fix(agent): v2.2.62 — 출력 degeneration 방어

- 재시작 감지: auto-continuation이 "이어쓰기" 대신 답변을 처음부터
  재생성하면 버림 → 분석이 두 번 나오던 문제 제거
- degeneration 정리 패스(cleanDegeneratedOutput): 문자 벽(같은 기호 8개+),
  (Note:…) 메타 노트, Candidate records 내부 지시문 누출,
  (질문 의도:…)/[핵심 확인 질문] 누출, 연속 중복 문단 제거

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

.astra/project-context/architecture.md

@@ -3,15 +3,15 @@
 <!-- ASTRA:AUTO-START -->
 
 ## Snapshot
-- **Workspace**: `connectai` `v2.2.61` _(absolute path varies by environment; resolved from the active VS Code workspace)_
+- **Workspace**: `connectai` `v2.2.62` _(absolute path varies by environment; resolved from the active VS Code workspace)_
 - **Description**: The personal intelligence layer for Antigravity and VS Code. A private cognitive partner for deep project context, memory, and proactive strategic decision-making.
 - **Stack**: TypeScript, Node.js, VS Code Extension, LM Studio SDK, Test runner
-- **Stats**: 272 source files, ~55,207 lines across 5 top-level modules.
+- **Stats**: 273 source files, ~55,326 lines across 5 top-level modules.
 
 ## Last Refresh
-- **Time**: 2026-05-22T06:25:57.973Z
+- **Time**: 2026-05-22T06:32:01.903Z
 - **Files newly analysed**: 1
-- **Files reused from cache**: 271
+- **Files reused from cache**: 272
 
 ## Directory Map
 ```mermaid
@@ -42,7 +42,7 @@ flowchart LR
     media["media/<br/>6 files"]
     tests["tests/<br/>35 files"]
     core_py["core_py/<br/>6 files"]
-    docs["docs/<br/>86 files"]
+    docs["docs/<br/>87 files"]
     tests --> src
 ```
 
@@ -65,7 +65,7 @@ flowchart LR
 
 ## Modules
 
-### `src/` — 139 files, ~38,185 lines
+### `src/` — 139 files, ~38,285 lines
 
 **Sub-directories**
 - `src/features/` (66) — Astra Office — public API. 다음 세션에서 추가될 OfficeSnapshot presenter / schema 도 같은 entry 로 노출 예정. 현재 노출: full webview panel H
@@ -93,7 +93,7 @@ flowchart LR
 - `src/retrieval/scoring.ts` (536 lines) — Scoring Engine — TF-IDF + Bilingual Tokenizer 단순 includes() 키워드 매칭을 넘어서, TF-IDF 가중치 기반의 문서 스코어링을 제공합니다. 한국어/영어 양국어 토크나이저를 포함합니다.
 - `src/skills/agentKnowledgeMap.ts` (374 lines)
 - `src/retrieval/lessonHelpers.ts` (325 lines) — Lesson / Experience Memory — pure helpers (no vscode dependency) "Lesson" = a markdown file in the active brain that captures a past mistake/risk and how to avoid repeating it. Identified by a lessons
-- `src/agent.ts` (3722 lines)
+- `src/agent.ts` (3822 lines)
 - `src/features/providers/types.ts` (63 lines) — Cloud LLM provider routing — model id prefix → provider id 매핑. Prefix 규칙: openrouter:anthropic/claude-3.5-sonnet → { provider: 'openrouter', model: 'anthropic/claude-3.5-sonnet' } anthropic:claude-3-5
 - `src/lib/engine.ts` (906 lines)
 - `src/retrieval/brainIndex.ts` (325 lines) — Brain Index — persistent, mtime-keyed tokenized cache of the Second Brain RAG 검색은 매 질의마다 브레인의 모든 .md 파일을 읽고 토크나이즈해서 TF-IDF 점수를 계산했습니다 — 파일 수가 많아지면 그게 병목입니다. 이 모듈은 <brainPath>/.astra/brain-index.json 에
@@ -161,10 +161,10 @@ flowchart LR
 - `core_py/optimizer.py` (55 lines)
 - `core_py/queue_worker.py` (82 lines)
 
-### `docs/` — 86 files, ~3,316 lines
+### `docs/` — 87 files, ~3,335 lines
 
 **Sub-directories**
-- `docs/records/` (73) — Bug: /Volumes/Data/project/Antigravity/ConnectAI 프로젝트 코드 리뷰 해줄 수 있어? 개선할 부분이 있는지, 그러고...
+- `docs/records/` (74) — Bug: /Volumes/Data/project/Antigravity/ConnectAI 프로젝트 코드 리뷰 해줄 수 있어? 개선할 부분이 있는지, 그러고...
 - `docs/docs/` (5) — Bug: Viewed integrationretrieval.test.ts:1-59 integrationretrieval.test.ts를 통해 ...
 - `docs/Meeting/` (0)
 
