fix(bridge): v0.3.12 approval state management — sawRunningAfterPending gate + approval-flow.md system doc

- processResponseFile: set sawRunningAfterPending=true instead of removing resets
  (prevents infinite pending loop AND known-issues L479 auto_resolve regression)
- Hoist sawRunningAfterPending to module level for cross-function access
- Add recentPendingSteps memory dedup Map (60s TTL) for file-deletion resilience
- Create docs/approval-flow.md: complete system flow guide with state diagram
- Update known-issues.md: 2 new entries (state reset fix, memory dedup)

extension/out/extension.js

@@ -82,6 +82,11 @@ let deterministicPort = 0; // derived from projectName, consistent across restar
 let watcher = null;
 let commandsWatcher = null;
 const sentPendingIds = new Set();
+// Memory-based dedup: tracks recently created pending step_indexes to prevent
+// regeneration after pending file deletion (by Collector/Bot response cycle).
+// Map<string, number> = `${conversationId}:${stepIndex}` → creation timestamp
+const recentPendingSteps = new Map();
+const PENDING_MEMORY_TTL_MS = 60_000; // 60 seconds memory retention
 // ─── Project Detection ───
 function detectProjectName() {
     const config = vscode.workspace.getConfiguration('gravityBridge');
@@ -706,6 +711,7 @@ let clickTrigger = null;
 let sessionStalled = false; // true when session is stalled waiting for approval
 let lastPendingStepIndex = -1; // dedup: don't re-create pending for same step (module-level for tryApprovalStrategies)
 let stallProbed = false; // prevent repeated step probes during same stall (module-level for processResponseFile reset)
+let sawRunningAfterPending = true; // gate: must see delta>0 before next pending (module-level for processResponseFile)
 // Deep inspect trigger: curl sets this, renderer picks it up and POSTs results back
 let deepInspectRequested = false;
 let deepInspectResult = null;
@@ -1733,7 +1739,7 @@ function setupMonitor() {
     // lastPendingStepIndex is module-level (above sessionStalled)
     let consecutiveIdleCount = 0; // debounce: require N consecutive stall polls
     let lastPendingTime = 0; // cooldown: minimum gap between pendings
-    let sawRunningAfterPending = true; // gate: must see delta>0 before next pending
+    // sawRunningAfterPending is module-level (used by processResponseFile to close auto_resolve gate)
     let lastModTime = ''; // track lastModifiedTime to distinguish thinking vs approval
     // stallProbed is module-level (used by processResponseFile to reset after approval)
     let lastRelayedTaskText = ''; // dedup TASK_BOUNDARY relay
@@ -1969,6 +1975,11 @@ function setupMonitor() {
                         writeChatSnapshot(`✅ **AG에서 직접 진행됨** (step ${lastPendingStepIndex})\n\n\`${primaryCommand.substring(0, 200)}\``);
                     }
                     lastPendingStepIndex = -1;
+                    // Clear memory dedup for this session (step progressed, new WAITING steps are allowed)
+                    for (const k of recentPendingSteps.keys()) {
+                        if (k.startsWith(activeSessionId + ':'))
+                            recentPendingSteps.delete(k);
+                    }
                 }
                 consecutiveIdleCount = 0;
                 sawRunningAfterPending = true;
@@ -2628,9 +2639,21 @@ async function processResponseFile(filePath) {
             logToFile(`[RESPONSE] strategy result: ${strategyResult}`);
         }
         logToFile(`[RESPONSE] ${approved ? 'approve' : 'reject'} done (${isDomObserver ? 'dom' : 'step_probe'})`);
-        // Reset stall probe gate so next WAITING step is detected immediately
-        stallProbed = false;
-        lastPendingStepIndex = -1;
+        // FIX v2 (2026-03-16): Correct state management after response processing.
+        //
+        // HISTORY: processResponseFile originally reset lastPendingStepIndex=-1 and stallProbed=false.
