feat(bridge): step-type-specific approval commands + SDK research

- tryApprovalStrategies: terminalCommand.run > terminalCommand.accept > command.accept > acceptAgentStep
- Step probe: immediate on first stall (5s), 775-limit detection with dynamic fallback
- NOTIFY filter: skip <50 chars, TASK dedup by taskName+taskStatus
- BTN-DUMP diagnostic removed from renderer
- Focus: agentPanel.focus + agentSidePanel.focus (verified SDK commands)
- known-issues: add step-type command mismatch finding

.agents/references/known-issues.md

@@ -164,3 +164,8 @@
 - **해결**: 항상 `extension.ts`의 `generateApprovalObserverScript()` 함수를 수정 → 컴파일 → 배포 → Reload
 - **주의**: HTML inline은 JS파일이 먼저 로드되어 `window.__agSDK` 가드에 의해 실행 안 됨. 실제 실행되는 것은 JS파일 경로의 스크립트
 
+### [2026-03-09] VS Code Accept — Run 버튼에 잘못된 명령 사용
+- **증상**: Discord 승인 → `acceptAgentStep` 실행 → "Silent Success" (실제 승인 안 됨)
+- **원인**: `acceptAgentStep`은 **코드 변경** 승인 전용. Run 버튼 = **터미널 명령** 승인으로 `terminalCommand.run` 또는 `terminalCommand.accept`가 올바른 명령
+- **해결**: SDK 7개 승인 명령을 step type별로 분기 시도 (`terminalCommand.run` → `terminalCommand.accept` → `command.accept` → `acceptAgentStep`)
+- **주의**: `terminalCommand.run`의 개별 동작 결과는 아직 미검증. devlog-004에서 순차 시도만 언급됨. AG 재시작 후 E2E 테스트 필요

docs/devlog/2026-03-09.md

@@ -0,0 +1,5 @@
+# 2026-03-09 Devlog
+
+| # | 시간 | 작업 설명 | 커밋 | 상태 |
+|---|------|----------|------|------|
+| 001 | 08:00~09:17 | 승인 실행 메커니즘 연구 + step-type별 VS Code 명령 분기 구현 | `pending` | 🔧 |

docs/devlog/entries/20260309-001.md

@@ -0,0 +1,34 @@
+# 승인 실행 메커니즘 연구 + step-type별 명령 분기
+
+- **시간**: 2026-03-09 08:00~09:17
+- **Commit**: `pending`
+- **Vikunja**: #263 진행중
+
+## 결정 사항
+
+### acceptAgentStep은 Run 버튼에 무효
+- SDK에 7개 별도 승인 명령 존재 (우리는 1개만 사용 중이었음)
+- Run 버튼 = **터미널 명령** → `terminalCommand.run` / `terminalCommand.accept`가 올바른 명령
+- `acceptAgentStep`은 **코드 변경** 승인 전용 → Run 버튼과 무관
+
+### 빌트인 Turbo Mode 발견
+- `TerminalExecutionPolicy.EAGER` = 항상 자동 실행 (AG Settings에서 on/off)
+- 사용자 판단: 통제권 유지가 중요 → Discord 승인 방식 유지, Turbo는 최후 수단
+
+### pywinauto 폐기
+- 크로스플랫폼 불가 (Linux 사용 가능성)
+- 창 겹침 시 동작 불가
+- 사용자 결정으로 폐기
+
+## 코드 변경
+- `extension.ts`: 
+  - `tryApprovalStrategies()` — HandleCascadeUserInteraction RPC 3가지 variant + VS Code 7개 명령 순차 시도
+  - Step probe: stall 1 poll (5초) 후 즉시 probe, 775-limit 감지 + 동적 fallback (20s/40s)
+  - NOTIFY 필터: 50자 미만 무시, TASK 중복 skip  
+  - BTN-DUMP 진단 로그 제거
+- `.agents/references/known-issues.md`: 1건 추가
+
+## 미완료
+1. **AG 재시작 후 E2E 테스트** — `terminalCommand.run` 개별 동작 확인 (핵심)
+2. **`extension.log` 분석** — 어떤 전략이 실제 승인을 트리거하는지 로그로 확인
+3. **실패 시 대안**: `executeCascadeAction` 파라미터 탐색 또는 Turbo Mode 활성화

extension/out/extension.js

@@ -689,23 +689,6 @@ function generateApprovalObserverScript(_port) {
     if(document.body)searchRoots.push(document.body);
     if(!searchRoots.length)return;
 
