tmux-team 3.2.1 → 3.2.2

package/package.json

@@ -1,25 +1,12 @@
 {
   "name": "tmux-team",
-  "version": "3.2.1",
+  "version": "3.2.2",
   "description": "CLI tool for AI agent collaboration in tmux - manage cross-pane communication",
   "type": "module",
   "bin": {
     "tmux-team": "./bin/tmux-team",
     "tmt": "./bin/tmux-team"
   },
-  "scripts": {
-    "dev": "tsx src/cli.ts",
-    "tmt": "./bin/tmux-team",
-    "test": "npm run test:run",
-    "test:watch": "vitest",
-    "test:run": "vitest run --coverage && node scripts/check-coverage.mjs --threshold 95",
-    "lint": "oxlint src/",
-    "lint:fix": "oxlint src/ --fix",
-    "format": "prettier --write src/",
-    "format:check": "prettier --check src/",
-    "type:check": "tsc --noEmit",
-    "check": "npm run type:check && npm run lint && npm run format:check"
-  },
   "keywords": [
     "tmux",
     "cli",
@@ -56,5 +43,18 @@
     "prettier": "^3.7.4",
     "typescript": "^5.3.0",
     "vitest": "^1.2.0"
+  },
+  "scripts": {
+    "dev": "tsx src/cli.ts",
+    "tmt": "./bin/tmux-team",
+    "test": "pnpm test:run",
+    "test:watch": "vitest",
+    "test:run": "vitest run --coverage && node scripts/check-coverage.mjs --threshold 95",
+    "lint": "oxlint src/",
+    "lint:fix": "oxlint src/ --fix",
+    "format": "prettier --write src/",
+    "format:check": "prettier --check src/",
+    "type:check": "tsc --noEmit",
+    "check": "pnpm type:check && pnpm lint && pnpm format:check"
   }
-}
+}

package/src/commands/talk.test.ts

@@ -14,9 +14,12 @@ import { cmdTalk } from './talk.js';
 // Constants
 // ─────────────────────────────────────────────────────────────
 
-// Regex to match new end marker format
+// Regex to match the END marker (as printed by agent) - used to find markers in output
 const END_MARKER_REGEX = /---RESPONSE-END-([a-f0-9]+)---/;
 
+// Regex to extract nonce from instruction (instruction says "RESPONSE-END-xxxx" without dashes)
+const INSTRUCTION_NONCE_REGEX = /RESPONSE-END-([a-f0-9]+)/;
+
 // ─────────────────────────────────────────────────────────────
 // Test utilities
 // ─────────────────────────────────────────────────────────────
@@ -457,11 +460,13 @@ describe('cmdTalk - --wait mode', () => {
   });
 
   // Helper: generate mock capture output with proper marker structure
-  // The end marker must appear TWICE: once in instruction, once from "agent"
-  // New format: ---RESPONSE-END-NONCE---
+  // New protocol: instruction describes the marker verbally (doesn't contain literal marker)
+  // Include the instruction line so extraction can anchor to it for clean output
   function mockCompleteResponse(nonce: string, response: string): string {
+    const instruction = `When you finish responding, output a completion marker on its own line: three dashes, RESPONSE-END-${nonce}, three dashes (no spaces).`;
     const endMarker = `---RESPONSE-END-${nonce}---`;
-    return `Hello\n\nWhen you finish responding, print this exact line:\n${endMarker}\n${response}\n${endMarker}`;
+    // Simulate: scrollback, user message with instruction, agent response, marker
+    return `Some scrollback content\nUser message here\n\n${instruction}\n${response}\n${endMarker}`;
   }
 
   it('appends nonce instruction to message', async () => {
@@ -471,9 +476,9 @@ describe('cmdTalk - --wait mode', () => {
     tmux.capture = () => {
       captureCount++;
       if (captureCount === 1) return ''; // Baseline
-      // Return marker on second capture - must include instruction AND agent's end marker
+      // Extract nonce from instruction and return agent response with marker
       const sent = tmux.sends[0]?.message || '';
-      const match = sent.match(END_MARKER_REGEX);
+      const match = sent.match(INSTRUCTION_NONCE_REGEX);
       return match ? mockCompleteResponse(match[1], 'Response here') : '';
     };
 
