From 2e32be96fe61446b08f7ad8d0ec4ba525699c10d Mon Sep 17 00:00:00 2001 From: Variet Worker Date: Wed, 15 Apr 2026 13:26:45 +0900 Subject: [PATCH] fix(observer/bridge): v13 _promptOnlySkipped fallback guard + generic button no-context filter (v0.5.46) #task-619 --- .agents/references/known-issues.md | 6 +++++ docs/devlog/2026-04-15.md | 1 + docs/devlog/entries/20260415-002.md | 30 +++++++++++++++++++++++ extension/package.json | 2 +- extension/src/http-bridge.ts | 13 +++++++++- extension/src/observer-script.ts | 38 +++++++++++++++++++++++------ 6 files changed, 80 insertions(+), 10 deletions(-) create mode 100644 docs/devlog/entries/20260415-002.md diff --git a/.agents/references/known-issues.md b/.agents/references/known-issues.md index 5d698d4..137ffdf 100644 --- a/.agents/references/known-issues.md +++ b/.agents/references/known-issues.md @@ -21,6 +21,12 @@ > 鍮꾩듂븳 臾몄젣媛 옱諛쒗븯硫 archive뿉꽌 寃깋븯꽭슂. +### [2026-04-15] [Extension] Observer fallback 컨텍스트가 채팅/UI 텍스트를 명령어로 추출 (v0.5.46) +- **증상**: Discord에 `cmd="실 동작검증을 해봐야하는데"`, `cmd="variet.gravity-bridge"` 등 사용자 채팅/AI 응답 텍스트가 명령어로 전송됨 +- **원인**: v0.5.45에서 `PROMPT_ONLY_RE`가 `code/pre` 요소 스킵 성공했으나, `extractContextFromNearby()`의 fallback(`span/div/p` 수집)이 여전히 작동하여 DOM 트리의 채팅 본문/UI 라벨을 명령어로 추출 +- **해결 (v0.5.46)**: Observer v13에 `_promptOnlySkipped` 플래그 — code 요소가 모두 prompt-only이면 fallback 비활성화. http-bridge에 generic button + no-context일 때 무조건 필터 +- **주의**: 프롬프트 스킵과 fallback 비활성화는 항상 연동해야 함. VSIX 설치 누락 방지를 위해 빌드 후 즉시 `code --install-extension` 확인 필수 + ### [2026-04-15] [Extension] PROMPT_ONLY_RE regex fixes — prompt-only terminal text leaking as enriched cmd (v0.5.45) - **증상**: Discord에 `cmd="…\gravity_control >"` (실제 명령어 없는 빈 터미널 프롬프트)가 전송됨. 명령어가 포함된 경우는 정상 작동 - **원인 1 (Observer)**: `PROMPT_ONLY_RE` 의 `\\\\s`(4중 backslash)가 template literal 안의 regex 리터럴에서 "literal backslash+s"가 되어 whitespace class가 아닌 문자열 매칭 diff --git a/docs/devlog/2026-04-15.md b/docs/devlog/2026-04-15.md index bc50464..3fb5327 100644 --- a/docs/devlog/2026-04-15.md +++ b/docs/devlog/2026-04-15.md @@ -3,3 +3,4 @@ | NNN | HH:MM | 작업 설명 | `커밋해시` | 완료? | |-------|-------|----------|-----------|----------| | 001 | 09:12 | PROMPT_ONLY_RE 근본원인 분석 및 수정 — Observer regex 이스케이핑(4중→2중 backslash) + http-bridge ellipsis prefix 지원, 16개 테스트 전체 통과, VSIX v0.5.45 빌드/배포 | `01539e9` | ✅ | +| 002 | 10:35 | Observer fallback 컨텍스트 추출 수정 — v0.5.45 VSIX 설치 누락 발견/수정 + v13 `_promptOnlySkipped` 플래그로 채팅/UI 텍스트 추출 차단 + bridge generic button 무조건 필터 (v0.5.46) | `b8cda27` | 🔧 | diff --git a/docs/devlog/entries/20260415-002.md b/docs/devlog/entries/20260415-002.md new file mode 100644 index 0000000..195d97e --- /dev/null +++ b/docs/devlog/entries/20260415-002.md @@ -0,0 +1,30 @@ +# Observer fallback 컨텍스트 추출 수정 (v0.5.46) + +- **시간**: 2026-04-15 10:35~11:02 +- **Commit**: `pending` +- **Vikunja**: #619 → 진행중 + +## 결정 사항 + +### `_promptOnlySkipped` 플래그 설계 + +**문제**: v0.5.45에서 PROMPT_ONLY_RE가 code/pre 요소의 프롬프트 텍스트를 정상 스킵했으나, `extractContextFromNearby()`의 **fallback 경로**(span/div/p 텍스트 수집)가 DOM 트리를 올라가면서 채팅 본문, UI 라벨, AI 응답을 명령어로 잘못 추출. + +**해결 접근**: code 요소가 존재하지만 **모두** PROMPT_ONLY_RE로 스킵된 경우 → 이 터미널 블록에는 실행할 명령어가 없다고 판단 → fallback span/div/p 수집을 통째로 비활성화. + +**대안 검토**: +- ❌ fallback 텍스트에 CJK/자연어 필터 추가 → false negative 위험 (한국어 명령어 경로명 등) +- ❌ fallback 수집 depth 제한 → DOM 구조가 바뀌면 다시 깨짐 +- ✅ **prompt-only 스킵과 fallback 비활성화 연동** → 가장 간결하고 확실 + +### VSIX 설치 누락 발견 + +이전 세션(fd78c28e)에서 v0.5.45 VSIX를 빌드했으나 **설치를 하지 않았음**. extensions.json 확인 결과 v0.5.43이 설치되어 있었음. 