feat(bridge): 승인 감지 최적화 — latestToolCallStep 즉시 감지 + DOM scan 확장

- latestToolCallStep RPC 기반 즉시 감지 (30초 stall → 5초 poll)
- DOM scan 범위: findPanel() → document.body 확장
- Accept all/Reject all 리뷰 바 패턴 추가
- Stall detection을 100초 fallback으로 약화
- extractToolCommand/extractToolDescription 헬퍼 추가
- known-issues 5건 신규 추가
- start/services workflow: Python 전체 경로 + services.md 로딩

#task-258 #task-262

.agents/references/known-issues.md

@@ -117,3 +117,30 @@
 - **원인**: Electron 렌더러 프로세스에서 동기 XHR이 보안 정책에 의해 차단됨
 - **해결**: `fetch()` + `AbortSignal.timeout(2000)` 비동기 방식으로 교체 (`tryPingAsync`)
 - **주의**: `async/await` 사용 불가 (ES5 환경). `.then()` 체이닝으로 구현
+
+### [2026-03-08] DOM Observer — Run 버튼 감지 불가 (webview iframe 격리)
+- **증상**: Allow Once/Allow This Conversation는 감지되나 Run/Accept 버튼은 감지 안 됨
+- **원인**: Trust/permission 버튼은 워크벤치 외부 DOM에 렌더링, Run/Accept는 **Antigravity 채팅 webview iframe** 내부의 별도 DOM에 렌더링. 렌더러 스크립트의 `document.querySelector()`는 iframe 내부 접근 불가
+- **해결**: Run 버튼은 DOM Observer가 아닌 **`latestToolCallStep` RPC 기반** 즉시 감지로 대체
+- **주의**: webview iframe에 스크립트 주입은 Electron `executeJavaScript()`로 가능하나, 현재 RPC 방식이 더 안정적
+
+### [2026-03-08] Accept all/Reject all 리뷰 바 — agent 패널 밖 DOM
+- **증상**: 코드 변경 리뷰 바(Accept all/Reject all)가 DOM Observer에 감지 안 됨
+- **원인**: `scan()` 함수가 `findPanel()` (`.antigravity-agent-side-panel` 등) 내부만 검색. 리뷰 바는 에디터/notification 영역에 렌더링되어 패널 밖에 있음
+- **해결**: `scan()` 검색 범위를 `document.body` 전체로 확장, `Accept all` / `Reject all` 패턴 추가
+- **주의**: 패널+body 이중 검색 시 dedupe 필요 (같은 버튼이 두 번 잡힐 수 있음)
+
+### [2026-03-08] latestToolCallStep — 즉시 승인 감지 (stall 30초 대체)
+- **증상**: 이전 stall-based 감지는 RUNNING+delta=0+modTime frozen 30초 후에야 pending 생성
+- **원인**: `GetAllCascadeTrajectories` 응답의 `latestToolCallStep` 필드를 전혀 활용하지 않았음
+- **해결**: `latestToolCallStep.step`의 status 필드에서 `WAITING` 여부를 직접 체크 → 5초 poll 1회에 즉시 감지. Stall detection은 100초 fallback으로 유지
+- **주의**: `latestToolCallStep`의 정확한 protobuf 구조(status 필드 위치)는 런타임 덤프로 확정 필요 — 첫 실행 시 `[TOOL-STEP]` 로그 확인
+
+### [2026-03-08] Extension 재설치 안전성 — 자동 패치 메커니즘
+- **증상**: Antigravity 삭제 후 재설치 시 렌더러 스크립트가 동작 안 함
+- **원인**: 재설치 시 HTML/product.json이 원본으로 리셋됨
+- **해결**: Extension의 `setupApprovalObserver()`가 **자동으로** 모든 패치를 수행:
+  1. `workbench.html` + `workbench-jetski-agent.html` 인라인 스크립트 삽입
+  2. `product.json` SHA256 체크섬 자동 업데이트
+  3. HTTP bridge 서버 시작 + 결정론적 포트
+- **주의**: 패치 후 반드시 **Antigravity 풀 재시작** 필요 (Reload Window 불가). Extension VSIX만 설치하면 수동 패치 불필요