-    // ── DIAGNOSTIC: dump ALL button-like elements EVERY scan cycle ──
-    {
-      var dumpBtns=[];
-      var totalChecked=0;
-      for(var dr=0;dr<searchRoots.length;dr++){
-        var dbs=searchRoots[dr].querySelectorAll('button,[role="button"]');
-        totalChecked+=dbs.length;
-        for(var di=0;di<dbs.length;di++){
-          var db=dbs[di];
-          var dt=(db.textContent||'').trim();
-          var dtShort=dt.replace(/\\s+/g,' ').substring(0,50);
-          dumpBtns.push(db.tagName+'['+dt.length+']:'+dtShort);
-        }
-      }
-      log('[BTN-DUMP] roots='+searchRoots.length+' total='+totalChecked+' btns='+dumpBtns.length+': '+JSON.stringify(dumpBtns.slice(0,15)));
-    }
-
     var seen={}; // dedupe buttons across search roots
     for(var r=0;r<searchRoots.length;r++){
     var allBtns=searchRoots[r].querySelectorAll('button');
@@ -936,6 +919,7 @@ function setupMonitor() {
     let sawRunningAfterPending = true; // gate: must see delta>0 before next pending
     let lastModTime = ''; // track lastModifiedTime to distinguish thinking vs approval
     let stallProbed = false; // prevent repeated step probes during same stall
+    let lastRelayedTaskText = ''; // dedup TASK_BOUNDARY relay
     setInterval(async () => {
         pollCount++;
         if (pollCount <= 3 || pollCount % 12 === 0) {
@@ -1012,9 +996,8 @@ function setupMonitor() {
                 logToFile(`[POLL#${pollCount}] status=${statusStr} steps=${currentCount} delta=${delta}`);
             }
             // ── PRIMARY: Step-probe-based approval detection ──
-            // latestToolCallStep does NOT exist in GetAllCascadeTrajectories response.
+            // On stall (idle=1, ~5s), probe GetCascadeTrajectorySteps to check WAITING.
-            // Instead, on early stall (idle=2, ~10s), probe GetCascadeTrajectorySteps
+            // 775-step limit: probe fails for long sessions → faster stall fallback.
-            // to fetch the latest step and check if it's a tool call awaiting approval.
             // ── STALL-BASED approval detection with step probe ──
             const currentModTime = bestSession.lastModifiedTime || bestSession.lastModifiedTimestamp || bestSession.modifiedTime || '';
             const modTimeChanged = currentModTime !== lastModTime;
@@ -1046,7 +1029,7 @@ function setupMonitor() {
                 // ── Step probe: on stall, fetch latest step via cascadeId (retry until WAITING found) ──
                 // CONFIRMED: param='cascadeId', id=sessionId (map key from trajectorySummaries)
                 // Retries every 2 polls (~10s) because RUN_COMMAND step may not be created yet
-                if (consecutiveIdleCount >= 1 && consecutiveIdleCount % 2 === 1 && !stallProbed) {
+                if (consecutiveIdleCount >= 1 && !stallProbed) {
                     try {
                         const stepsResp = await sdk.ls.rawRPC('GetCascadeTrajectorySteps', {
                             cascadeId: bestSessionId,
@@ -1055,6 +1038,9 @@ function setupMonitor() {
                             const steps = stepsResp.steps;
                             // Diagnostic: compare returned steps vs trajectory stepCount
                             logToFile(`[STEP-PROBE] returned=${steps.length} vs trajectory.stepCount=${currentCount}`);
+                            if (steps.length < currentCount) {
+                                logToFile(`[STEP-PROBE] ⚠️ 775-limit hit! steps=${steps.length} < stepCount=${currentCount}`);
+                            }
                             // Scan last 5 steps backwards to find WAITING (RUN_COMMAND may not be last)
                             let foundWaiting = false;
                             for (let si = steps.length - 1; si >= Math.max(0, steps.length - 5); si--) {
@@ -1116,8 +1102,9 @@ function setupMonitor() {
                 }
                 const now = Date.now();
                 const cooldownOk = (now - lastPendingTime) > 60_000;
-                if (consecutiveIdleCount >= 8 && sawRunningAfterPending && cooldownOk) {
+                const fallbackThreshold = (currentCount > 770) ? 4 : 8; // 775-limit: faster fallback
-                    // 8 polls × 5s = 40 seconds — fallback (reduced from 100s)
+                if (consecutiveIdleCount >= fallbackThreshold && sawRunningAfterPending && cooldownOk) {
+                    // Dynamic fallback: 20s (>770 steps) or 40s (normal)
                     lastPendingStepIndex = currentCount;
                     lastPendingTime = now;
                     sawRunningAfterPending = false;
@@ -1126,7 +1113,7 @@ function setupMonitor() {
                     logToFile(`[STALL-FALLBACK] step=${currentCount} frozenCount=${consecutiveIdleCount} → pending`);
                     writePendingApproval({ conversation_id: activeSessionId, command, description, source: 'stall_fallback' });
                 }
-                else if (consecutiveIdleCount === 8) {
+                else if (consecutiveIdleCount === fallbackThreshold) {
                     const reasons = [];
                     if (!sawRunningAfterPending)
                         reasons.push('needDelta>0');
@@ -1145,10 +1132,14 @@ function setupMonitor() {
             if (notifyStep && notifyStep.stepIndex > lastNotifyStepIndex) {
                 lastNotifyStepIndex = notifyStep.stepIndex;
                 const content = notifyStep.step?.notifyUser?.notificationContent || '';
-                if (content.length > 10) {
+                // Filter: only relay meaningful notifications (skip trivial ones)
+                if (content.length > 50) {
                     writeChatSnapshot(`📣 **알림** (step ${notifyStep.stepIndex})\n\n${content}`);
                     console.log(`Gravity Bridge: [POLL#${pollCount}] NOTIFY step=${notifyStep.stepIndex} ${content.length} chars`);
