Variet/gravity_control
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(bridge): approval flow tuning — dedup + text cleanup + stall fallback removal + safe reject #task-256

Browse Source
This commit is contained in:
authentik Default Admin
2026-03-09 22:31:44 +09:00
parent 520d36ea43
commit 18b3734c02
5 changed files with 117 additions and 91 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
.agents/references/known-issues.md
View File

@@ -249,3 +249,15 @@
- **해결**: CSP의 `script-src`에 `'unsafe-inline'` 추가. Extension `setupApprovalObserver()`에 CSP 자동 패치 로직 영구 추가
- **주의**: `style-src`에는 `'unsafe-inline'`이 있어 스타일은 동작 → 스크립트만 차단되는 것이 함정. `connect-src`에 `http://127.0.0.1:*` 존재하여 HTTP fetch는 허용됨 (CSP 통과 시 통신은 문제없음). 이 CSP 이슈는 **V8 CachedData 이슈와 동시에 존재**하여 진단이 매우 어려웠음
### [2026-03-09] 중복 승인 요청 — DOM scan + Step probe 동시 발동
- **증상**: Discord에 같은 명령에 대해 승인 요청이 2개 도착
- **원인**: DOM Observer가 `RunAlt+↵` 버튼을 감지하여 pending 생성 + Step probe가 `CORTEX_STEP_STATUS_WAITING` 감지하여 pending 생성 → 동일 step에 대해 2개 파일
- **해결**: `writePendingApproval()`에 15초 dedup 윈도우 추가 — 최근 DOM observer pending이 있으면 step probe pending 스킵
- **주의**: DOM scan은 제거 불가 — `Allow Once`, `Allow This Conversation` 등 step probe가 감지 못하는 UI 버튼 존재. 버튼 텍스트도 `(Alt|Ctrl|Shift|Meta)\+.*` 정규식으로 키보드 단축키 정제
### [2026-03-09] Step probe reject → ResolveOutstandingSteps가 AI 작업 취소
- **증상**: Discord에서 거부 클릭 → AI의 현재 step뿐 아니라 진행 중인 작업 전체가 중단
- **원인**: step probe 경로의 `tryApprovalStrategies(approved=false)` → `ResolveOutstandingSteps` RPC 호출 → 이것은 step을 **CANCEL**하는 파괴적 동작
- **해결**: step probe 경로에서 reject 시 `tryApprovalStrategies` 호출 제거. reject은 로그만 남기고 파기적 동작 없음. DOM observer 경로만 실제 Reject 버튼 클릭 허용
- **주의**: `ResolveOutstandingSteps`는 이름과 달리 "해결"이 아닌 "취소". 승인에 절대 사용 금지 (이전 이슈 #55~59 참조)

3
docs/devlog/2026-03-09.md
View File

@@ -9,4 +9,5 @@
| 005 | 18:30~19:28 | workbench.html inline v3 패치 누락 수정 + pre-patch 검증 | `b61cff1` | ✅ |
| 006 | 19:38~19:56 | V8 CachedData 진단 + 캐시 삭제 (renderer 미실행 근본 원인) | docs only | ✅ |
| 007 | 20:04~20:28 | CSP script-src `'unsafe-inline'` 패치 (renderer 미실행 진짜 근본 원인) | `08077e8` | ✅ |
| 008 | 21:00~21:30 | **E2E 승인 플로우 성공 검증** — AG 재시작 후 renderer v3 실행 확인 + Discord 승인→명령 실행 | pending | ✅ |
| 008 | 21:00~21:30 | **E2E 승인 플로우 성공 검증** — AG 재시작 후 renderer v3 실행 확인 + Discord 승인→명령 실행 | `520d36e` | ✅ |
| 009 | 21:33~22:28 | 승인 플로우 튜닝 — dedup + 텍스트 정제 + stall fallback 제거 + reject 안전화 | pending | 🔧 |