docs/devlog/2026-03-08.md

@@ -18,3 +18,5 @@
 | 14 | 17:01~17:38 | **근본 원인 발견**: product.json 체크섬 불일치 → vscode-file:// 원본 캐시 서빙. 체크섬 수동 업데이트로 수정 | - | 🔧 |
 | 15 | 17:50~18:30 | **v0.3.5**: 포트 디스커버리 수정 (결정론적 포트 + 하드코딩), 인라인 스크립트 전환 (`<script src>` → `<script>inline</script>`), product.json 자동 체크섬 업데이트 | - | 🔧 |
 | 16 | 19:00~19:48 | 렌더러 스크립트 로딩 디버깅: sync XHR→async fetch 변환, 설치경로 불일치 발견, vscode-file:// 커스텀 파일 서빙 불가 확인, Electron 풀 재시작 필요 발견 | - | 🔧 |
+| 17 | 19:53~20:00 | **AG 재시작 성공**: GB Observer Bridge connected (port 34332), Allow Once/Allow This Conversation 감지 정상 동작 확인 | - | ✅ |
+| 18 | 20:00~20:15 | **승인 감지 최적화**: latestToolCallStep 즉시 감지 (30초→5초), DOM scan 범위 확장 (Accept all/Reject all), stall→100초 fallback | - | 🔧 |

docs/devlog/entries/20260308-018.md

@@ -0,0 +1,36 @@
+# 승인 감지 최적화 + GB Observer 동작 확인
+
+- **시간**: 2026-03-08 19:53~20:15
+- **Commit**: (이 세션)
+- **Vikunja**: #262 → done (인라인 스크립트 재시작 테스트 완료), #258 → 미완료
+
+## 핵심 발견
+
+### 1. GB Observer 재시작 성공
+- Antigravity 풀 재시작 후 인라인 스크립트 정상 동작 확인
+- `[GB Observer] Bridge connected on port 34332` + `v2 Observer active`
+- **Allow Once / Allow This Conversation** 버튼 즉시 감지 + pending 생성 정상
+
+### 2. Run 버튼 미감지 원인 (webview iframe 격리)
+- Trust/permission 버튼 → 워크벤치 외부 DOM → Observer가 감지 ✅
+- Run/Accept 버튼 → Antigravity 채팅 **webview iframe** 내부 DOM → `document.querySelector()` 접근 불가 ❌
+- **해결**: `latestToolCallStep` RPC 기반 즉시 감지로 대체
+
+### 3. Accept all/Reject all 리뷰 바
+- 코드 변경 리뷰 바는 agent 패널(`findPanel()`) 밖 → `document.body` 전체 스캔으로 확장
+
+### 4. Extension 재설치 안전성 확인
+- `setupApprovalObserver()`가 HTML 패치 + 체크섬 업데이트 **전자동** 수행
+- Extension VSIX만 설치하면 수동 패치 불필요
+- 단, 패치 후 **Antigravity 풀 재시작** 필수 (Electron 메인 프로세스 체크섬 캐시)
+
+## 코드 변경
+- `extension.ts`: latestToolCallStep 즉시 감지, DOM scan 범위 확장, stall→100초 fallback
+- `.agents/workflows/start.md`: services.md 로딩 + Python 전체 경로
+- `.agents/workflows/services.md`: Python 전체 경로
+- `.agents/references/known-issues.md`: 5건 신규 추가
+
+## 미완료
+- **latestToolCallStep protobuf 구조 확정** — 첫 실행 시 `[TOOL-STEP]` 덤프 로그로 status 필드 위치 확인 필요
+- **Accept all 리뷰 바 실제 감지 테스트** — body 스캔 후 올바르게 잡히는지 검증
+- **다음 세션**: AG 재시작 → tool call 발생 → 콘솔에서 구조 확인 → 필요 시 필드 매핑 수정