                 }
+                else if (content.length > 0) {
+                    logToFile(`[POLL] NOTIFY skipped (too short: ${content.length} chars): ${content.substring(0, 80)}`);
+                }
             }
             // ── Process latestTaskBoundaryStep ──
             const taskStep = bestSession.latestTaskBoundaryStep;
@@ -1157,8 +1148,16 @@ function setupMonitor() {
                 const tb = taskStep.step?.taskBoundary;
                 if (tb?.taskName) {
                     const mode = tb.mode ? tb.mode.replace('AGENT_MODE_', '') : '';
-                    writeChatSnapshot(`📋 **[${mode}] ${tb.taskName}**\n${tb.taskStatus || ''}\n\n${tb.taskSummary || ''}`);
+                    // Filter: skip status-only updates with same task name (noise)
-                    console.log(`Gravity Bridge: [POLL#${pollCount}] TASK step=${taskStep.stepIndex} "${tb.taskName}"`);
+                    const taskText = `${tb.taskName}|${tb.taskStatus || ''}`;
+                    if (taskText !== lastRelayedTaskText) {
+                        lastRelayedTaskText = taskText;
+                        writeChatSnapshot(`📋 **[${mode}] ${tb.taskName}**\n${tb.taskStatus || ''}\n\n${tb.taskSummary || ''}`);
+                        console.log(`Gravity Bridge: [POLL#${pollCount}] TASK step=${taskStep.stepIndex} "${tb.taskName}"`);
+                    }
+                    else {
+                        logToFile(`[POLL] TASK skipped (duplicate): "${tb.taskName}"`);
+                    }
                 }
             }
         }
@@ -1225,52 +1224,23 @@ async function processResponseFile(filePath) {
             }
             catch { }
         }
-        // ═══ APPROVAL STRATEGY ═══
+        // ═══ MULTI-STRATEGY APPROVAL (v2.1) ═══
-        // DOM observer approvals: renderer handles clicking via pollResponse — skip VS Code commands
+        // Tries multiple methods sequentially with detailed logging.
-        // Stall-detection approvals: use VS Code commands as fallback (focus-dependent)
+        // DOM observer: renderer handles clicking via pollResponse
+        // Step probe/stall: try RPC → VS Code commands → log results
         const approved = resp.approved;
         if (isDomObserver) {
             // DOM observer path: renderer polls /response/:rid and clicks directly
             logToFile(`[RESPONSE] renderer-handled approval (rid=${resp.request_id})`);
         }
         else {
-            // Step probe / stall path: relay approval to DOM observer pending files
+            // Step probe / stall path: try all approval strategies
-            // The renderer polls /response/:rid and can click the actual button
+            logToFile(`[RESPONSE] step_probe/stall → trying multi-strategy approval`);
-            logToFile(`[RESPONSE] step_probe/stall → relaying to DOM observer pending files`);
+            const approvalResult = await tryApprovalStrategies(approved, sessionId);
-            const pendingDir = path.join(bridgePath, 'pending');
+            logToFile(`[RESPONSE] approval result: ${approvalResult}`);
-            const responseDir = path.join(bridgePath, 'response');
-            if (!fs.existsSync(responseDir))
-                fs.mkdirSync(responseDir, { recursive: true });
-            let relayCount = 0;
-            try {
-                const files = fs.readdirSync(pendingDir).filter((f) => f.endsWith('.json'));
-                for (const f of files) {
-                    try {
-                        const pd = JSON.parse(fs.readFileSync(path.join(pendingDir, f), 'utf-8'));
-                        if (pd.source === 'dom_observer' && pd.status === 'pending') {
-                            // Write response file for this DOM observer pending
-                            const responsePayload = {
-                                request_id: pd.request_id,
-                                approved,
-                                timestamp: Date.now() / 1000,
-                            };
-                            fs.writeFileSync(path.join(responseDir, `${pd.request_id}.json`), JSON.stringify(responsePayload, null, 2), 'utf-8');
-                            relayCount++;
-                            logToFile(`[RESPONSE] relayed to DOM pending: ${pd.request_id} approved=${approved}`);
-                        }
-                    }
-                    catch { }
-                }
-            }
-            catch (e) {
-                logToFile(`[RESPONSE] relay scan error: ${e.message}`);
-            }
-            logToFile(`[RESPONSE] relayed to ${relayCount} DOM observer pending files`);
         }
-        logToFile(`[RESPONSE] ${approved ? 'approve' : 'reject'} done (${isDomObserver ? 'dom' : 'stall'})`);
+        logToFile(`[RESPONSE] ${approved ? 'approve' : 'reject'} done (${isDomObserver ? 'dom' : 'step_probe'})`);
         // Cleanup response file — BUT NOT for DOM observer!
-        // DOM observer: renderer's pollResponse needs to find it via HTTP GET /response/:rid
-        // Non-DOM (stall/step_probe relay): watcher already handled it, safe to delete
         if (!isDomObserver) {
             try {
                 fs.unlinkSync(filePath);
@@ -1437,6 +1407,129 @@ function writePendingApproval(data) {
         console.log(`Gravity Bridge: pending write error: ${e.message}`);
     }
 }
+// ─── Multi-Strategy Approval Execution ───
+/**
+ * Try multiple approval methods sequentially.
+ * Returns a string describing which method succeeded (or all failed).
+ *
+ * Strategy order (most reliable first):
+ * 1. HandleCascadeUserInteraction RPC (cross-platform, no focus)
+ * 2. VS Code accept/reject commands (focus-dependent)
+ * 3. Log failure for manual intervention
+ */
+async function tryApprovalStrategies(approved, sessionId) {