74
extension/out/extension.js
View File

@@ -980,6 +980,9 @@ function generateApprovalObserverScript(_port) {
var txt=(b.textContent||'').trim();
if(!txt)continue;
// Strip keyboard shortcut suffixes (e.g. "RunAlt+↵" → "Run")
txt=txt.replace(/(Alt|Ctrl|Shift|Meta)\+.*/,'').trim();
if(!txt)continue;
// Match against patterns
var matchedType=null;
@@ -1480,28 +1483,8 @@ function setupMonitor() {
logToFile(`[STEP-PROBE] error: ${e.message}`);
}
}
const now = Date.now();
const cooldownOk = (now - lastPendingTime) > 60_000;
const fallbackThreshold = (currentCount > 770) ? 4 : 8; // 775-limit: faster fallback
if (consecutiveIdleCount >= fallbackThreshold && sawRunningAfterPending && cooldownOk) {
// Dynamic fallback: 20s (>770 steps) or 40s (normal)
lastPendingStepIndex = currentCount;
lastPendingTime = now;
sawRunningAfterPending = false;
const command = `Stall at step ${currentCount} (fallback)`;
const description = `승인 대기 감지 — fallback (${consecutiveIdleCount * 5}초 정지), Title: "${currentTitle}"`;
logToFile(`[STALL-FALLBACK] step=${currentCount} frozenCount=${consecutiveIdleCount} → pending`);
writePendingApproval({ conversation_id: activeSessionId, command, description, source: 'stall_fallback' });
}
else if (consecutiveIdleCount === fallbackThreshold) {
const reasons = [];
if (!sawRunningAfterPending)
reasons.push('needDelta>0');
if (!cooldownOk)
reasons.push(`cooldown(${Math.round((60000 - (now - lastPendingTime)) / 1000)}s)`);
if (reasons.length > 0)
logToFile(`[STALL] SKIP: ${reasons.join(', ')}`);
}
// Stall fallback REMOVED — step probe is sole fallback source
// (stall fallback was generating false positives and is now redundant)
}
else if (!isRunning) {
consecutiveIdleCount = 0;
@@ -1614,19 +1597,22 @@ async function processResponseFile(filePath) {
logToFile(`[RESPONSE] renderer-handled approval (rid=${resp.request_id})`);
}
else {
// Step probe / stall path: try all approval strategies
logToFile(`[RESPONSE] step_probe/stall → trying multi-strategy approval`);
const approvalResult = await tryApprovalStrategies(approved, sessionId);
logToFile(`[RESPONSE] approval result: ${approvalResult}`);
// Step probe path: approve → trigger renderer click, reject → log only
if (approved) {
logToFile(`[RESPONSE] step_probe → approve via trigger-click`);
clickTrigger = { action: 'approve', timestamp: Date.now() };
}
else {
// SAFE: reject logs only — NO ResolveOutstandingSteps (it cancels AI work!)
logToFile(`[RESPONSE] step_probe → reject (log only, no destructive action)`);
}
}
logToFile(`[RESPONSE] ${approved ? 'approve' : 'reject'} done (${isDomObserver ? 'dom' : 'step_probe'})`);
// Cleanup response file — BUT NOT for DOM observer!
if (!isDomObserver) {
try {
fs.unlinkSync(filePath);
}
catch { }
// Cleanup response file
try {
fs.unlinkSync(filePath);
}
catch { }
}
catch (e) {
const log = `[RESPONSE] error: ${e.message}`;
@@ -1762,13 +1748,33 @@ function writePendingApproval(data) {
if (!fs.existsSync(pendingDir)) {
fs.mkdirSync(pendingDir, { recursive: true });
}
const id = Date.now().toString();
// ── Dedup: skip if DOM observer already created a pending for same action recently ──
const nowMs = Date.now();
const DEDUP_WINDOW_MS = 15_000; // 15 second dedup window
try {
const existingFiles = fs.readdirSync(pendingDir).filter((f) => f.endsWith('.json'));
for (const ef of existingFiles) {
const efPath = path.join(pendingDir, ef);
const existing = JSON.parse(fs.readFileSync(efPath, 'utf-8'));
if (existing.source === 'dom_observer' && existing.status === 'pending') {
const age = nowMs - (existing.timestamp * 1000);
if (age < DEDUP_WINDOW_MS && age >= 0) {
logToFile(`[DEDUP] skip step_probe pending — DOM observer pending exists: ${ef} (${Math.round(age / 1000)}s ago)`);
return;
}
}
}
}
catch (dedupErr) {
logToFile(`[DEDUP] check error (non-fatal): ${dedupErr.message}`);
}
const id = nowMs.toString();
const payload = {
request_id: id,
conversation_id: data.conversation_id,
command: data.command,
description: data.description,
timestamp: Date.now() / 1000,
timestamp: nowMs / 1000,
status: 'pending',
discord_message_id: 0,
project_name: projectName,