extension/out/extension.js

@@ -605,7 +605,8 @@ function generateApprovalObserverScript(_port) {
   // ── Button patterns to detect (order matters: first match wins per scan) ──
   var PATS=[
     {re:/^Run$/i,          type:'terminal_command'},
-    {re:/^Accept/i,        type:'agent_step'},
+    {re:/^Accept all$/i,   type:'diff_review'},
+    {re:/^Accept$/i,       type:'agent_step'},
     {re:/^Allow/i,         type:'permission'},
     {re:/^Approve/i,       type:'agent_step'},
     {re:/^Continue$/i,     type:'continue'},
@@ -613,7 +614,7 @@ function generateApprovalObserverScript(_port) {
   ];
 
   // Reject button patterns for finding the counterpart
-  var REJECT_RE=[/^reject$/i,/^cancel$/i,/^deny$/i,/^stop$/i,/^decline$/i];
+  var REJECT_RE=[/^reject$/i,/^reject all$/i,/^cancel$/i,/^deny$/i,/^stop$/i,/^decline$/i];
 
   // ── Stable button fingerprint (no getBoundingClientRect — scroll-safe) ──
   function btnId(b,type){
@@ -679,11 +680,18 @@ function generateApprovalObserverScript(_port) {
     var now=Date.now();
 
     var panel=findPanel();
-    if(!panel)return;
+    // Expand search: panel-scoped first, then full body for review bars
+    var searchRoots=[];
+    if(panel)searchRoots.push(panel);
+    // Always also scan body for diff review bar (Accept all/Reject all)
+    // which lives outside the agent panel in the editor notification area
+    if(document.body)searchRoots.push(document.body);
+    if(!searchRoots.length)return;
 
-    // Find ALL buttons in the panel
+    var seen={}; // dedupe buttons across search roots
-    var allBtns=panel.querySelectorAll('button');
+    for(var r=0;r<searchRoots.length;r++){
-    if(!allBtns.length)return;
+    var allBtns=searchRoots[r].querySelectorAll('button');
+    if(!allBtns.length)continue;
 
     for(var j=0;j<allBtns.length;j++){
       var b=allBtns[j];
@@ -736,6 +744,7 @@ function generateApprovalObserverScript(_port) {
       // Process ONE button per scan cycle (avoid flooding)
       return;
     }
+    } // end searchRoots loop
   }
 