+    const action = approved ? 'APPROVE' : 'REJECT';
+    logToFile(`[APPROVAL] Starting ${action} strategies for session ${sessionId.substring(0, 8)}`);
+    // ── Strategy 1: HandleCascadeUserInteraction RPC ──
+    if (sdk) {
+        // Try variant A: { cascadeId, approved }
+        try {
+            logToFile(`[APPROVAL-1A] HandleCascadeUserInteraction { cascadeId, approved: ${approved} }`);
+            const rpcResult = await sdk.ls.rawRPC('HandleCascadeUserInteraction', {
+                cascadeId: sessionId,
+                approved: approved,
+            });
+            logToFile(`[APPROVAL-1A] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
+            return `RPC-1A:HandleCascadeUserInteraction(approved=${approved})`;
+        }
+        catch (e) {
+            logToFile(`[APPROVAL-1A] ❌ FAIL: ${e.message}`);
+        }
+        // Try variant B: { cascadeId, stepAction }
+        try {
+            const stepAction = approved ? 'STEP_ACTION_ACCEPT' : 'STEP_ACTION_REJECT';
+            logToFile(`[APPROVAL-1B] HandleCascadeUserInteraction { cascadeId, stepAction: '${stepAction}' }`);
+            const rpcResult = await sdk.ls.rawRPC('HandleCascadeUserInteraction', {
+                cascadeId: sessionId,
+                stepAction: stepAction,
+            });
+            logToFile(`[APPROVAL-1B] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
+            return `RPC-1B:HandleCascadeUserInteraction(stepAction=${stepAction})`;
+        }
+        catch (e) {
+            logToFile(`[APPROVAL-1B] ❌ FAIL: ${e.message}`);
+        }
+        // Try variant C: { cascadeId, userAction } (experimental)
+        try {
+            const userAction = approved ? 'USER_ACTION_APPROVED' : 'USER_ACTION_REJECTED';
+            logToFile(`[APPROVAL-1C] HandleCascadeUserInteraction { cascadeId, userAction: '${userAction}' }`);
+            const rpcResult = await sdk.ls.rawRPC('HandleCascadeUserInteraction', {
+                cascadeId: sessionId,
+                userAction: userAction,
+            });
+            logToFile(`[APPROVAL-1C] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
+            return `RPC-1C:HandleCascadeUserInteraction(userAction=${userAction})`;
+        }
+        catch (e) {
+            logToFile(`[APPROVAL-1C] ❌ FAIL: ${e.message}`);
+        }
+    }
+    // ── Strategy 2: VS Code Commands — step-type-specific (focus-dependent) ──
+    // Per SDK research (2026-03-09):
+    //   Run button = terminal command → terminalCommand.run / terminalCommand.accept
+    //   Code changes = agent step → agent.acceptAgentStep
+    //   General commands = command.accept
+    // Previously only tried acceptAgentStep (Silent Success) — now tries ALL 7
+    // Try to focus the panel first (required for command.accept / acceptAgentStep)
+    try {
+        logToFile(`[APPROVAL-2] focusing panel...`);
+        await vscode.commands.executeCommand('antigravity.agentPanel.focus');
+        await new Promise(resolve => setTimeout(resolve, 300));
+        await vscode.commands.executeCommand('antigravity.agentSidePanel.focus');
+        await new Promise(resolve => setTimeout(resolve, 300));
+        logToFile(`[APPROVAL-2] panel focus attempted (agentPanel + agentSidePanel)`);
+    }
+    catch (e) {
+        logToFile(`[APPROVAL-2] panel focus failed: ${e.message}`);
+    }
+    // All 7 approval commands in priority order (terminal first for Run button)
+    const commands = approved
+        ? [
+            // Terminal commands (Run button) — UNTESTED INDIVIDUALLY (devlog-004)
+            'antigravity.terminalCommand.run', // SDK: TERMINAL_RUN
+            'antigravity.terminalCommand.accept', // SDK: TERMINAL_ACCEPT
+            // General command approval
+            'antigravity.command.accept', // SDK: COMMAND_ACCEPT
+            // Agent step approval (known: Silent Success with focus)
+            'antigravity.agent.acceptAgentStep', // SDK: ACCEPT_AGENT_STEP
+            // Cascade action (experimental)
+            // 'antigravity.executeCascadeAction',   // SDK: needs action param — skip
+        ]
+        : [
+            'antigravity.terminalCommand.reject', // SDK: TERMINAL_REJECT
+            'antigravity.command.reject', // SDK: COMMAND_REJECT
+            'antigravity.agent.rejectAgentStep', // SDK: REJECT_AGENT_STEP
+        ];
+    for (let i = 0; i < commands.length; i++) {
+        const cmd = commands[i];
+        try {
+            const t0 = Date.now();
+            logToFile(`[APPROVAL-2${String.fromCharCode(65 + i)}] executing: ${cmd}`);
+            const result = await vscode.commands.executeCommand(cmd);
+            const dt = Date.now() - t0;
+            logToFile(`[APPROVAL-2${String.fromCharCode(65 + i)}] returned: ${JSON.stringify(result)} (${dt}ms)`);
+        }
+        catch (e) {
+            logToFile(`[APPROVAL-2${String.fromCharCode(65 + i)}] ❌ FAIL: ${e.message}`);
+        }
+    }
+    // ── Strategy 3: ResolveOutstandingSteps (REJECT ONLY — this CANCELS!) ──
+    if (!approved && sdk) {
+        try {
+            logToFile(`[APPROVAL-3] ResolveOutstandingSteps (REJECT/CANCEL only!)`);
+            const rpcResult = await sdk.ls.rawRPC('ResolveOutstandingSteps', {
+                cascadeId: sessionId,
+            });
+            logToFile(`[APPROVAL-3] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
+            return `RPC-3:ResolveOutstandingSteps(cancel)`;
+        }
+        catch (e) {
+            logToFile(`[APPROVAL-3] ❌ FAIL: ${e.message}`);
+        }
+    }
+    logToFile(`[APPROVAL] ⚠️ ALL strategies attempted — check logs for results`);
+    return `ALL_ATTEMPTED:${action}`;
+}
 // ─── Activation ───
 async function activate(context) {
     console.log('Gravity Bridge: activating...');