@@ -494,8 +499,11 @@ describe('cmdTalk - --wait mode', () => {
     await cmdTalk(ctx, 'claude', 'Hello');
 
     expect(tmux.sends).toHaveLength(1);
-    expect(tmux.sends[0].message).toContain('When you finish responding, print this exact line:');
-    expect(tmux.sends[0].message).toMatch(/---RESPONSE-END-[a-f0-9]+---/);
+    // New protocol: instruction describes marker verbally, doesn't contain literal marker
+    expect(tmux.sends[0].message).toContain('output a completion marker on its own line');
+    expect(tmux.sends[0].message).toContain('three dashes, RESPONSE-END-');
+    // Should NOT contain the literal marker format
+    expect(tmux.sends[0].message).not.toMatch(/---RESPONSE-END-[a-f0-9]+---/);
   });
 
   it('detects nonce marker and extracts response', async () => {
@@ -507,9 +515,9 @@ describe('cmdTalk - --wait mode', () => {
     tmux.capture = () => {
       captureCount++;
       if (captureCount === 1) return 'baseline content';
-      // Extract nonce from sent message and return matching marker
+      // Extract nonce from instruction and return agent response with marker
       const sent = tmux.sends[0]?.message || '';
-      const match = sent.match(END_MARKER_REGEX);
+      const match = sent.match(INSTRUCTION_NONCE_REGEX);
       if (match) {
         return mockCompleteResponse(match[1], 'Agent response here');
       }
@@ -575,20 +583,20 @@ describe('cmdTalk - --wait mode', () => {
     expect(output.error).toContain('Timed out');
   });
 
-  it('isolates response using end markers in scrollback', async () => {
+  it('isolates response using end marker in scrollback', async () => {
     const tmux = createMockTmux();
     const ui = createMockUI();
 
     const oldContent = 'Previous conversation\nOld content here';
 
     tmux.capture = () => {
-      // Simulate scrollback with old content, then our instruction (with end marker), response, and agent's end marker
+      // Simulate scrollback with old content, then agent response with marker
       const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        const endMarker = `---RESPONSE-END-${endMatch[1]}---`;
-        // Must include end marker TWICE: once in instruction, once from "agent"
-        return `${oldContent}\n\nMessage content here\n\nWhen you finish responding, print this exact line:\n${endMarker}\nNew response content\n\n${endMarker}`;
+      const nonceMatch = sent.match(INSTRUCTION_NONCE_REGEX);
+      if (nonceMatch) {
+        const endMarker = `---RESPONSE-END-${nonceMatch[1]}---`;
+        // Only ONE marker from agent
+        return `${oldContent}\nNew response content\n\n${endMarker}`;
       }
       return oldContent;
     };
@@ -612,9 +620,6 @@ describe('cmdTalk - --wait mode', () => {
 
     const output = ui.jsonOutput[0] as Record<string, unknown>;
     expect(output.status).toBe('completed');
-    // Response should NOT include old content
-    expect(output.response).not.toContain('Previous conversation');
-    expect(output.response).not.toContain('Old content here');
     // Response should contain the actual response content
     expect(output.response).toContain('New response content');
   });
@@ -627,7 +632,7 @@ describe('cmdTalk - --wait mode', () => {
       captureCount++;
       if (captureCount === 1) return '';
       const sent = tmux.sends[0]?.message || '';
-      const match = sent.match(END_MARKER_REGEX);
+      const match = sent.match(INSTRUCTION_NONCE_REGEX);
       return match ? mockCompleteResponse(match[1], 'Done') : '';
     };
 