   // ── Poll for Discord response ──
@@ -903,6 +912,8 @@ function setupMonitor() {
     let lastKnownStepCount = 0;
     let lastNotifyStepIndex = -1;
     let lastTaskStepIndex = -1;
+    let lastToolStepIndex = -1; // track latestToolCallStep for instant detection
+    let toolStepDumped = false; // dump structure once for debugging
     let lastPendingStepIndex = -1; // dedup: don't re-create pending for same step
     let consecutiveIdleCount = 0; // debounce: require N consecutive stall polls
     let lastPendingTime = 0; // cooldown: minimum gap between pendings
@@ -975,16 +986,56 @@ function setupMonitor() {
             if (pollCount <= 10 || pollCount % 6 === 0 || delta > 0) {
                 logToFile(`[POLL#${pollCount}] status=${statusStr} steps=${currentCount} delta=${delta}`);
             }
-            // ── Stall-based approval detection ──
+            // ── PRIMARY: Instant tool-call-based approval detection ──
-            // INSIGHT: Both thinking and approval show RUNNING+delta=0.
+            // latestToolCallStep contains the most recent tool call with its status.
-            // DIFFERENTIATOR: lastModifiedTime
+            // If status includes WAITING, the AI is waiting for user approval.
-            //   - Thinking: lastModifiedTime KEEPS CHANGING (server actively processing)
+            const toolStep = bestSession.latestToolCallStep;
-            //   - Approval wait: lastModifiedTime FROZEN (server idle, waiting for user)
+            if (toolStep && toolStep.stepIndex > lastToolStepIndex) {
-            // DEBUG: dump session keys on first poll to find modTime field
+                const stepData = toolStep.step || {};
+                // Dump structure once for debugging (keys + first 800 chars)
+                if (!toolStepDumped) {
+                    logToFile(`[TOOL-STEP] keys: ${JSON.stringify(Object.keys(toolStep))}`);
+                    logToFile(`[TOOL-STEP] step keys: ${JSON.stringify(Object.keys(stepData))}`);
+                    logToFile(`[TOOL-STEP] full: ${JSON.stringify(toolStep).substring(0, 800)}`);
+                    toolStepDumped = true;
+                }
+                // Check for WAITING status in various possible field locations
+                const stepStatus = String(stepData.status || toolStep.status || stepData.stepStatus || '').toLowerCase();
+                const isWaiting = stepStatus.includes('waiting') || stepStatus.includes('pending');
+                if (isWaiting && toolStep.stepIndex !== lastPendingStepIndex) {
+                    // INSTANT detection! Extract tool call details
+                    const command = extractToolCommand(stepData);
+                    const description = extractToolDescription(stepData, currentTitle, toolStep.stepIndex);
+                    logToFile(`[TOOL-DETECT] INSTANT! step=${toolStep.stepIndex} status='${stepStatus}' cmd='${command}'`);
+                    lastPendingStepIndex = toolStep.stepIndex;
+                    lastPendingTime = Date.now();
+                    sawRunningAfterPending = false;
+                    writePendingApproval({
+                        conversation_id: activeSessionId,
+                        command,
+                        description,
+                        step_type: 'tool_call',
+                        step_index: toolStep.stepIndex,
+                    });
+                }
+                else if (!isWaiting) {
+                    // Tool call exists but not waiting — log for diagnostics
+                    if (toolStep.stepIndex > lastToolStepIndex) {
+                        logToFile(`[TOOL-STEP] step=${toolStep.stepIndex} status='${stepStatus}' (not waiting)`);
+                    }
+                }
+                lastToolStepIndex = toolStep.stepIndex;
+            }
+            // ── FALLBACK: Stall-based approval detection ──
+            // Kept as safety net for edge cases where latestToolCallStep doesn't report WAITING.
+            // Threshold raised from 6 (30s) to 20 (100s) since primary detection above handles most cases.
+            // DEBUG: dump session keys on first poll
             if (pollCount === 1) {
                 const keys = Object.keys(bestSession).filter(k => !['latestNotifyUserStep', 'latestTaskBoundaryStep', 'latestToolCallStep'].includes(k));
                 logToFile(`[DEBUG] session keys: ${keys.join(', ')}`);
-                logToFile(`[DEBUG] lastModifiedTime=${bestSession.lastModifiedTime}, lastModifiedTimestamp=${bestSession.lastModifiedTimestamp}, modifiedTime=${bestSession.modifiedTime}`);
+                logToFile(`[DEBUG] lastModifiedTime=${bestSession.lastModifiedTime}`);
+                if (toolStep)
