fix(bridge): resolve websocket zombie connection and bounding memory leaks

.agents/references/known-issues.md

@@ -40,6 +40,12 @@
 
 ## 미해결 이슈
 
+### [2026-03-23] 평생 지속되는 WebSocket 좀비 커넥션 — ping 전송은 성공하나 pong 누락
+- **증상**: 네트워크 환경이 바뀌거나 컴퓨터 절전 모드 복구 시, 서버와의 실연결이 끊어졌음에도 `ws-client.ts`는 이를 인지하지 못하고 연결됨 상태로 표기하며(Zombie) 에이전트 승인 신호가 도착하지 못함.
+- **원인**: Node.js `ws` 라이브러리의 `ws.ping()`은 로컬 OS 버퍼에만 패킷을 찔러넣고 반환을 성공으로 처리함. 서버로부터 돌아오는 `pong`을 기다려 타임아웃시키는 Watchdog이 빠져 있어 영구적인 반쪽짜리 소켓(Half-open state)이 됨.
+- **해결** (v0.5.6): `ping()` 직후 10초 `setTimeout`을 걸고 `on('pong')`에서 해제. 초과 시 로컬 소켓을 강제 `ws.terminate()`하여 `_scheduleReconnect` 재연결 루프를 강제로 트리거함.
+- **주의**: WebSocket 기반 앱을 설계할 땐 반드시 한쪽이 아닌 '양방향' 응답 검증(Watchdog Ping-Pong) 타이머 구조를 가져야 좀비 소켓을 막을 수 있음.
+
 ### [2026-03-11] rejectAgentStep / !stop — AG 미등록 커맨드 + 렌더러 전용 함수 + 스테일 프리미티브
 - **증상**: `!stop` 명령이 AI를 멈추지 못함. 로그: "No active cascade" / "no session tracked yet"
 - **원인**: (1) `antigravity.agent.rejectAgentStep`은 AG 미등록 커맨드. (2) 대체한 `getActiveCascadeId()`는 **렌더러(DOM) 전용 함수** — Extension host에서 항상 `undefined` 반환. (3) **v0.4.5 수정도 실패**: `extension.ts`의 `getActiveSessionId: () => activeSessionId`가 module-level 스트링 프리미티브를 참조 — step-probe가 `ctx.activeSessionId`를 업데이트해도 extension.ts의 변수는 불변 (프리미티브 복사)

bot.py

@@ -202,7 +202,7 @@ class GravityBot(commands.Bot):
         self.conv_to_project: dict[str, str] = {}  # conv_id → project
         self.channel_to_project: dict[int, str] = {}  # channel.id → project
         self.session_status_messages: dict[str, int] = {}  # conv_id → msg_id
-        self._sent_approval_ids: set[str] = set()
+        self._sent_approval_ids: dict[str, bool] = {} # request_id → bool 
         self._deferred_ids: dict[str, int] = {}  # request_id → defer count
         self._sent_commands: dict[str, str] = {}  # request_id → command text (for MERGE edit detection)
         self._ready_event = asyncio.Event()
@@ -255,6 +255,13 @@ class GravityBot(commands.Bot):
             cmd_data["id"] = cmd_data.get("id", str(int(_time.time() * 1000)))
             self.gateway.push_command(project, cmd_data)
 
+    def _cap_dict(self, d: dict, max_size: int = 5000):
+        """Prevent memory leaks by capping dictionary sizes using insertion order (oldest first)."""
+        if len(d) >= max_size:
+            to_remove = len(d) - max_size + max_size // 10  # remove 10%
+            for k in list(d.keys())[:to_remove]:
+                d.pop(k, None)
+
     @staticmethod
     def _make_channel_name(project_name: str) -> str:
         """ag-gravity_control, ag-deriva, etc."""
@@ -650,11 +657,13 @@ class GravityBot(commands.Bot):
                     reject_commands = {"deny", "reject", "cancel", "decline", "dismiss", "stop"}
                     if req.command.strip().lower() in reject_commands:
                         logger.warning(f"Auto-approve BLOCKED: command='{req.command}' is reject-word — skipping")
-                        self._sent_approval_ids.add(req.request_id)
+                        self._cap_dict(self._sent_approval_ids)
+                        self._sent_approval_ids[req.request_id] = True
                         phase1_processed += 1
                         continue
 
-                    self._sent_approval_ids.add(req.request_id)
+                    self._cap_dict(self._sent_approval_ids)
+                    self._sent_approval_ids[req.request_id] = True
 