extension/src/extension.ts

@@ -660,23 +660,6 @@ function generateApprovalObserverScript(_port: number): string {
     if(document.body)searchRoots.push(document.body);
     if(!searchRoots.length)return;
 
-    // ── DIAGNOSTIC: dump ALL button-like elements EVERY scan cycle ──
-    {
-      var dumpBtns=[];
-      var totalChecked=0;
-      for(var dr=0;dr<searchRoots.length;dr++){
-        var dbs=searchRoots[dr].querySelectorAll('button,[role="button"]');
-        totalChecked+=dbs.length;
-        for(var di=0;di<dbs.length;di++){
-          var db=dbs[di];
-          var dt=(db.textContent||'').trim();
-          var dtShort=dt.replace(/\\s+/g,' ').substring(0,50);
-          dumpBtns.push(db.tagName+'['+dt.length+']:'+dtShort);
-        }
-      }
-      log('[BTN-DUMP] roots='+searchRoots.length+' total='+totalChecked+' btns='+dumpBtns.length+': '+JSON.stringify(dumpBtns.slice(0,15)));
-    }
-
     var seen={}; // dedupe buttons across search roots
     for(var r=0;r<searchRoots.length;r++){
     var allBtns=searchRoots[r].querySelectorAll('button');
@@ -906,6 +889,7 @@ function setupMonitor() {
     let sawRunningAfterPending = true; // gate: must see delta>0 before next pending
     let lastModTime = ''; // track lastModifiedTime to distinguish thinking vs approval
     let stallProbed = false; // prevent repeated step probes during same stall
+    let lastRelayedTaskText = ''; // dedup TASK_BOUNDARY relay
 
     setInterval(async () => {
         pollCount++;
@@ -988,9 +972,8 @@ function setupMonitor() {
             }
 
             // ── PRIMARY: Step-probe-based approval detection ──
-            // latestToolCallStep does NOT exist in GetAllCascadeTrajectories response.
+            // On stall (idle=1, ~5s), probe GetCascadeTrajectorySteps to check WAITING.
-            // Instead, on early stall (idle=2, ~10s), probe GetCascadeTrajectorySteps
+            // 775-step limit: probe fails for long sessions → faster stall fallback.
-            // to fetch the latest step and check if it's a tool call awaiting approval.
 