원인: 이전 세션에서 `code --install-extension` 실행 없이 AG 재시작만 수행. + +→ known-issues에 "빌드 후 즉시 install 확인 필수" 주의사항 추가 + +## 미완료 + +- AG 재시작 후 v0.5.46 실제 동작 검증 필요 +- Discord에 빈 프롬프트/채팅 텍스트가 전송되지 않는지 확인 +- 검증 완료 후 devlog에 커밋 해시 업데이트 + Vikunja #619 완료 처리 diff --git a/extension/package.json b/extension/package.json index 5d57b13..7ed88b4 100644 --- a/extension/package.json +++ b/extension/package.json @@ -2,7 +2,7 @@ "name": "gravity-bridge", "displayName": "Gravity Bridge", "description": "Discord-based unified approval system for Antigravity AI interactions.", - "version": "0.5.45", + "version": "0.5.46", "publisher": "variet", "engines": { "vscode": "^1.100.0" diff --git a/extension/src/http-bridge.ts b/extension/src/http-bridge.ts index a8b32ea..b80eb30 100644 --- a/extension/src/http-bridge.ts +++ b/extension/src/http-bridge.ts @@ -279,9 +279,20 @@ function _handlePending(req: any, res: any, ctx: HttpBridgeContext) { enrichedDesc = `[${rawCmd}] ${rawDesc}`; ctx.logToFile(`[HTTP] command enriched: "${rawCmd}" → "${enrichedCmd.substring(0, 60)}"`); } else { - // Keep original button text as command, but use desc as-is + // v13: Enrichment failed — description is prompt-only or empty + // If cmd is still generic button text, unconditionally filter it ctx.logToFile(`[HTTP] enrichment skipped (prompt-only): "${rawCmd}" desc="${rawDesc.substring(0, 60)}"`); + ctx.logToFile(`[HTTP] filtered generic+prompt-only: "${rawCmd}"`); + res.writeHead(200, { 'Content-Type': 'application/json' }); + res.end(JSON.stringify({ ok: false, filtered: true, reason: 'generic_btn_prompt_only' })); + return; } + } else if (GENERIC_BTN_RE.test(rawCmd) && (rawDesc.length <= 10 || rawDesc === rawCmd)) { + // v13: Generic button with no useful description (observer prompt-only context) + ctx.logToFile(`[HTTP] filtered generic button no-context: "${rawCmd}" desc="${rawDesc.substring(0, 30)}"`); + res.writeHead(200, { 'Content-Type': 'application/json' }); + res.end(JSON.stringify({ ok: false, filtered: true, reason: 'generic_btn_no_context' })); + return; } // ── Server-side false positive filter (uses enriched cmd) ── diff --git a/extension/src/observer-script.ts b/extension/src/observer-script.ts index 6bec43a..2011084 100644 --- a/extension/src/observer-script.ts +++ b/extension/src/observer-script.ts @@ -1,7 +1,7 @@ export function generateApprovalObserverScript(_port: number): string { return ` -// ── Gravity Bridge v12: Prompt-Skip + Multi-CodeEl Context Extraction ── -// v12: Skip prompt-only code text, try all codeEls, improved enrichment fallback +// ── Gravity Bridge v13: Prompt-Skip + Fallback Guard ── +// v13: When all code els are prompt-only, disable fallback text collection entirely (function(){ 'use strict'; var BASE='',_obs=false,_sent={},_ready=false; @@ -10,7 +10,7 @@ export function generateApprovalObserverScript(_port: number): string { var CLEANUP_MS=300000; function log(m){console.log('[GB Observer] '+m);} - log('v12 Script loaded — Prompt-Skip Context Extraction'); + log('v13 Script loaded — Prompt-Skip + Fallback Guard'); // DIAGNOSTIC BEACON: immediate POST to confirm script execution in renderer try { @@ -109,30 +109,36 @@ export function generateApprovalObserverScript(_port: number): string { return null; } - // v12: Climb DOM tree to find context near the button + // v13: Climb DOM tree to find context near the button // Priority: pre/code (skip prompt-only) > substantial span/div/p text > aria-label > button text - // PROMPT_ONLY: matches terminal prompts like "...\project >" or "PS C:\...