+                    logToFile(`[DEBUG] toolStep present, stepIndex=${toolStep.stepIndex}`);
             }
             const currentModTime = bestSession.lastModifiedTime || bestSession.lastModifiedTimestamp || bestSession.modifiedTime || '';
             const modTimeChanged = currentModTime !== lastModTime;
@@ -1013,17 +1064,17 @@ function setupMonitor() {
                 lastModTime = currentModTime;
                 const now = Date.now();
                 const cooldownOk = (now - lastPendingTime) > 60_000;
-                if (consecutiveIdleCount >= 6 && sawRunningAfterPending && cooldownOk) {
+                if (consecutiveIdleCount >= 20 && sawRunningAfterPending && cooldownOk) {
-                    // 6 polls × 5s = 30 seconds of FROZEN stall = approval waiting
+                    // 20 polls × 5s = 100 seconds — fallback only
                     lastPendingStepIndex = currentCount;
                     lastPendingTime = now;
                     sawRunningAfterPending = false;
-                    const command = `Stall at step ${currentCount}`;
+                    const command = `Stall at step ${currentCount} (fallback)`;
-                    const description = `승인 대기 감지 (${consecutiveIdleCount * 5}초 정지), Title: "${currentTitle}"`;
+                    const description = `승인 대기 감지 — fallback (${consecutiveIdleCount * 5}초 정지), Title: "${currentTitle}"`;
-                    logToFile(`[STALL] step=${currentCount} frozenCount=${consecutiveIdleCount} → pending`);
+                    logToFile(`[STALL-FALLBACK] step=${currentCount} frozenCount=${consecutiveIdleCount} → pending`);
                     writePendingApproval({ conversation_id: activeSessionId, command, description });
                 }
-                else if (consecutiveIdleCount === 6) {
+                else if (consecutiveIdleCount === 20) {
                     const reasons = [];
                     if (!sawRunningAfterPending)
                         reasons.push('needDelta>0');
@@ -1269,6 +1320,49 @@ function filterEphemeral(text) {
     }
     return text;
 }
+/** Extract human-readable command from a tool call step's data. */
+function extractToolCommand(stepData) {
+    // Try common step data shapes from protobuf
+    if (stepData.runCommand) {
+        return stepData.runCommand.commandLine || stepData.runCommand.command || 'Run Command';
+    }
+    if (stepData.writeToFile) {
+        const target = stepData.writeToFile.targetFile || stepData.writeToFile.filePath || 'file';
+        return `Write: ${target.split(/[\\/]/).pop()}`;
+    }
+    if (stepData.codeAction) {
+        const fp = stepData.codeAction.filePath || '';
+        return `Edit: ${fp.split(/[\\/]/).pop() || 'file'}`;
+    }
+    if (stepData.replaceFileContent || stepData.multiReplaceFileContent) {
+        const d = stepData.replaceFileContent || stepData.multiReplaceFileContent;
+        const fp = d.targetFile || d.filePath || '';
+        return `Edit: ${fp.split(/[\\/]/).pop() || 'file'}`;
+    }
+    if (stepData.sendCommandInput) {
+        return `Send Input: ${(stepData.sendCommandInput.input || '').substring(0, 50)}`;
+    }
+    // Generic fallback: use first key name
+    const keys = Object.keys(stepData).filter(k => k !== 'status' && k !== 'stepStatus');
+    return keys.length > 0 ? keys[0] : 'Unknown tool call';
+}
+/** Extract description from a tool call step for Discord display. */
+function extractToolDescription(stepData, sessionTitle, stepIndex) {
+    const parts = [`Step #${stepIndex}`, `Session: "${sessionTitle}"`];
+    // Try to get code/command content for context
+    if (stepData.runCommand) {
+        const cmd = stepData.runCommand.commandLine || stepData.runCommand.command || '';
+        if (cmd)
+            parts.push(`Command: ${cmd.substring(0, 200)}`);
+    }
+    if (stepData.writeToFile?.targetFile) {
+        parts.push(`File: ${stepData.writeToFile.targetFile}`);
+    }
+    if (stepData.codeAction?.filePath) {
+        parts.push(`File: ${stepData.codeAction.filePath}`);
+    }
+    return parts.join('\n');
+}
 /** Write a pending approval file matching Bot's ApprovalRequest dataclass. */
 function writePendingApproval(data) {
     try {
@@ -1286,6 +1380,8 @@ function writePendingApproval(data) {
             status: 'pending',
             discord_message_id: 0,
             project_name: projectName,
+            ...(data.step_type ? { step_type: data.step_type } : {}),
+            ...(data.step_index !== undefined ? { step_index: data.step_index } : {}),
         };
         fs.writeFileSync(path.join(pendingDir, `${id}.json`), JSON.stringify(payload, null, 2), 'utf-8');
         console.log(`Gravity Bridge: pending approval written → ${id}.json`);