             // ── STALL-BASED approval detection with step probe ──
 
@@ -1025,7 +1008,7 @@ function setupMonitor() {
                 // ── Step probe: on stall, fetch latest step via cascadeId (retry until WAITING found) ──
                 // CONFIRMED: param='cascadeId', id=sessionId (map key from trajectorySummaries)
                 // Retries every 2 polls (~10s) because RUN_COMMAND step may not be created yet
-                if (consecutiveIdleCount >= 1 && consecutiveIdleCount % 2 === 1 && !stallProbed) {
+                if (consecutiveIdleCount >= 1 && !stallProbed) {
                     try {
                         const stepsResp = await sdk.ls.rawRPC('GetCascadeTrajectorySteps', {
                             cascadeId: bestSessionId,
@@ -1034,6 +1017,9 @@ function setupMonitor() {
                             const steps = stepsResp.steps;
                             // Diagnostic: compare returned steps vs trajectory stepCount
                             logToFile(`[STEP-PROBE] returned=${steps.length} vs trajectory.stepCount=${currentCount}`);
+                            if (steps.length < currentCount) {
+                                logToFile(`[STEP-PROBE] ⚠️ 775-limit hit! steps=${steps.length} < stepCount=${currentCount}`);
+                            }
                             
                             // Scan last 5 steps backwards to find WAITING (RUN_COMMAND may not be last)
                             let foundWaiting = false;
@@ -1096,8 +1082,9 @@ function setupMonitor() {
                 const now = Date.now();
                 const cooldownOk = (now - lastPendingTime) > 60_000;
 
-                if (consecutiveIdleCount >= 8 && sawRunningAfterPending && cooldownOk) {
+                const fallbackThreshold = (currentCount > 770) ? 4 : 8; // 775-limit: faster fallback
-                    // 8 polls × 5s = 40 seconds — fallback (reduced from 100s)
+                if (consecutiveIdleCount >= fallbackThreshold && sawRunningAfterPending && cooldownOk) {
+                    // Dynamic fallback: 20s (>770 steps) or 40s (normal)
                     lastPendingStepIndex = currentCount;
                     lastPendingTime = now;
                     sawRunningAfterPending = false;
@@ -1107,7 +1094,7 @@ function setupMonitor() {
 
                     logToFile(`[STALL-FALLBACK] step=${currentCount} frozenCount=${consecutiveIdleCount} → pending`);
                     writePendingApproval({ conversation_id: activeSessionId, command, description, source: 'stall_fallback' });
-                } else if (consecutiveIdleCount === 8) {
+                } else if (consecutiveIdleCount === fallbackThreshold) {
                     const reasons = [];
                     if (!sawRunningAfterPending) reasons.push('needDelta>0');
                     if (!cooldownOk) reasons.push(`cooldown(${Math.round((60000 - (now - lastPendingTime)) / 1000)}s)`);
@@ -1123,9 +1110,12 @@ function setupMonitor() {
             if (notifyStep && notifyStep.stepIndex > lastNotifyStepIndex) {
                 lastNotifyStepIndex = notifyStep.stepIndex;
                 const content = notifyStep.step?.notifyUser?.notificationContent || '';
-                if (content.length > 10) {
+                // Filter: only relay meaningful notifications (skip trivial ones)
+                if (content.length > 50) {
                     writeChatSnapshot(`📣 **알림** (step ${notifyStep.stepIndex})\n\n${content}`);
                     console.log(`Gravity Bridge: [POLL#${pollCount}] NOTIFY step=${notifyStep.stepIndex} ${content.length} chars`);
+                } else if (content.length > 0) {
+                    logToFile(`[POLL] NOTIFY skipped (too short: ${content.length} chars): ${content.substring(0, 80)}`);
                 }
             }
 
@@ -1136,8 +1126,15 @@ function setupMonitor() {
                 const tb = taskStep.step?.taskBoundary;
                 if (tb?.taskName) {
                     const mode = tb.mode ? tb.mode.replace('AGENT_MODE_', '') : '';
-                    writeChatSnapshot(`📋 **[${mode}] ${tb.taskName}**\n${tb.taskStatus || ''}\n\n${tb.taskSummary || ''}`);
+                    // Filter: skip status-only updates with same task name (noise)
-                    console.log(`Gravity Bridge: [POLL#${pollCount}] TASK step=${taskStep.stepIndex} "${tb.taskName}"`);
+                    const taskText = `${tb.taskName}|${tb.taskStatus || ''}`;
+                    if (taskText !== lastRelayedTaskText) {
+                        lastRelayedTaskText = taskText;
+                        writeChatSnapshot(`📋 **[${mode}] ${tb.taskName}**\n${tb.taskStatus || ''}\n\n${tb.taskSummary || ''}`);
+                        console.log(`Gravity Bridge: [POLL#${pollCount}] TASK step=${taskStep.stepIndex} "${tb.taskName}"`);
+                    } else {
+                        logToFile(`[POLL] TASK skipped (duplicate): "${tb.taskName}"`);
+                    }
                 }
             }
         } catch (e: any) {
@@ -1207,9 +1204,10 @@ async function processResponseFile(filePath: string) {
             } catch { }
         }
 