                     # Smart button_index: read buttons array from pending file
                     # file_permission buttons = [Allow Once(0), Allow This Conv(1), Deny(2)]
@@ -724,7 +733,9 @@ class GravityBot(commands.Bot):
 
                 channel = await self._get_channel(project)
                 if channel:
-                    self._sent_approval_ids.add(req.request_id)
+                    self._cap_dict(self._sent_approval_ids)
+                    self._sent_approval_ids[req.request_id] = True
+                    self._cap_dict(self._sent_commands)
                     self._sent_commands[req.request_id] = req.command
                     await self._send_approval_request(channel, req)
                     phase1_processed += 1
@@ -767,7 +778,7 @@ class GravityBot(commands.Bot):
                         self._deferred_ids.pop(rid, None)
                         self._sent_commands.pop(rid, None)
                         self._approval_messages.pop(rid, None)
-                        self._sent_approval_ids.discard(rid)
+                        self._sent_approval_ids.pop(rid, None)
                         phase2_processed += 1
 
                     elif status == "expired":
@@ -790,7 +801,7 @@ class GravityBot(commands.Bot):
                         f.unlink()
                         self._deferred_ids.pop(rid, None)
                         self._sent_commands.pop(rid, None)
-                        self._sent_approval_ids.discard(rid)
+                        self._sent_approval_ids.pop(rid, None)
                         phase2_processed += 1
 
                     elif status == "pending":
@@ -885,6 +896,7 @@ class GravityBot(commands.Bot):
                 pass
 
         logger.info(f"Sent approval request: {request.request_id[:12]}")
+        self._cap_dict(self._approval_messages)
         self._approval_messages[request.request_id] = msg.id  # FIX #4: Track msg_id for auto_resolved lookup
 
     # ─── Discord → IDE Text Relay + Multi-PC UX ───────────────────────────
@@ -1073,7 +1085,10 @@ class GravityBot(commands.Bot):
             view = ApprovalView(self.bridge, request, buttons=buttons, hub=self.hub)
             msg = await channel.send(embed=embed, view=view)
 
-            self._sent_approval_ids.add(request_id)
+            self._cap_dict(self._sent_approval_ids)
+            self._sent_approval_ids[request_id] = True
+            
+            self._cap_dict(self._approval_messages)
             self._approval_messages[request_id] = msg.id
             logger.info(f"[HUB-PENDING] Sent approval: {request_id[:12]} project={project}")
 
@@ -1082,7 +1097,8 @@ class GravityBot(commands.Bot):
 
     async def _auto_approve_via_hub(self, request: ApprovalRequest):
         """Auto-approve a pending request via Hub."""
-        self._sent_approval_ids.add(request.request_id)
+        self._cap_dict(self._sent_approval_ids)
+        self._sent_approval_ids[request.request_id] = True
 
         delivered = False
         if self.hub:

docs/devlog/2026-03-23.md

@@ -0,0 +1,5 @@
+# 2026-03-23 Devlog
+
+| NNN | HH:MM | 작업 설명 | `커밋해시` | ✅ 또는 🔧 |
+|-----|-------|----------|-----------|-----------|
+| 001 | 21:09 | WebSocket 좀비 커넥션 해결 및 통신망 메모리 누수 패치 | `TBD` | ✅ |

docs/devlog/entries/20260323-001.md

@@ -0,0 +1,12 @@
+# WebSocket 좀비 커넥션 해결 및 통신망 메모리 누수 구조 패치
+
+- **시간**: 2026-03-23 21:09~21:20
+- **Commit**: TBD
+- **Vikunja**: #510 → done
+
+## 결정 사항
+- **ws-client.ts 핑퐁 와치독(Ping-Pong Watchdog)**: 단순 에러 캐치가 아니라 `ws.terminate()`를 통해 무반응 소켓을 강제 종료하여 자체 재연결 로직(`_onDisconnect`)을 활성화하도록 설계.
+- **통신망 추적 변수 캡핑(Bounded Cap)**: `hub.py`의 `pending_owners` 및 `bot.py`의 `_sent_approval_ids` 등 무한히 쌓일 수 있는 파이썬 딕셔너리에 LRU(오래된 순 삭제) 로직을 추가. 비록 당장 OOM을 유발하진 않지만 이 구조적 메모리 누수(Leak)를 원천적으로 차단하여 시스템 안정성을 극대화함.
+
+## 미완료
+- 없음