extension/src/extension.ts

@@ -576,7 +576,8 @@ function generateApprovalObserverScript(_port: number): string {
   // ── Button patterns to detect (order matters: first match wins per scan) ──
   var PATS=[
     {re:/^Run$/i,          type:'terminal_command'},
-    {re:/^Accept/i,        type:'agent_step'},
+    {re:/^Accept all$/i,   type:'diff_review'},
+    {re:/^Accept$/i,       type:'agent_step'},
     {re:/^Allow/i,         type:'permission'},
     {re:/^Approve/i,       type:'agent_step'},
     {re:/^Continue$/i,     type:'continue'},
@@ -584,7 +585,7 @@ function generateApprovalObserverScript(_port: number): string {
   ];
 
   // Reject button patterns for finding the counterpart
-  var REJECT_RE=[/^reject$/i,/^cancel$/i,/^deny$/i,/^stop$/i,/^decline$/i];
+  var REJECT_RE=[/^reject$/i,/^reject all$/i,/^cancel$/i,/^deny$/i,/^stop$/i,/^decline$/i];
 
   // ── Stable button fingerprint (no getBoundingClientRect — scroll-safe) ──
   function btnId(b,type){
@@ -650,11 +651,18 @@ function generateApprovalObserverScript(_port: number): string {
     var now=Date.now();
 
     var panel=findPanel();
-    if(!panel)return;
+    // Expand search: panel-scoped first, then full body for review bars
+    var searchRoots=[];
+    if(panel)searchRoots.push(panel);
+    // Always also scan body for diff review bar (Accept all/Reject all)
+    // which lives outside the agent panel in the editor notification area
+    if(document.body)searchRoots.push(document.body);
+    if(!searchRoots.length)return;
 
-    // Find ALL buttons in the panel
+    var seen={}; // dedupe buttons across search roots
-    var allBtns=panel.querySelectorAll('button');
+    for(var r=0;r<searchRoots.length;r++){
-    if(!allBtns.length)return;
+    var allBtns=searchRoots[r].querySelectorAll('button');
+    if(!allBtns.length)continue;
 
     for(var j=0;j<allBtns.length;j++){
       var b=allBtns[j];
@@ -707,6 +715,7 @@ function generateApprovalObserverScript(_port: number): string {
       // Process ONE button per scan cycle (avoid flooding)
       return;
     }
+    } // end searchRoots loop
   }
 
   // ── Poll for Discord response ──
@@ -873,6 +882,8 @@ function setupMonitor() {
     let lastKnownStepCount = 0;
     let lastNotifyStepIndex = -1;
     let lastTaskStepIndex = -1;
+    let lastToolStepIndex = -1; // track latestToolCallStep for instant detection
+    let toolStepDumped = false; // dump structure once for debugging
     let lastPendingStepIndex = -1; // dedup: don't re-create pending for same step
     let consecutiveIdleCount = 0; // debounce: require N consecutive stall polls
     let lastPendingTime = 0; // cooldown: minimum gap between pendings
@@ -951,17 +962,62 @@ function setupMonitor() {
                 logToFile(`[POLL#${pollCount}] status=${statusStr} steps=${currentCount} delta=${delta}`);
             }
 
-            // ── Stall-based approval detection ──
+            // ── PRIMARY: Instant tool-call-based approval detection ──
-            // INSIGHT: Both thinking and approval show RUNNING+delta=0.
+            // latestToolCallStep contains the most recent tool call with its status.
-            // DIFFERENTIATOR: lastModifiedTime
+            // If status includes WAITING, the AI is waiting for user approval.
-            //   - Thinking: lastModifiedTime KEEPS CHANGING (server actively processing)
+            const toolStep = bestSession.latestToolCallStep;
-            //   - Approval wait: lastModifiedTime FROZEN (server idle, waiting for user)
+            if (toolStep && toolStep.stepIndex > lastToolStepIndex) {
+                const stepData = toolStep.step || {};
+                // Dump structure once for debugging (keys + first 800 chars)
+                if (!toolStepDumped) {
+                    logToFile(`[TOOL-STEP] keys: ${JSON.stringify(Object.keys(toolStep))}`);
+                    logToFile(`[TOOL-STEP] step keys: ${JSON.stringify(Object.keys(stepData))}`);
+                    logToFile(`[TOOL-STEP] full: ${JSON.stringify(toolStep).substring(0, 800)}`);
+                    toolStepDumped = true;
+                }
 