-        // ═══ APPROVAL STRATEGY ═══
+        // ═══ MULTI-STRATEGY APPROVAL (v2.1) ═══
-        // DOM observer approvals: renderer handles clicking via pollResponse — skip VS Code commands
+        // Tries multiple methods sequentially with detailed logging.
-        // Stall-detection approvals: use VS Code commands as fallback (focus-dependent)
+        // DOM observer: renderer handles clicking via pollResponse
+        // Step probe/stall: try RPC → VS Code commands → log results
 
         const approved = resp.approved;
 
@@ -1217,47 +1215,15 @@ 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: relay approval to DOM observer pending files
+            // Step probe / stall path: try all approval strategies
-            // The renderer polls /response/:rid and can click the actual button
+            logToFile(`[RESPONSE] step_probe/stall → trying multi-strategy approval`);
-            logToFile(`[RESPONSE] step_probe/stall → relaying to DOM observer pending files`);
+            const approvalResult = await tryApprovalStrategies(approved, sessionId);
-
+            logToFile(`[RESPONSE] approval result: ${approvalResult}`);
-            const pendingDir = path.join(bridgePath, 'pending');
-            const responseDir = path.join(bridgePath, 'response');
-            if (!fs.existsSync(responseDir)) fs.mkdirSync(responseDir, { recursive: true });
-
-            let relayCount = 0;
-            try {
-                const files = fs.readdirSync(pendingDir).filter((f: string) => f.endsWith('.json'));
-                for (const f of files) {
-                    try {
-                        const pd = JSON.parse(fs.readFileSync(path.join(pendingDir, f), 'utf-8'));
-                        if (pd.source === 'dom_observer' && pd.status === 'pending') {
-                            // Write response file for this DOM observer pending
-                            const responsePayload = {
-                                request_id: pd.request_id,
-                                approved,
-                                timestamp: Date.now() / 1000,
-                            };
-                            fs.writeFileSync(
-                                path.join(responseDir, `${pd.request_id}.json`),
-                                JSON.stringify(responsePayload, null, 2), 'utf-8'
-                            );
-                            relayCount++;
-                            logToFile(`[RESPONSE] relayed to DOM pending: ${pd.request_id} approved=${approved}`);
-                        }
-                    } catch { }
-                }
-            } catch (e: any) {
-                logToFile(`[RESPONSE] relay scan error: ${e.message}`);
-            }
-            logToFile(`[RESPONSE] relayed to ${relayCount} DOM observer pending files`);
         }
 
-        logToFile(`[RESPONSE] ${approved ? 'approve' : 'reject'} done (${isDomObserver ? 'dom' : 'stall'})`);
+        logToFile(`[RESPONSE] ${approved ? 'approve' : 'reject'} done (${isDomObserver ? 'dom' : 'step_probe'})`);
 
         // Cleanup response file — BUT NOT for DOM observer!
-        // DOM observer: renderer's pollResponse needs to find it via HTTP GET /response/:rid
-        // Non-DOM (stall/step_probe relay): watcher already handled it, safe to delete
         if (!isDomObserver) {
             try { fs.unlinkSync(filePath); } catch { }
         }
@@ -1409,6 +1375,134 @@ function writePendingApproval(data: { conversation_id: string; command: string;
     }
 }
 