extension/out/extension.js

@@ -44,6 +44,7 @@ var __importStar = (this && this.__importStar) || (function () {
     };
 })();
 Object.defineProperty(exports, "__esModule", { value: true });
+exports.fixLSConnection = fixLSConnection;
 exports.activate = activate;
 exports.deactivate = deactivate;
 const vscode = __importStar(require("vscode"));
@@ -251,18 +252,24 @@ async function initSDK(context) {
  * found (wrong workspace).
  */
 async function fixLSConnection() {
-    if (!sdk?.ls)
-        return;
+    if (!sdk?.ls) {
+        logToFile('[LS-FIX] skipped: sdk.ls not available');
+        return false;
+    }
     try {
         const folders = vscode.workspace.workspaceFolders;
-        if (!folders || folders.length === 0)
-            return;
+        if (!folders || folders.length === 0) {
+            logToFile('[LS-FIX] skipped: no workspace folders');
+            return false;
+        }
         // Generate the workspace hint the same way SDK does, but we'll match case-insensitively
         const folder = folders[0].uri.fsPath;
         const parts = folder.replace(/\\/g, '/').split('/');
         const hint = parts.slice(-2).join('_').replace(/[-.\s]/g, '_').toLowerCase();
-        if (!hint)
-            return;
+        if (!hint) {
+            logToFile('[LS-FIX] skipped: empty hint');
+            return false;
+        }
         // Find all language_server processes with csrf_token
         const { exec } = cp;
         const { promisify } = require('util');
@@ -274,12 +281,18 @@ async function fixLSConnection() {
             const result = await execAsync(`powershell.exe -NoProfile -EncodedCommand ${encoded}`, { encoding: 'utf8', timeout: 15000, windowsHide: true });
             output = result.stdout;
         }
-        catch {
-            return; // Can't discover processes — leave SDK's choice
+        catch (psErr) {
+            logToFile(`[LS-FIX] skipped: PowerShell failed — ${psErr.message?.substring(0, 100)}`);
+            return false;
         }
         const lines = output.split('\n').filter((l) => l.trim().length > 0);
-        if (lines.length <= 1)
-            return; // Only one LS — no ambiguity
+        if (lines.length === 0) {
+            logToFile('[LS-FIX] skipped: no LS processes found');
+            return false;
+        }
+        // NOTE: Do NOT skip on single LS — SDK may have fallen back to wrong LS
+        // due to case-sensitive hint mismatch, even when only one process exists.
+        logToFile(`[LS-FIX] found ${lines.length} LS process(es), hint="${hint}"`);
         // Find the line whose workspace_id matches our workspace (case-insensitive)
         let matchedLine = null;
         for (const line of lines) {
@@ -296,7 +309,7 @@ async function fixLSConnection() {
         }
         if (!matchedLine) {
             logToFile(`[LS-FIX] No LS process matched hint="${hint}" (${lines.length} processes)`);
-            return;
+            return false;
         }
         // Extract port and csrf_token from matched line
         const csrfMatch = matchedLine.match(/--csrf_token[= ](\S+)/);
@@ -304,7 +317,7 @@ async function fixLSConnection() {
         const pidMatch = matchedLine.split('|')[0]?.trim();
         if (!csrfMatch || !extPortMatch) {
             logToFile(`[LS-FIX] Matched LS but missing csrf/port args`);
-            return;
+            return false;
         }
         const csrfToken = csrfMatch[1];
         const extPort = parseInt(extPortMatch[1], 10);
@@ -312,7 +325,7 @@ async function fixLSConnection() {
         // Check if SDK already connected to this LS
         if (sdk.ls.port === extPort) {
             logToFile(`[LS-FIX] SDK already on correct LS port=${extPort}`);
-            return;
+            return false;
         }
         // Find ConnectRPC port via netstat (same as SDK logic)
         let netstatOutput;
@@ -325,7 +338,7 @@ async function fixLSConnection() {
             logToFile(`[LS-FIX] netstat failed, using ext_port=${extPort} for PID=${pid}`);
             sdk.ls.setConnection(extPort, csrfToken, false);
             logToFile(`[LS-FIX] ✅ Reconnected to correct LS: port=${extPort} hint="${hint}" PID=${pid}`);
-            return;
+            return true;
         }
         const portMatches = netstatOutput.matchAll(/127\.0\.0\.1:(\d+)/g);
         const ports = [];
@@ -358,7 +371,7 @@ async function fixLSConnection() {
                     if (ok) {
                         sdk.ls.setConnection(port, csrfToken, useTls);
                         logToFile(`[LS-FIX] ✅ Reconnected to correct LS: port=${port} ${proto} hint="${hint}" PID=${pid}`);
-                        return;
+                        return true;
                     }
                 }
                 catch { /* try next */ }
@@ -367,9 +380,11 @@ async function fixLSConnection() {
         // Last resort: use extension_server_port
         sdk.ls.setConnection(extPort, csrfToken, false);
         logToFile(`[LS-FIX] ✅ Reconnected via ext_port=${extPort} hint="${hint}" PID=${pid}`);
+        return true;
     }
     catch (err) {
         logToFile(`[LS-FIX] error: ${err.message}`);
+        return false;
     }
 }
 // ─── Approval Observer + Product.json Checksums extracted to ./html-patcher.ts ───
@@ -515,6 +530,7 @@ async function activate(context) {
             recentDiscordSentTexts,
             writeChatSnapshot,
             writeChatSnapshotWithFiles,
+            fixLSConnection,
         });
         // Start HTTP bridge with live step-probe state (prevents stale primitive bug)
         const httpBridgeCtx = {