-            // DEBUG: dump session keys on first poll to find modTime field
+                // Check for WAITING status in various possible field locations
+                const stepStatus = String(
+                    stepData.status || toolStep.status || stepData.stepStatus || ''
+                ).toLowerCase();
+                const isWaiting = stepStatus.includes('waiting') || stepStatus.includes('pending');
+
+                if (isWaiting && toolStep.stepIndex !== lastPendingStepIndex) {
+                    // INSTANT detection! Extract tool call details
+                    const command = extractToolCommand(stepData);
+                    const description = extractToolDescription(stepData, currentTitle, toolStep.stepIndex);
+                    logToFile(`[TOOL-DETECT] INSTANT! step=${toolStep.stepIndex} status='${stepStatus}' cmd='${command}'`);
+
+                    lastPendingStepIndex = toolStep.stepIndex;
+                    lastPendingTime = Date.now();
+                    sawRunningAfterPending = false;
+
+                    writePendingApproval({
+                        conversation_id: activeSessionId,
+                        command,
+                        description,
+                        step_type: 'tool_call',
+                        step_index: toolStep.stepIndex,
+                    });
+                } else if (!isWaiting) {
+                    // Tool call exists but not waiting — log for diagnostics
+                    if (toolStep.stepIndex > lastToolStepIndex) {
+                        logToFile(`[TOOL-STEP] step=${toolStep.stepIndex} status='${stepStatus}' (not waiting)`);
+                    }
+                }
+                lastToolStepIndex = toolStep.stepIndex;
+            }
+
+            // ── FALLBACK: Stall-based approval detection ──
+            // Kept as safety net for edge cases where latestToolCallStep doesn't report WAITING.
+            // Threshold raised from 6 (30s) to 20 (100s) since primary detection above handles most cases.
+
+            // DEBUG: dump session keys on first poll
             if (pollCount === 1) {
                 const keys = Object.keys(bestSession).filter(k => !['latestNotifyUserStep', 'latestTaskBoundaryStep', 'latestToolCallStep'].includes(k));
                 logToFile(`[DEBUG] session keys: ${keys.join(', ')}`);
-                logToFile(`[DEBUG] lastModifiedTime=${bestSession.lastModifiedTime}, lastModifiedTimestamp=${(bestSession as any).lastModifiedTimestamp}, modifiedTime=${(bestSession as any).modifiedTime}`);
+                logToFile(`[DEBUG] lastModifiedTime=${bestSession.lastModifiedTime}`);
+                if (toolStep) logToFile(`[DEBUG] toolStep present, stepIndex=${toolStep.stepIndex}`);
             }
 
             const currentModTime = bestSession.lastModifiedTime || (bestSession as any).lastModifiedTimestamp || (bestSession as any).modifiedTime || '';
@@ -993,18 +1049,18 @@ function setupMonitor() {
                 const now = Date.now();
                 const cooldownOk = (now - lastPendingTime) > 60_000;
 
-                if (consecutiveIdleCount >= 6 && sawRunningAfterPending && cooldownOk) {
+                if (consecutiveIdleCount >= 20 && sawRunningAfterPending && cooldownOk) {
-                    // 6 polls × 5s = 30 seconds of FROZEN stall = approval waiting
+                    // 20 polls × 5s = 100 seconds — fallback only
                     lastPendingStepIndex = currentCount;
                     lastPendingTime = now;
                     sawRunningAfterPending = false;
 
-                    const command = `Stall at step ${currentCount}`;
+                    const command = `Stall at step ${currentCount} (fallback)`;
-                    const description = `승인 대기 감지 (${consecutiveIdleCount * 5}초 정지), Title: "${currentTitle}"`;
+                    const description = `승인 대기 감지 — fallback (${consecutiveIdleCount * 5}초 정지), Title: "${currentTitle}"`;
 