@@ -695,7 +700,7 @@ describe('cmdTalk - --wait mode', () => {
 
     // Mock send to capture the nonce for each pane
     tmux.send = (pane: string, msg: string) => {
-      const match = msg.match(END_MARKER_REGEX);
+      const match = msg.match(INSTRUCTION_NONCE_REGEX);
       if (match) {
         noncesByPane[pane] = match[1];
       }
@@ -743,17 +748,18 @@ describe('cmdTalk - --wait mode', () => {
     const noncesByPane: Record<string, string> = {};
 
     tmux.send = (pane: string, msg: string) => {
-      const match = msg.match(END_MARKER_REGEX);
+      const match = msg.match(INSTRUCTION_NONCE_REGEX);
       if (match) {
         noncesByPane[pane] = match[1];
       }
     };
 
-    // Only pane 10.1 responds, 10.2 times out
+    // Only pane 10.1 responds with end marker, 10.2 never has end marker
     tmux.capture = (pane: string) => {
       if (pane === '10.1' && noncesByPane[pane]) {
         return mockCompleteResponse(noncesByPane[pane], 'Response from codex');
       }
+      // gemini has no end marker - still typing
       return 'still working...';
     };
 
@@ -763,10 +769,10 @@ describe('cmdTalk - --wait mode', () => {
       ui,
       tmux,
       paths,
-      flags: { wait: true, timeout: 0.1, json: true },
+      flags: { wait: true, timeout: 0.5, json: true },
       config: {
         defaults: {
-          timeout: 0.1,
+          timeout: 0.5,
           pollInterval: 0.02,
           captureLines: 100,
           preambleEvery: 3,
@@ -781,7 +787,7 @@ describe('cmdTalk - --wait mode', () => {
     try {
       await cmdTalk(ctx, 'all', 'Hello');
     } catch {
-      // Expected timeout exit
+      // Expected timeout exit for gemini
     }
 
     // Should have JSON output with both results
@@ -801,7 +807,7 @@ describe('cmdTalk - --wait mode', () => {
     const nonces: string[] = [];
 
     tmux.send = (_pane: string, msg: string) => {
-      const match = msg.match(END_MARKER_REGEX);
+      const match = msg.match(INSTRUCTION_NONCE_REGEX);
       if (match) {
         nonces.push(match[1]);
       }
@@ -910,24 +916,17 @@ describe('cmdTalk - nonce collision handling', () => {
       if (captureCount === 1) {
         return `Old question\nOld response\n${oldEndMarker}`;
       }
-      // New capture still has old markers but new request markers not complete yet
+      // New capture still has old markers but agent hasn't responded yet
       if (captureCount === 2) {
-        const sent = tmux.sends[0]?.message || '';
-        const endMatch = sent.match(END_MARKER_REGEX);
-        if (endMatch) {
-          const newEndMarker = `---RESPONSE-END-${endMatch[1]}---`;
-          // Old content + new instruction (only one occurrence of new marker so far)
-          return `Old question\nOld response\n${oldEndMarker}\n\nNew question asked\n\nWhen you finish responding, print this exact line:\n${newEndMarker}`;
-        }
         return `Old question\nOld response\n${oldEndMarker}`;
       }
-      // Finally, new end marker appears - must have TWO occurrences of new end marker
+      // Finally, new end marker appears from agent
       const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        const newEndMarker = `---RESPONSE-END-${endMatch[1]}---`;
-        // Old markers in scrollback + new instruction (with end marker) + response + agent's end marker
-        return `Old question\nOld response\n${oldEndMarker}\n\nNew question asked\n\nWhen you finish responding, print this exact line:\n${newEndMarker}\nNew response\n\n${newEndMarker}`;
+      const nonceMatch = sent.match(INSTRUCTION_NONCE_REGEX);
+      if (nonceMatch) {
+        const newEndMarker = `---RESPONSE-END-${nonceMatch[1]}---`;
+        // Old markers in scrollback + new response + agent's end marker
+        return `Old question\nOld response\n${oldEndMarker}\nNew response\n\n${newEndMarker}`;
       }
       return `Old question\nOld response\n${oldEndMarker}`;
     };
@@ -951,10 +950,12 @@ describe('cmdTalk - nonce collision handling', () => {
 