2
extension/out/extension.js.map
View File

File diff suppressed because one or more lines are too long

111
extension/src/extension.ts
View File

@@ -569,29 +569,29 @@ function startObserverHttpBridge(): Promise<number> {
// Listen on deterministic port (derived from projectName), fallback to random
deterministicPort = getDeterministicPort(projectName);
const tryListen = (targetPort: number) => {
server.listen(targetPort, '127.0.0.1', () => {
const port = server.address().port;
observerHttpServer = server;
logToFile(`[HTTP] bridge server started on port ${port}`);
server.listen(targetPort, '127.0.0.1', () => {
const port = server.address().port;
observerHttpServer = server;
logToFile(`[HTTP] bridge server started on port ${port}`);
// Write port to shared ports JSON (multi-bridge support)
const patcher = (sdk.integration as any)?._patcher;
if (patcher && typeof patcher.getWorkbenchDir === 'function') {
const workbenchDir = patcher.getWorkbenchDir();
const portsFile = path.join(workbenchDir, 'ag-bridge-ports.json');
let portsData: Record<string, number> = {};
try {
if (fs.existsSync(portsFile)) {
portsData = JSON.parse(fs.readFileSync(portsFile, 'utf-8'));
}
} catch { }
portsData[projectName] = port;
fs.writeFileSync(portsFile, JSON.stringify(portsData), 'utf-8');
logToFile(`[HTTP] ports JSON updated → ${portsFile} (${projectName}=${port})`);
}
// Write port to shared ports JSON (multi-bridge support)
const patcher = (sdk.integration as any)?._patcher;
if (patcher && typeof patcher.getWorkbenchDir === 'function') {
const workbenchDir = patcher.getWorkbenchDir();
const portsFile = path.join(workbenchDir, 'ag-bridge-ports.json');
let portsData: Record<string, number> = {};
try {
if (fs.existsSync(portsFile)) {
portsData = JSON.parse(fs.readFileSync(portsFile, 'utf-8'));
}
} catch { }
portsData[projectName] = port;
fs.writeFileSync(portsFile, JSON.stringify(portsData), 'utf-8');
logToFile(`[HTTP] ports JSON updated → ${portsFile} (${projectName}=${port})`);
}
resolve(port);
});
resolve(port);
});
};
server.on('error', (e: any) => {
@@ -957,6 +957,9 @@ function generateApprovalObserverScript(_port: number): string {
var txt=(b.textContent||'').trim();
if(!txt)continue;
// Strip keyboard shortcut suffixes (e.g. "RunAlt+↵" → "Run")
txt=txt.replace(/(Alt|Ctrl|Shift|Meta)\+.*/,'').trim();
if(!txt)continue;
// Match against patterns
var matchedType=null;
@@ -1465,27 +1468,8 @@ function setupMonitor() {
}
}
const now = Date.now();
const cooldownOk = (now - lastPendingTime) > 60_000;
const fallbackThreshold = (currentCount > 770) ? 4 : 8; // 775-limit: faster fallback
if (consecutiveIdleCount >= fallbackThreshold && sawRunningAfterPending && cooldownOk) {
// Dynamic fallback: 20s (>770 steps) or 40s (normal)
lastPendingStepIndex = currentCount;
lastPendingTime = now;
sawRunningAfterPending = false;
const command = `Stall at step ${currentCount} (fallback)`;
const description = `승인 대기 감지 — fallback (${consecutiveIdleCount * 5}초 정지), Title: "${currentTitle}"`;
logToFile(`[STALL-FALLBACK] step=${currentCount} frozenCount=${consecutiveIdleCount} → pending`);
writePendingApproval({ conversation_id: activeSessionId, command, description, source: 'stall_fallback' });