@@ -175,7 +175,7 @@ flowchart LR
 - `docs/EXPERIENCE_MEMORY_PLAN.md` (122 lines) — Experience Memory (Mistake / Lesson Loop) — Implementation Plan
 - `docs/records/ConnectAI/development/2026-05-02_connectai_project_knowledge_overview.md` (121 lines) — Astra Project Knowledge Overview
 - `docs/records/ConnectAI/development/2026-05-03_connectai_project_knowledge_overview.md` (121 lines) — Astra Project Knowledge Overview
-- `docs/records/ConnectAI/timeline.md` (170 lines) — Project Timeline
+- `docs/records/ConnectAI/timeline.md` (173 lines) — Project Timeline
 - `docs/Advanced_Features_Implementation_Guide.md` (40 lines) — Advanced Features Implementation Guide
 - `docs/docs/records/docs/bugs/BUG-0001-viewed-integration-retrieval-test-ts-1-59-integration-retrie.md` (16 lines) — Bug: Viewed integrationretrieval.test.ts:1-59 integrationretrieval.test.ts를 통해 ...
 - `docs/docs/records/docs/chronicle.config.json` (11 lines) — JSON configuration
@@ -336,7 +336,7 @@ Astra는 대표님의 명시적인 승인 하에 로컬 시스템의 강력한
 **Designed for High-Performance Decision Making.**
 Copyright (C) **g1nation**. All rights reserved.
 
-_Last auto-scan: 2026-05-22T06:25:57.973Z · signature `32c602bb`_
+_Last auto-scan: 2026-05-22T06:32:01.903Z · signature `26f10f3d`_
 <!-- ASTRA:AUTO-END -->
 
 ## Purpose

.astra/project-context/scan-cache.json

@@ -1,11 +1,11 @@
 {
   "version": 1,
-  "generatedAt": "2026-05-22T06:25:58.013Z",
+  "generatedAt": "2026-05-22T06:32:01.938Z",
   "files": {
     "src/agent.ts": {
-      "mtimeMs": 1779430182152.1404,
+      "mtimeMs": 1779431481976.8423,
-      "size": 217774,
+      "size": 220892,
-      "lines": 3722,
+      "lines": 3822,
       "role": "",
       "imports": [
         "src/utils",
@@ -1977,7 +1977,7 @@
       "imports": []
     },
     "docs/records/ConnectAI/chronicle.config.json": {
-      "mtimeMs": 1779430497353.3203,
+      "mtimeMs": 1779431169044.1296,
       "size": 371,
       "lines": 11,
       "role": "JSON configuration",
@@ -2333,6 +2333,13 @@
       "role": "Discussion: 최성연 알아? 칼리버스에서 근무한데",
       "imports": []
     },
+    "docs/records/ConnectAI/discussions/2026-05-22_e-wiki-connectai-git에-커밋하고-푸쉬해줘.md": {
+      "mtimeMs": 1779431169034.8508,
+      "size": 745,
+      "lines": 16,
+      "role": "Discussion: E:\\Wiki\\connectai git에 커밋하고 푸쉬해줘.",
+      "imports": []
+    },
     "docs/records/ConnectAI/planning/2026-05-02_project-chronicle-guard.md": {
       "mtimeMs": 1778028987339.2349,
       "size": 3004,
@@ -2383,9 +2390,9 @@
       "imports": []
     },
     "docs/records/ConnectAI/timeline.md": {
-      "mtimeMs": 1779430149729.8076,
+      "mtimeMs": 1779431169036.8623,
-      "size": 11074,
+      "size": 11199,
-      "lines": 170,
+      "lines": 173,
       "role": "Project Timeline",
       "imports": []
     },

.astra/tests/stress/.astra/cache/259a37934ead3910a8722b82054d46d2ca2057b05c488be1dcf439166ac5a9a1.json

@@ -1,5 +1,5 @@
 {
   "result": "Final report with inconsistencies. This should be long enough to pass validation.",
-  "createdAt": 1779430407981,
+  "createdAt": 1779431564297,
   "modelVersion": "unknown"
 }

.astra/tests/stress/.astra/cache/65775be352df43297b63c7af59c9f4f39d2bc368f77456c37b5eef9a94a66b5c.json