-                    logToFile(`[STALL] step=${currentCount} frozenCount=${consecutiveIdleCount} → pending`);
+                    logToFile(`[STALL-FALLBACK] step=${currentCount} frozenCount=${consecutiveIdleCount} → pending`);
                     writePendingApproval({ conversation_id: activeSessionId, command, description });
-                } else if (consecutiveIdleCount === 6) {
+                } else if (consecutiveIdleCount === 20) {
                     const reasons = [];
                     if (!sawRunningAfterPending) reasons.push('needDelta>0');
                     if (!cooldownOk) reasons.push(`cooldown(${Math.round((60000 - (now - lastPendingTime)) / 1000)}s)`);
@@ -1239,8 +1295,52 @@ function filterEphemeral(text: string): string | null {
     return text;
 }
 
+/** Extract human-readable command from a tool call step's data. */
+function extractToolCommand(stepData: any): string {
+    // Try common step data shapes from protobuf
+    if (stepData.runCommand) {
+        return stepData.runCommand.commandLine || stepData.runCommand.command || 'Run Command';
+    }
+    if (stepData.writeToFile) {
+        const target = stepData.writeToFile.targetFile || stepData.writeToFile.filePath || 'file';
+        return `Write: ${target.split(/[\\/]/).pop()}`;
+    }
+    if (stepData.codeAction) {
+        const fp = stepData.codeAction.filePath || '';
+        return `Edit: ${fp.split(/[\\/]/).pop() || 'file'}`;
+    }
+    if (stepData.replaceFileContent || stepData.multiReplaceFileContent) {
+        const d = stepData.replaceFileContent || stepData.multiReplaceFileContent;
+        const fp = d.targetFile || d.filePath || '';
+        return `Edit: ${fp.split(/[\\/]/).pop() || 'file'}`;
+    }
+    if (stepData.sendCommandInput) {
+        return `Send Input: ${(stepData.sendCommandInput.input || '').substring(0, 50)}`;
+    }
+    // Generic fallback: use first key name
+    const keys = Object.keys(stepData).filter(k => k !== 'status' && k !== 'stepStatus');
+    return keys.length > 0 ? keys[0] : 'Unknown tool call';
+}
+
+/** Extract description from a tool call step for Discord display. */
+function extractToolDescription(stepData: any, sessionTitle: string, stepIndex: number): string {
+    const parts = [`Step #${stepIndex}`, `Session: "${sessionTitle}"`];
+    // Try to get code/command content for context
+    if (stepData.runCommand) {
+        const cmd = stepData.runCommand.commandLine || stepData.runCommand.command || '';
+        if (cmd) parts.push(`Command: ${cmd.substring(0, 200)}`);
+    }
+    if (stepData.writeToFile?.targetFile) {
+        parts.push(`File: ${stepData.writeToFile.targetFile}`);
+    }
+    if (stepData.codeAction?.filePath) {
+        parts.push(`File: ${stepData.codeAction.filePath}`);
+    }
+    return parts.join('\n');
+}
+
 /** Write a pending approval file matching Bot's ApprovalRequest dataclass. */
-function writePendingApproval(data: { conversation_id: string; command: string; description: string }) {
+function writePendingApproval(data: { conversation_id: string; command: string; description: string; step_type?: string; step_index?: number }) {
     try {
         const pendingDir = path.join(bridgePath, 'pending');
         if (!fs.existsSync(pendingDir)) { fs.mkdirSync(pendingDir, { recursive: true }); }
@@ -1254,6 +1354,8 @@ function writePendingApproval(data: { conversation_id: string; command: string;
             status: 'pending',
             discord_message_id: 0,
             project_name: projectName,
+            ...(data.step_type ? { step_type: data.step_type } : {}),
+            ...(data.step_index !== undefined ? { step_index: data.step_index } : {}),
         };
         fs.writeFileSync(path.join(pendingDir, `${id}.json`), JSON.stringify(payload, null, 2), 'utf-8');
         console.log(`Gravity Bridge: pending approval written → ${id}.json`);