extension/package.json

@@ -2,7 +2,7 @@
     "name": "gravity-bridge",
     "displayName": "Gravity Bridge",
     "description": "Antigravity ↔ Discord 브리지 연동 확장",
-    "version": "0.5.5",
+    "version": "0.5.6",
     "publisher": "variet",
     "engines": {
         "vscode": "^1.100.0"

extension/src/step-probe.ts

@@ -92,6 +92,7 @@ export function resetPendingState(): void {
  * re-detection of WAITING steps whose pending was lost during disconnect.
  */
 export function resetPendingStateForReconnect(): void {
+    if (!ctx) return; // Prevent startup race conditions
     ctx.lastPendingStepIndex = -1;
     ctx.stallProbed = false;
     ctx.sawRunningAfterPending = false;

extension/src/ws-client.ts

@@ -122,6 +122,7 @@ export class WSBridgeClient {
     private reconnectDelay = INITIAL_RECONNECT_DELAY;
     private reconnectTimer: NodeJS.Timeout | null = null;
     private heartbeatTimer: NodeJS.Timeout | null = null;
+    private pongTimeoutTimer: NodeJS.Timeout | null = null;
     private authTimer: NodeJS.Timeout | null = null;
 
     // Message queue (survives reconnection)
@@ -245,6 +246,10 @@ export class WSBridgeClient {
 
                 ws.on('pong', () => {
                     // Server responded to our ping — connection is alive
+                    if (this.pongTimeoutTimer) {
+                        clearTimeout(this.pongTimeoutTimer);
+                        this.pongTimeoutTimer = null;
+                    }
                 });
             } else {
                 // ─── Browser-style WebSocket API (.onopen / .onmessage) ───
@@ -469,6 +474,16 @@ export class WSBridgeClient {
                     // Node.js ws has .ping(), browser WebSocket doesn't
                     if (typeof this.ws.ping === 'function') {
                         this.ws.ping();
+
+                        // Set timeout waiting for pong
+                        if (this.pongTimeoutTimer) clearTimeout(this.pongTimeoutTimer);
+                        this.pongTimeoutTimer = setTimeout(() => {
+                            this.logFn('[WS] Heartbeat timeout — no pong received, terminating connection');
+                            if (this.ws) {
+                                try { this.ws.terminate(); } catch { try { this.ws.close(); } catch { } }
+                            }
+                            this._onDisconnect();
+                        }, 10000); // 10s timeout
                     } else {
                         // Fallback: send heartbeat as JSON message
                         this.ws.send(JSON.stringify({ type: 'heartbeat' }));
@@ -485,6 +500,10 @@ export class WSBridgeClient {
             clearInterval(this.heartbeatTimer);
             this.heartbeatTimer = null;
         }
+        if (this.pongTimeoutTimer) {
+            clearTimeout(this.pongTimeoutTimer);
+            this.pongTimeoutTimer = null;
+        }
     }
 
     // ─── Reconnection ───
@@ -540,6 +559,11 @@ export class WSBridgeClient {
             this.reconnectTimer = null;
         }
 
+        if (this.pongTimeoutTimer) {
+            clearTimeout(this.pongTimeoutTimer);
+            this.pongTimeoutTimer = null;
+        }
+
         if (this.ws) {
             try {
                 this.ws.close();

hub.py

@@ -546,6 +546,11 @@ class WSHub:
             payload = data.get("data", {})
             request_id = payload.get("request_id", "")
             if request_id:
+                # Prevent slow memory leak for stranded requests
+                if len(self.pending_owners) > 10000:
+                    oldest = next(iter(self.pending_owners))
+                    self.pending_owners.pop(oldest, None)
+                
                 # Track ownership for response routing
                 self.pending_owners[request_id] = conn.conn_id
                 # Add source metadata