@@ -1,5 +1,5 @@
 {
   "result": "[CONFLICT WARNING] 성능이 200% 증가했습니다. vs 그러나 동시에 50% 감소했습니다. 최적화와 성능 저하가 동시에 발견됨.",
-  "createdAt": 1779430407979,
+  "createdAt": 1779431564294,
   "modelVersion": "unknown"
 }

.astra/tests/stress/.astra/cache/6894d26c5b0a55d25d756a473225c7a44d7661af673b24e3f49551a7a2e50280.json

@@ -1,5 +1,5 @@
 {
   "result": "Detailed Execution Plan: 1. Research 2. Analyze 3. Write report with high quality.",
-  "createdAt": 1779430407976,
+  "createdAt": 1779431564292,
   "modelVersion": "unknown"
 }

.astra/tests/stress/.astra/cache/88cb61499f88ed38165b64bd3e8adc543795e4b427b64540a49c9ab27c7fe213.json

@@ -1,5 +1,5 @@
 {
-  "result": "---\nid: stress_conflict_1779430407960\ndate: 2026-05-22T06:13:27.983Z\ntype: knowledge_artifact\nstandard: P-Reinforce v3.0\ntags: [automated, connect_ai, brain_sync]\n---\n\n## 📌 Brief Summary\nFinal report with inconsistencies. This should be long enough to pass validation.\n\nFinal report with inconsistencies. This should be long enough to pass validation.\n\n---\n## 💡 Astra의 선제적 제안 (Proactive Next Actions)\nFinal report with inconsistencies. This should be long enough to pass validation.\n---\n## 🛡️ Reliability & Audit Summary\n> [!NOTE]\n> 이 문서는 ConnectAI의 **Intelligent Resilience** 엔진에 의해 검증 및 정제되었습니다.\n\n| Metric | Value | Status |\n| :--- | :--- | :--- |\n| **Conflict Risk** | `60/100` | ⚠️ Medium |\n| **Fallbacks Used** | `0` | ✅ None |\n| **Auto Retries** | `0` | ✅ Stable |\n| **Deduplication** | `0` | Standard |\n| **Processing Time** | `0.0s` | ✅ Fast |\n\n### 🔍 Decision Audit Trail\n- **[PLANNER]** 전략 수립 중... (15ms)\n- **[RESEARCHER]** 핵심 정보 수집 및 분석 중... (2ms)\n- **[WRITER]** 최종 리포트 작성 및 편집 중... (4ms)\n",
+  "result": "---\nid: stress_conflict_1779431564276\ndate: 2026-05-22T06:32:44.299Z\ntype: knowledge_artifact\nstandard: P-Reinforce v3.0\ntags: [automated, connect_ai, brain_sync]\n---\n\n## 📌 Brief Summary\nFinal report with inconsistencies. This should be long enough to pass validation.\n\nFinal report with inconsistencies. This should be long enough to pass validation.\n\n---\n## 💡 Astra의 선제적 제안 (Proactive Next Actions)\nFinal report with inconsistencies. This should be long enough to pass validation.\n---\n## 🛡️ Reliability & Audit Summary\n> [!NOTE]\n> 이 문서는 ConnectAI의 **Intelligent Resilience** 엔진에 의해 검증 및 정제되었습니다.\n\n| Metric | Value | Status |\n| :--- | :--- | :--- |\n| **Conflict Risk** | `60/100` | ⚠️ Medium |\n| **Fallbacks Used** | `0` | ✅ None |\n| **Auto Retries** | `0` | ✅ Stable |\n| **Deduplication** | `0` | Standard |\n| **Processing Time** | `0.0s` | ✅ Fast |\n\n### 🔍 Decision Audit Trail\n- **[PLANNER]** 전략 수립 중... (14ms)\n- **[RESEARCHER]** 핵심 정보 수집 및 분석 중... (3ms)\n- **[WRITER]** 최종 리포트 작성 및 편집 중... (3ms)\n",
-  "createdAt": 1779430407983,
+  "createdAt": 1779431564299,
   "modelVersion": "unknown"
 }

