamChatOnce 등)
  · executeActions 415줄 → 8 handler 그룹 (file/run/list/brain/calendar/sheets/tasks)
  · handlePrompt 1100줄 → 7 phase 모듈 (system prompt + budget + autoContinue 등)

R50–R55: extension.ts 1145→349줄 (telegram/settings/provider commands 분리)

Stocks feature 신규: /stocks slash command (v2.2.152~158)
  · .astra/stocks.json 저장소 + Yahoo Finance 현재가 갱신
  · 8 키워드 필터 (ROE/성장성/유동성/수익성/영업효율/기술력/안정성/PBR)
  · Naver 시가총액 페이지 JSON API (m.stock.naver.com) 발굴
  · LLM Top 5 매력도 분석 + Telegram 자동 보고서
  · KST 09:00/15:00 watcher 자동 모니터링

대화 연속성 (v2.2.150~157):
  · [PRIOR TURN CONCLUSION] block 으로 직전 결론 anchor
  · thin follow-up 분류 → boilerplate 헤더 suppression
  · slash 명령 결과 chatHistory mirror (capture wrapper)
  · echo/parrot 금지 system prompt rule

기타: /stocks 슬래시 자동완성 dropdown UI, Naver JSON API 전환 (cheerio 제거)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
g1nation
2026-05-25 09:59:32 +09:00
parent 4153f640c2
commit 0a97324f1b
149 changed files with 14628 additions and 6927 deletions
@@ -0,0 +1,161 @@
import { logInfo, logError } from '../../utils';
import type { ChatMessage } from '../../agent';
import {
estimateTokens,
estimateMessagesTokens,
computeOutputBudget,
trimHistoryToBudget,
truncateSystemPromptContext,
estimateModelParamsB,
type ContextLimits,
} from '../../lib/contextManager';
import { buildDroppedHistorySummary } from '../../lib/contextBuilders/droppedHistorySummary';
export interface ComputeBudgetedRequestInput {
fullSystemPrompt: string;
/** Caller is expected to have run `capChatHistory` on this already. */
reqMessages: ChatMessage[];
actualModel: string;
/** Result of `getConfig()` — reads contextLength, maxOutputTokens, contextSafetyMargin, smallModelContextCap, autoCompactHistory. */
config: any;
imageCount: number;
}
export interface ComputeBudgetedRequestResult {
messagesForRequest: ChatMessage[];
ctxLimits: ContextLimits;
inputTokens: number;
maxOutputTokens: number;
systemTokens: number;
systemTruncated: boolean;
droppedHistoryCount: number;
budgetedHistoryLength: number;
/** Exact return shape of `computeOutputBudget`. */
outputBudget: { maxOutputTokens: number; available: number; tight: boolean };
modelParamB: number | null;
cappedForSmallModel: boolean;
}
/**
* 입력(시스템 프롬프트 + 대화 기록 + 이미지)을 컨텍스트 윈도우 예산에 맞게 정리하고
* 최종 요청 메시지 배열과 동적 출력 상한을 계산합니다.
*
* 호출 측에서 미리 capChatHistory 로 메시지 개수를 캡한 뒤 넘겨주는 것을 전제로 합니다
* (AgentExecutor.MAX_RETAINED_MESSAGES 같은 정적 한도는 이 함수의 관심사가 아닙니다).
*/
export function computeBudgetedRequest(input: ComputeBudgetedRequestInput): ComputeBudgetedRequestResult {
const { fullSystemPrompt, reqMessages, actualModel, config, imageCount } = input;
// ──────────────────────────────────────────────────────────────────
// [Context Limit Manager] context length 는 "답변을 그만큼 길게 써도 된다"
// 는 뜻이 아니다: 시스템 프롬프트 + 대화 기록 + 입력 + 생성될 답변 + 여유분 ≤ context length.
// 요청을 보내기 전에 입력 토큰을 추정해서
// (1) 시스템 프롬프트가 과하면 [CONTEXT] 블록을 마지막 수단으로 줄이고
// (2) 대화 기록을 남은 예산에 맞게 압축하고 (UI 표시용 chatHistory 는 건드리지 않음)
// (3) 동적으로 출력 상한(maxOutputTokens)을 계산한다.
// ──────────────────────────────────────────────────────────────────
// Optional opt-in guard (g1nation.smallModelContextCap, OFF/0 by default): some very small
// models (≤3B) emit EOS as the first token when the prompt is near their context window
// even though it nominally fits. If the user opted in, budget ≤3B models against that
// smaller effective window. Never applied to 4B+ models, and never when the setting is 0 —
// capping squeezes the output-token budget, so it's a knob, not a default.
const modelParamB = estimateModelParamsB(actualModel);
const smallModelCap = config.smallModelContextCap; // 0 = disabled (default)
const cappedForSmallModel = smallModelCap > 0