     const output = ui.jsonOutput[0] as Record<string, unknown>;
     expect(output.status).toBe('completed');
-    // Response should be from the new markers, not triggered by old markers
-    expect(output.response as string).not.toContain('Old response');
-    expect(output.response as string).not.toContain('Old question');
+    // The key behavior: old markers with different nonce don't trigger completion
+    // We waited for the NEW marker with correct nonce before completing
+    // Note: With new protocol, response includes N lines before marker (may include scrollback)
     expect(output.response as string).toContain('New response');
+    // Verify we polled multiple times (waiting for correct marker, not triggered by old one)
+    expect(captureCount).toBeGreaterThan(2);
   });
 });
 
@@ -977,10 +978,9 @@ describe('cmdTalk - JSON output contract', () => {
 
     tmux.capture = () => {
       const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        // Must have TWO end markers: one in instruction, one from "agent"
-        return mockCompleteResponse(endMatch[1], 'Response');
+      const nonceMatch = sent.match(INSTRUCTION_NONCE_REGEX);
+      if (nonceMatch) {
+        return mockCompleteResponse(nonceMatch[1], 'Response');
       }
       return '';
     };
@@ -1013,9 +1013,11 @@ describe('cmdTalk - JSON output contract', () => {
   });
 
   // Helper moved to describe scope for JSON output tests
+  // Include instruction line for proper extraction anchoring
   function mockCompleteResponse(nonce: string, response: string): string {
+    const instruction = `When you finish responding, output a completion marker on its own line: three dashes, RESPONSE-END-${nonce}, three dashes (no spaces).`;
     const endMarker = `---RESPONSE-END-${nonce}---`;
-    return `Hello\n\nWhen you finish responding, print this exact line:\n${endMarker}\n${response}\n${endMarker}`;
+    return `Some scrollback\n${instruction}\n${response}\n${endMarker}`;
   }
 
   it('includes required fields in timeout response', async () => {
@@ -1054,20 +1056,14 @@ describe('cmdTalk - JSON output contract', () => {
     expect(output).toHaveProperty('endMarker');
   });
 
-  it('captures partialResponse on timeout when agent started responding', async () => {
+  it('captures partialResponse on timeout even when no marker visible', async () => {
     const tmux = createMockTmux();
     const ui = createMockUI();
 
-    // Simulate agent started responding but didn't finish (only ONE end marker in instruction, no second from agent)
+    // Agent is writing but hasn't printed any marker yet
+    // New behavior: we capture the last N lines as partial response
     tmux.capture = () => {
-      const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        const endMarker = `---RESPONSE-END-${endMatch[1]}---`;
-        // Only one end marker (in instruction), agent started writing but didn't finish
-        return `Hello\n\nWhen you finish responding, print this exact line:\n${endMarker}\nThis is partial content\nStill writing...`;
-      }
-      return 'random content';
+      return `This is partial content\nStill writing...`;
     };
 
     const ctx = createContext({
@@ -1093,16 +1089,17 @@ describe('cmdTalk - JSON output contract', () => {
 
     const output = ui.jsonOutput[0] as Record<string, unknown>;
     expect(output).toHaveProperty('status', 'timeout');
-    expect(output).toHaveProperty('partialResponse');
+    // Fallback: capture last N lines as partial response
     expect(output.partialResponse).toContain('This is partial content');
     expect(output.partialResponse).toContain('Still writing...');
   });
 
-  it('returns null partialResponse when nothing captured', async () => {
+  it('returns scrollback as partialResponse when no instruction visible', async () => {
     const tmux = createMockTmux();
     const ui = createMockUI();
 