.astra/tests/stress/.astra/missions/stress_conflict_1779431564276.json

@@ -1,7 +1,7 @@
 {
-  "missionId": "stress_conflict_1779430407960",
+  "missionId": "stress_conflict_1779431564276",
   "status": "completed",
-  "startTime": "2026-05-22T06:13:27.960Z",
+  "startTime": "2026-05-22T06:32:44.276Z",
   "totalElapsedMs": 24,
   "results": {
     "planner": "Detailed Execution Plan: 1. Research 2. Analyze 3. Write report with high quality.",
@@ -16,30 +16,30 @@
     {
       "from": "idle",
       "to": "planner",
-      "durationMs": 15,
+      "durationMs": 14,
       "message": "전략 수립 중...",
-      "ts": "2026-05-22T06:13:27.975Z"
+      "ts": "2026-05-22T06:32:44.290Z"
     },
     {
       "from": "planner",
       "to": "researcher",
-      "durationMs": 2,
+      "durationMs": 3,
       "message": "핵심 정보 수집 및 분석 중...",
-      "ts": "2026-05-22T06:13:27.977Z"
+      "ts": "2026-05-22T06:32:44.293Z"
     },
     {
       "from": "researcher",
       "to": "writer",
-      "durationMs": 4,
+      "durationMs": 3,
       "message": "최종 리포트 작성 및 편집 중...",
-      "ts": "2026-05-22T06:13:27.981Z"
+      "ts": "2026-05-22T06:32:44.296Z"
     },
     {
       "from": "writer",
       "to": "completed",
-      "durationMs": 3,
+      "durationMs": 4,
       "message": "미션 완료",
-      "ts": "2026-05-22T06:13:27.984Z"
+      "ts": "2026-05-22T06:32:44.300Z"
     }
   ],
   "resilienceMetrics": {

PATCHNOTES.md

@@ -1,5 +1,18 @@
 # Astra Patch Notes
 
+## v2.2.62 (2026-05-22)
+### 🧹 모델 출력 붕괴(degeneration) 방어 — 가독성 보호
+- **문제:** 긴 세션의 분석 답변에서 ①같은 분석이 통째로 두 번 생성됨 ②밑줄 `___` 수백 개 등 문자 벽 ③`(Note: I am acting as ...)` 메타 노트 반복 ④내부 지시문 `Candidate records for this discussion...` 누출 ⑤`[핵심 확인 질문]`·`(질문 의도: …)` 누출. 마크다운/테이블이 깨져 렌더링 불가.
+- **원인:** 소형 로컬 모델이 긴 컨텍스트에서 붕괴. 특히 auto-continuation이 "이어서 쓰라"는 지시를 무시하고 답변을 처음부터 **재생성** → `mergeContinuationParts`가 중복을 못 걸러 통째로 이어 붙임.
+- **수정 1 — 재시작 감지.** continuation 결과가 기존 답변과 같은 도입부로 시작하면(앞부분 12자+ 일치) "이어쓰기"가 아닌 "재시작"으로 판정해 **버림**. 분석이 두 번 나오는 문제 제거.
+- **수정 2 — degeneration 정리 패스.** 최종 출력에서 문자 벽(같은 기호 8개+), `(Note: …)` 메타 노트, `Candidate records…` 내부 지시문, `(질문 의도: …)`·`[핵심 확인 질문]` 누출, 연속 중복 문단을 제거. 모델이 붕괴해도 답변은 읽을 수 있게 유지.
+- **한계:** 한글-한글 토큰 깨짐(`붕괴`→`붕점`, `핵심`→`핵점`)은 멀쩡한 단어처럼 보여 탐지 불가 — 이는 모델 자체 한계. 긴 세션에서는 `/newChat`으로 초기화 권장.
+- **신규 패키징:** `astra-2.2.62.vsix`.
+
+---
+
+
+
 ## v2.2.61 (2026-05-22)
 ### 🔁 출력 없는 액션엔 2번째 LLM 호출 생략 — "빈 응답" 근본 원인 제거
 - **문제:** `git 커밋해줘` 한 번에 LLM이 2번 호출됨 — ①`<run_command>` 액션을 뱉는 호출, ②결과를 정리하는 continuation 호출. 2번째가 빈 응답(`eosFound`)으로 실패.

docs/records/ConnectAI/chronicle.config.json

@@ -7,5 +7,5 @@
   "corePurpose": "",
   "detailLevel": "standard",
   "createdAt": "2026-05-20T09:42:40.003Z",
-  "updatedAt": "2026-05-22T06:26:09.044Z"
+  "updatedAt": "2026-05-22T07:11:47.201Z"
 }

package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "astra",
-  "version": "2.2.61",
+  "version": "2.2.62",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "astra",
-      "version": "2.2.61",
+      "version": "2.2.62",
       "license": "MIT",
       "dependencies": {
         "@lmstudio/sdk": "^1.5.0",