+// ─── Multi-Strategy Approval Execution ───
+
+/**
+ * Try multiple approval methods sequentially.
+ * Returns a string describing which method succeeded (or all failed).
+ * 
+ * Strategy order (most reliable first):
+ * 1. HandleCascadeUserInteraction RPC (cross-platform, no focus)
+ * 2. VS Code accept/reject commands (focus-dependent)
+ * 3. Log failure for manual intervention
+ */
+async function tryApprovalStrategies(approved: boolean, sessionId: string): Promise<string> {
+    const action = approved ? 'APPROVE' : 'REJECT';
+    logToFile(`[APPROVAL] Starting ${action} strategies for session ${sessionId.substring(0, 8)}`);
+
+    // ── Strategy 1: HandleCascadeUserInteraction RPC ──
+    if (sdk) {
+        // Try variant A: { cascadeId, approved }
+        try {
+            logToFile(`[APPROVAL-1A] HandleCascadeUserInteraction { cascadeId, approved: ${approved} }`);
+            const rpcResult = await sdk.ls.rawRPC('HandleCascadeUserInteraction', {
+                cascadeId: sessionId,
+                approved: approved,
+            });
+            logToFile(`[APPROVAL-1A] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
+            return `RPC-1A:HandleCascadeUserInteraction(approved=${approved})`;
+        } catch (e: any) {
+            logToFile(`[APPROVAL-1A] ❌ FAIL: ${e.message}`);
+        }
+
+        // Try variant B: { cascadeId, stepAction }
+        try {
+            const stepAction = approved ? 'STEP_ACTION_ACCEPT' : 'STEP_ACTION_REJECT';
+            logToFile(`[APPROVAL-1B] HandleCascadeUserInteraction { cascadeId, stepAction: '${stepAction}' }`);
+            const rpcResult = await sdk.ls.rawRPC('HandleCascadeUserInteraction', {
+                cascadeId: sessionId,
+                stepAction: stepAction,
+            });
+            logToFile(`[APPROVAL-1B] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
+            return `RPC-1B:HandleCascadeUserInteraction(stepAction=${stepAction})`;
+        } catch (e: any) {
+            logToFile(`[APPROVAL-1B] ❌ FAIL: ${e.message}`);
+        }
+
+        // Try variant C: { cascadeId, userAction } (experimental)
+        try {
+            const userAction = approved ? 'USER_ACTION_APPROVED' : 'USER_ACTION_REJECTED';
+            logToFile(`[APPROVAL-1C] HandleCascadeUserInteraction { cascadeId, userAction: '${userAction}' }`);
+            const rpcResult = await sdk.ls.rawRPC('HandleCascadeUserInteraction', {
+                cascadeId: sessionId,
+                userAction: userAction,
+            });
+            logToFile(`[APPROVAL-1C] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
+            return `RPC-1C:HandleCascadeUserInteraction(userAction=${userAction})`;
+        } catch (e: any) {
+            logToFile(`[APPROVAL-1C] ❌ FAIL: ${e.message}`);
+        }
+    }
+
+    // ── Strategy 2: VS Code Commands — step-type-specific (focus-dependent) ──
+    // Per SDK research (2026-03-09):
+    //   Run button = terminal command → terminalCommand.run / terminalCommand.accept
+    //   Code changes = agent step → agent.acceptAgentStep
+    //   General commands = command.accept
+    // Previously only tried acceptAgentStep (Silent Success) — now tries ALL 7
+
+    // Try to focus the panel first (required for command.accept / acceptAgentStep)
+    try {
+        logToFile(`[APPROVAL-2] focusing panel...`);
+        await vscode.commands.executeCommand('antigravity.agentPanel.focus');
+        await new Promise(resolve => setTimeout(resolve, 300));
+        await vscode.commands.executeCommand('antigravity.agentSidePanel.focus');
+        await new Promise(resolve => setTimeout(resolve, 300));
+        logToFile(`[APPROVAL-2] panel focus attempted (agentPanel + agentSidePanel)`);
+    } catch (e: any) {
+        logToFile(`[APPROVAL-2] panel focus failed: ${e.message}`);
+    }
+
+    // All 7 approval commands in priority order (terminal first for Run button)
+    const commands = approved
+        ? [
+            // Terminal commands (Run button) — UNTESTED INDIVIDUALLY (devlog-004)
+            'antigravity.terminalCommand.run',      // SDK: TERMINAL_RUN
+            'antigravity.terminalCommand.accept',   // SDK: TERMINAL_ACCEPT
+            // General command approval
+            'antigravity.command.accept',            // SDK: COMMAND_ACCEPT
+            // Agent step approval (known: Silent Success with focus)
+            'antigravity.agent.acceptAgentStep',     // SDK: ACCEPT_AGENT_STEP
+            // Cascade action (experimental)
+            // 'antigravity.executeCascadeAction',   // SDK: needs action param — skip
+        ]
+        : [
+            'antigravity.terminalCommand.reject',   // SDK: TERMINAL_REJECT
+            'antigravity.command.reject',            // SDK: COMMAND_REJECT
+            'antigravity.agent.rejectAgentStep',     // SDK: REJECT_AGENT_STEP
+        ];
+
+    for (let i = 0; i < commands.length; i++) {
+        const cmd = commands[i];
+        try {
+            const t0 = Date.now();
+            logToFile(`[APPROVAL-2${String.fromCharCode(65 + i)}] executing: ${cmd}`);
+            const result = await vscode.commands.executeCommand(cmd);
+            const dt = Date.now() - t0;
+            logToFile(`[APPROVAL-2${String.fromCharCode(65 + i)}] returned: ${JSON.stringify(result)} (${dt}ms)`);
+        } catch (e: any) {
+            logToFile(`[APPROVAL-2${String.fromCharCode(65 + i)}] ❌ FAIL: ${e.message}`);
+        }
+    }
+
+    // ── Strategy 3: ResolveOutstandingSteps (REJECT ONLY — this CANCELS!) ──
+    if (!approved && sdk) {
+        try {
+            logToFile(`[APPROVAL-3] ResolveOutstandingSteps (REJECT/CANCEL only!)`);
+            const rpcResult = await sdk.ls.rawRPC('ResolveOutstandingSteps', {
+                cascadeId: sessionId,
+            });
+            logToFile(`[APPROVAL-3] ✅ SUCCESS: ${JSON.stringify(rpcResult).substring(0, 200)}`);
+            return `RPC-3:ResolveOutstandingSteps(cancel)`;
+        } catch (e: any) {
+            logToFile(`[APPROVAL-3] ❌ FAIL: ${e.message}`);
+        }
+    }
+
+    logToFile(`[APPROVAL] ⚠️ ALL strategies attempted — check logs for results`);
+    return `ALL_ATTEMPTED:${action}`;
+}
+
 // ─── Activation ───
 
 export async function activate(context: vscode.ExtensionContext) {