} else if (consecutiveIdleCount === fallbackThreshold) {
const reasons = [];
if (!sawRunningAfterPending) reasons.push('needDelta>0');
if (!cooldownOk) reasons.push(`cooldown(${Math.round((60000 - (now - lastPendingTime)) / 1000)}s)`);
if (reasons.length > 0) logToFile(`[STALL] SKIP: ${reasons.join(', ')}`);
}
// Stall fallback REMOVED — step probe is sole fallback source
// (stall fallback was generating false positives and is now redundant)
} else if (!isRunning) {
consecutiveIdleCount = 0;
lastModTime = currentModTime;
@@ -1601,18 +1585,20 @@ async function processResponseFile(filePath: string) {
// DOM observer path: renderer polls /response/:rid and clicks directly
logToFile(`[RESPONSE] renderer-handled approval (rid=${resp.request_id})`);
} else {
// Step probe / stall path: try all approval strategies
logToFile(`[RESPONSE] step_probe/stall → trying multi-strategy approval`);
const approvalResult = await tryApprovalStrategies(approved, sessionId);
logToFile(`[RESPONSE] approval result: ${approvalResult}`);
// Step probe path: approve → trigger renderer click, reject → log only
if (approved) {
logToFile(`[RESPONSE] step_probe → approve via trigger-click`);
clickTrigger = { action: 'approve' as const, timestamp: Date.now() };
} else {
// SAFE: reject logs only — NO ResolveOutstandingSteps (it cancels AI work!)
logToFile(`[RESPONSE] step_probe → reject (log only, no destructive action)`);
}
}
logToFile(`[RESPONSE] ${approved ? 'approve' : 'reject'} done (${isDomObserver ? 'dom' : 'step_probe'})`);
// Cleanup response file — BUT NOT for DOM observer!
if (!isDomObserver) {
try { fs.unlinkSync(filePath); } catch { }
}
// Cleanup response file
try { fs.unlinkSync(filePath); } catch { }
} catch (e: any) {
const log = `[RESPONSE] error: ${e.message}`;
console.log(`Gravity Bridge: ${log}`);
@@ -1738,13 +1724,34 @@ function writePendingApproval(data: { conversation_id: string; command: string;
try {
const pendingDir = path.join(bridgePath, 'pending');
if (!fs.existsSync(pendingDir)) { fs.mkdirSync(pendingDir, { recursive: true }); }
const id = Date.now().toString();
// ── Dedup: skip if DOM observer already created a pending for same action recently ──
const nowMs = Date.now();
const DEDUP_WINDOW_MS = 15_000; // 15 second dedup window
try {
const existingFiles = fs.readdirSync(pendingDir).filter((f: string) => f.endsWith('.json'));
for (const ef of existingFiles) {
const efPath = path.join(pendingDir, ef);
const existing = JSON.parse(fs.readFileSync(efPath, 'utf-8'));
if (existing.source === 'dom_observer' && existing.status === 'pending') {
const age = nowMs - (existing.timestamp * 1000);
if (age < DEDUP_WINDOW_MS && age >= 0) {
logToFile(`[DEDUP] skip step_probe pending — DOM observer pending exists: ${ef} (${Math.round(age/1000)}s ago)`);
return;
}
}
}
} catch (dedupErr: any) {
logToFile(`[DEDUP] check error (non-fatal): ${dedupErr.message}`);
}
const id = nowMs.toString();
const payload = {
request_id: id,
conversation_id: data.conversation_id,
command: data.command,
description: data.description,
timestamp: Date.now() / 1000,
timestamp: nowMs / 1000,
status: 'pending',
discord_message_id: 0,
project_name: projectName,