&& modelParamB !== null && modelParamB <= 3
&& config.contextLength > smallModelCap;
const effectiveContextLength = cappedForSmallModel ? smallModelCap : config.contextLength;
if (cappedForSmallModel) {
logInfo('Small model detected — capping effective context window for budgeting.', {
model: actualModel, paramB: modelParamB,
nominalContext: config.contextLength, effectiveContext: effectiveContextLength,
});
}
const ctxLimits: ContextLimits = {
contextLength: effectiveContextLength,
maxOutputTokens: config.maxOutputTokens,
safetyMargin: config.contextSafetyMargin,
minOutputTokens: 512,
};
const imageTokenReserve = imageCount * 1024;
// Output budget we ACTUALLY reserve before trimming — not the bare
// minOutputTokens floor (512). If we only reserve 512, a long session
// is allowed to grow the prompt until ~512-1k tokens remain for the
// answer; small/MoE local models (e.g. gemma 4B-active) then emit EOS
// as the first token and return an empty response. Reserving ~10% of
// the window (>=2048) forces history/system trimming to keep a real
// answer-sized hole open. Capped at maxOutputTokens.
const preferredOutputReserve = Math.min(
ctxLimits.maxOutputTokens,
Math.max(2048, Math.floor(ctxLimits.contextLength * 0.1))
);
// (1) 시스템 프롬프트는 예산의 ~65%까지만 허용 — 그 이상이면 [CONTEXT] 블록부터 잘라낸다.
const systemCapTokens = Math.max(
1024,
Math.floor((ctxLimits.contextLength - ctxLimits.safetyMargin - preferredOutputReserve - imageTokenReserve) * 0.65)
);
const { prompt: budgetedSystemPrompt, truncated: systemTruncated } =
truncateSystemPromptContext(fullSystemPrompt, systemCapTokens);
if (systemTruncated) {
logInfo('System prompt context truncated to fit the context window.', { model: actualModel, systemCapTokens });
}
const systemTokens = estimateTokens(budgetedSystemPrompt) + 4;
// (2) 대화 기록 압축.
const historyBudget = Math.max(
256,
ctxLimits.contextLength - systemTokens - ctxLimits.safetyMargin - preferredOutputReserve - imageTokenReserve
);
let budgetedHistory: ChatMessage[] = reqMessages;
if (config.autoCompactHistory) {
// v2.2.69 — dropped 메시지를 받아 heuristic 요약을 만든 뒤 한 system 메시지로 prepend.
// 단순 count 마커는 "이전에 무슨 얘기를 했는지" 를 전혀 알려주지 않아 후속 턴에서 모델이
// 맥락을 잃어버리는 회귀를 낳았다. 이제는 U1/A1/U2/A2 골자가 남아 sliding window 가 동작.
const trim = trimHistoryToBudget<ChatMessage>(reqMessages, historyBudget, (_n, dropped) => ({
role: 'system',
content: buildDroppedHistorySummary(dropped),
internal: true,
}));
budgetedHistory = trim.messages;
if (trim.droppedCount > 0) {
logInfo('Conversation history compacted to fit the context window (with summary).', {
model: actualModel, droppedCount: trim.droppedCount, historyBudget,
});
}
}
const messagesForRequest: ChatMessage[] = [
{ role: 'system', content: budgetedSystemPrompt, internal: true },
...budgetedHistory
];
// (3) 동적 출력 상한.
const inputTokens = estimateMessagesTokens(messagesForRequest) + imageTokenReserve;
const outputBudget = computeOutputBudget(inputTokens, ctxLimits);
const maxOutputTokens = outputBudget.maxOutputTokens;
if (outputBudget.tight) {
logError('Prompt nearly fills the context window — output budget is at the minimum.', {
model: actualModel, contextLength: ctxLimits.contextLength, inputTokens, maxOutputTokens,
});
}
logInfo('Context budget computed.', {
model: actualModel, contextLength: ctxLimits.contextLength,
inputTokens, maxOutputTokens, droppedHistory: reqMessages.length - budgetedHistory.length,
});
return {
messagesForRequest,
ctxLimits,
inputTokens,
maxOutputTokens,
systemTokens,
systemTruncated,
droppedHistoryCount: reqMessages.length - budgetedHistory.length,
budgetedHistoryLength: budgetedHistory.length,
outputBudget,
modelParamB,
cappedForSmallModel,
};
}