>" with no actual command - var PROMPT_ONLY_RE = /^[^\\n]*[\\/>\xbb$#]\\s*$/; + // PROMPT_ONLY: matches terminal prompts like "...\project >" or "PS C:\...">" with no actual command + var PROMPT_ONLY_RE = /^[^\\n]*[\\/\>\xbb$#]\\s*$/; function extractContextFromNearby(btn) { var node = btn; var _debugTrail = []; var _fallbackText = ''; // Best span/div/p text found (used if no pre/code) var _bestCodeText = ''; // Best code text found across all depths var _bestCodeHeader = ''; + var _sawCodeEls = false; // v13: did we encounter any code/pre elements? + var _promptOnlySkipped = false; // v13: were ALL code elements prompt-only? for (var depth = 0; depth < 20 && node; depth++) { if (!node.querySelector) { _debugTrail.push('d'+depth+':noQS'); node = node.parentElement; continue; } // Look for code/pre blocks (actual command text) var codeEls = node.querySelectorAll('pre, code, [class*="terminal"]'); _debugTrail.push('d'+depth+':tag='+((node.tagName||'?').toLowerCase())+',cls='+(((typeof node.className==='string')?node.className:'').substring(0,60))+',codeEls='+codeEls.length); + var _depthHadCode = false; + var _depthAllPrompt = true; for (var ci = 0; ci < codeEls.length; ci++) { var codeText = cleanLines((codeEls[ci].textContent || '').trim().substring(0, 500)); if (!codeText || codeText.length <= 5) continue; if (/^Running\\s*\\d/i.test(codeText)) continue; + _depthHadCode = true; // v12: Skip prompt-only text (e.g. "...\gravity_control >") - no actual command if (PROMPT_ONLY_RE.test(codeText.trim())) { _debugTrail.push('skip_prompt_ci='+ci+':'+codeText.substring(0,30)); continue; } + _depthAllPrompt = false; // This code element has actual content - capture it if (!_bestCodeText || codeText.length > _bestCodeText.length) { _bestCodeText = codeText; @@ -148,6 +154,11 @@ export function generateApprovalObserverScript(_port: number): string { } } } + // v13: Track whether code elements were seen and all were prompt-only + if (_depthHadCode) { + _sawCodeEls = true; + if (_depthAllPrompt) _promptOnlySkipped = true; + } // v12: If we found a good code text, return it immediately at this depth if (_bestCodeText) { var parts = []; @@ -157,8 +168,12 @@ export function generateApprovalObserverScript(_port: number): string { _lastContextDebug = _debugTrail.join(' > '); return parts.join(' — '); } - // v11: Look for substantial text in span/div/p as fallback context - if (depth <= 8 && !_fallbackText) { + // v13: If we saw code elements but all were prompt-only, do NOT collect fallback text + // This prevents capturing chat messages / UI text from nearby DOM elements + if (_promptOnlySkipped) { + _debugTrail.push('fallback_blocked_prompt_only'); + } else if (depth <= 8 && !_fallbackText) { + // v11: Look for substantial text in span/div/p as fallback context var textEls = node.querySelectorAll('span, div, p'); for (var ti = 0; ti < Math.min(textEls.length, 20); ti++) { var tText = (textEls[ti].textContent || '').trim(); @@ -175,6 +190,13 @@ export function generateApprovalObserverScript(_port: number): string { } node = node.parentElement; } + // v13: If code elements existed but were all prompt-only, return just button text + // This ensures http-bridge's Run/Always run filter can handle it properly + if (_promptOnlySkipped) { + log('CONTEXT-PROMPT-ONLY trail='+_debugTrail.join(' > ')); + _lastContextDebug = _debugTrail.join(' > '); + return cleanButtonText(btn); + } // v11: Use fallback text from span/div/p if available if (_fallbackText && _fallbackText.length > 5) { log('CONTEXT-OK src=fallback trail='+_debugTrail.join(' > '));