+        // v0.3.11 removed both resets entirely → caused infinite pending loop (step_probe re-detects
+        // same WAITING step because lastPendingStepIndex=-1 makes si!=lastPendingStepIndex true).
+        // v0.3.12 attempt removed resets entirely → conflicts with known-issues.md L479
+        // (auto_resolve duplicate notification on delta>0 because sawRunningAfterPending is false).
+        //
+        // CORRECT FIX: Set sawRunningAfterPending=true to close the auto_resolve gate.
+        // - lastPendingStepIndex: KEEP (prevents step_probe from re-creating pending for same step)
+        // - stallProbed: KEEP (prevents re-probe during same stall)
+        // - sawRunningAfterPending = true: CRITICAL — tells auto_resolve "Discord handled this,
+        //   don't generate duplicate 직접 진행됨 notification" (prevents known-issues L479 regression)
+        // All three reset naturally at delta > 0 (L1972-1980) when step actually progresses.
+        sawRunningAfterPending = true;
         // Cleanup response file
         // CRITICAL: DOM observer responses must NOT be deleted here!
         // The renderer polls GET /response/:rid to discover the approval.
@@ -2780,6 +2803,22 @@ function writePendingApproval(data) {
         // ── Dedup: if DOM observer already created a "Run"-only pending, MERGE detailed info into it ──
         const nowMs = Date.now();
         const DEDUP_WINDOW_MS = 15_000; // 15 second dedup window
+        // ── FIX: Memory-based dedup (survives pending file deletion by Collector/Bot) ──
+        // Pending files are deleted when Bot writes a response (bridge.py L461, collector.py L259).
+        // File-based dedup alone fails after deletion → same step_index creates new pending → loop.
+        if (data.step_index !== undefined && data.conversation_id) {
+            const memKey = `${data.conversation_id}:${data.step_index}`;
+            const prevTs = recentPendingSteps.get(memKey);
+            if (prevTs && (nowMs - prevTs) < PENDING_MEMORY_TTL_MS) {
+                logToFile(`[DEDUP-MEM] skip: step_index ${data.step_index} already created ${Math.round((nowMs - prevTs) / 1000)}s ago`);
+                return;
+            }
+            // Cleanup stale entries (keep map small)
+            for (const [k, ts] of recentPendingSteps) {
+                if (nowMs - ts > PENDING_MEMORY_TTL_MS)
+                    recentPendingSteps.delete(k);
+            }
+        }
         try {
             const existingFiles = fs.readdirSync(pendingDir).filter((f) => f.endsWith('.json'));
             for (const ef of existingFiles) {
@@ -2799,6 +2838,10 @@ function writePendingApproval(data) {
                         existing.source = 'dom_observer+step_probe'; // mark as merged
                         fs.writeFileSync(efPath, JSON.stringify(existing, null, 2), 'utf-8');
                         logToFile(`[DEDUP] MERGED step_probe info into DOM pending: ${ef} cmd="${data.command.substring(0, 60)}"`);
+                        // Record in memory dedup
+                        if (data.step_index !== undefined && data.conversation_id) {
+                            recentPendingSteps.set(`${data.conversation_id}:${data.step_index}`, nowMs);
+                        }
                         return;
                     }
                 }
@@ -2845,6 +2888,10 @@ function writePendingApproval(data) {
         };
         fs.writeFileSync(path.join(pendingDir, `${id}.json`), JSON.stringify(payload, null, 2), 'utf-8');
         console.log(`Gravity Bridge: pending approval written → ${id}.json`);
+        // Record in memory dedup cache (survives file deletion by Collector/Bot)
+        if (data.step_index !== undefined && data.conversation_id) {
+            recentPendingSteps.set(`${data.conversation_id}:${data.step_index}`, nowMs);
+        }
         // Register session → project mapping (correct because projectName is per-window)
         if (data.conversation_id) {
             writeRegistration(data.conversation_id);