package.json

@@ -2,7 +2,7 @@
   "name": "astra",
   "displayName": "Astra",
   "description": "The personal intelligence layer for Antigravity and VS Code. A private cognitive partner for deep project context, memory, and proactive strategic decision-making.",
-  "version": "2.2.61",
+  "version": "2.2.62",
   "publisher": "g1nation",
   "license": "MIT",
   "icon": "assets/icon.png",

src/agent.ts

@@ -269,7 +269,7 @@ export class AgentExecutor {
     }
 
     private sanitizeAssistantContent(text: string): string {
-        return text
+        const stripped = text
             .replace(/<rationale>[\s\S]*?<\/rationale>/gi, '')
             .replace(/^\s*\[PROBLEM\][\s\S]*?\[GOAL\][\s\S]*?\[REASONING\][^\n]*(?:\n+|$)/i, '')
             .replace(/^\s*\[PROBLEM\][\s\S]*?(?:\n\s*\n|$)/i, '')
@@ -283,6 +283,54 @@ export class AgentExecutor {
             .replace(/<\|?channel\|?>\s*final\b\s*(?:<\|?message\|?>)?/gi, '')
             .replace(/<\|?(?:end|return|start|message)\|?>/gi, '')
             .trim();
+        return this.cleanDegeneratedOutput(stripped);
+    }
+
+    /**
+     * Contain the blast radius when a small local model degenerates on a long /
+     * near-full context: runaway character walls, leaked English meta-notes,
+     * leaked internal directives, stuttered duplicate paragraphs. This does not
+     * fix the model failure — it just keeps a bad answer readable instead of a
+     * wall of noise.
+     */
+    private cleanDegeneratedOutput(text: string): string {
+        let s = text;
+        // Runaway character walls — 8+ of the same separator. A real markdown rule
+        // is exactly 3 (`---`/`***`/`___`); nothing legitimate repeats a char 8+ times.
+        s = s.replace(/([_=~.*])\1{7,}/g, '');
+        // Leaked English meta-commentary the model narrates about itself.
+        s = s.replace(/\(Note:[^)]*\)/gi, '');
+        // Leaked project-chronicle directive (inline form — the `## ` heading form is stripped elsewhere).
+        s = s.replace(/\n?Candidate records for this discussion[^\n]*/gi, '');
+        // Leaked follow-up-question scaffolding the system prompt forbids.
+        s = s.replace(/\(\s*질문\s*의도\s*[:：][^)]*\)/g, '');
+        s = s.replace(/\[\s*핵심\s*확인\s*질문\s*\]\s*/g, '');
+        // Collapse consecutive duplicate paragraphs (model stutter).
+        const paras = s.split(/\n{2,}/);
+        const deduped: string[] = [];
+        for (const p of paras) {
+            const norm = p.trim().replace(/\s+/g, ' ');
+            const prevNorm = deduped.length ? deduped[deduped.length - 1].trim().replace(/\s+/g, ' ') : '';
+            if (norm && norm === prevNorm) { continue; }
+            deduped.push(p);
+        }
+        return deduped.join('\n\n').replace(/[ \t]+\n/g, '\n').replace(/\n{3,}/g, '\n\n').trim();
+    }
+
+    /**
+     * True when a continuation round restarted the whole answer from the top
+     * instead of continuing it — a small model commonly ignores "continue from
+     * here". Detected by a shared leading prefix with the answer so far.
+     */
+    private isRestartedAnswer(soFar: string, cont: string): boolean {
+        const norm = (x: string) => x.trim().replace(/\s+/g, ' ').toLowerCase();
+        const a = norm(soFar);
+        const b = norm(cont);
+        if (a.length < 40 || b.length < 40) { return false; }
+        let i = 0;
+        const max = Math.min(a.length, b.length, 80);
+        while (i < max && a[i] === b[i]) { i++; }
+        return i >= 12;
     }
 
     private async restoreLastSession() {
@@ -1103,6 +1151,13 @@ export class AgentExecutor {
                             logInfo('Continuation produced no visible text — stopping.', { model: actualModel, round: continuationCount });
                             break;
                         }
+                        // A weak model often ignores "continue from here" and re-generates the
+                        // whole answer from the top. Discard such a restart instead of merging
+                        // it — otherwise the user gets the entire analysis twice.
+                        if (this.isRestartedAnswer(cleaned.visible, ccl.visible)) {
+                            logInfo('Continuation restarted the answer instead of continuing — discarding it.', { model: actualModel, round: continuationCount });
+                            break;
+                        }
                         const before = cleaned.visible;
                         cleaned = { ...cleaned, visible: mergeContinuationParts(cleaned.visible, ccl.visible), wasThoughtOnly: false };
                         lastOutputTokens = estimateTokens(ccl.visible);