-    // Nothing meaningful in the capture
+    // Capture shows scrollback but no instruction marker
+    // Fallback returns last N lines
     tmux.capture = () => 'random scrollback content';
 
     const ctx = createContext({
@@ -1128,31 +1125,30 @@ describe('cmdTalk - JSON output contract', () => {
 
     const output = ui.jsonOutput[0] as Record<string, unknown>;
     expect(output).toHaveProperty('status', 'timeout');
-    expect(output.partialResponse).toBeNull();
+    // Fallback captures last N lines even without instruction visible
+    expect(output.partialResponse).toBe('random scrollback content');
   });
 
-  it('captures partialResponse in broadcast timeout', async () => {
+  it('handles broadcast with mixed completion and timeout', async () => {
     const tmux = createMockTmux();
     const ui = createMockUI();
     const markersByPane: Record<string, string> = {};
 
     tmux.send = (pane: string, msg: string) => {
-      const match = msg.match(END_MARKER_REGEX);
+      const match = msg.match(INSTRUCTION_NONCE_REGEX);
       if (match) markersByPane[pane] = match[1];
     };
 
-    // codex completes, gemini times out with partial response
+    // codex completes with end marker, gemini has no end marker (still typing)
     tmux.capture = (pane: string) => {
       if (pane === '10.1') {
         const nonce = markersByPane['10.1'];
         const endMarker = `---RESPONSE-END-${nonce}---`;
-        // Complete response: two end markers
-        return `Msg\n\nWhen you finish responding, print this exact line:\n${endMarker}\nResponse\n${endMarker}`;
+        // Complete response with end marker (only ONE marker in new protocol)
+        return `Response\n${endMarker}`;
       }
-      // gemini has partial response - only one end marker (in instruction)
-      const nonce = markersByPane['10.2'];
-      const endMarker = `---RESPONSE-END-${nonce}---`;
-      return `Msg\n\nWhen you finish responding, print this exact line:\n${endMarker}\nPartial gemini output...`;
+      // gemini has no end marker at all - agent is still responding
+      return `Gemini is still typing this response and hasn't finished yet...`;
     };
 
     const paths = createTestPaths(testDir);
@@ -1160,10 +1156,10 @@ describe('cmdTalk - JSON output contract', () => {
       ui,
       tmux,
       paths,
-      flags: { wait: true, timeout: 0.1, json: true },
+      flags: { wait: true, timeout: 0.5, json: true },
       config: {
         defaults: {
-          timeout: 0.1,
+          timeout: 0.5,
           pollInterval: 0.02,
           captureLines: 100,
           preambleEvery: 3,
@@ -1178,7 +1174,7 @@ describe('cmdTalk - JSON output contract', () => {
     try {
       await cmdTalk(ctx, 'all', 'Hello');
     } catch {
-      // Expected timeout exit
+      // Expected timeout exit for gemini
     }
 
     const result = ui.jsonOutput[0] as {
@@ -1186,17 +1182,20 @@ describe('cmdTalk - JSON output contract', () => {
         agent: string;
         status: string;
         response?: string;
-        partialResponse?: string;
+        partialResponse?: string | null;
       }>;
     };
     const codexResult = result.results.find((r) => r.agent === 'codex');
     const geminiResult = result.results.find((r) => r.agent === 'gemini');
 
+    // Codex should complete (has end marker, output stable)
     expect(codexResult?.status).toBe('completed');
     expect(codexResult?.response).toContain('Response');
 
+    // Gemini times out (no end marker in output)
     expect(geminiResult?.status).toBe('timeout');
-    expect(geminiResult?.partialResponse).toContain('Partial gemini output');
+    // Fallback captures the output even without marker
+    expect(geminiResult?.partialResponse).toContain('Gemini is still typing');
   });
 });
 
@@ -1218,22 +1217,24 @@ describe('cmdTalk - end marker detection', () => {
   });
 
   // Helper: generate mock capture output with proper marker structure
-  // The end marker must appear TWICE: once in instruction, once from "agent"
+  // Include instruction line for proper extraction anchoring
   function mockResponse(nonce: string, response: string): string {
+    const instruction = `When you finish responding, output a completion marker on its own line: three dashes, RESPONSE-END-${nonce}, three dashes (no spaces).`;
     const endMarker = `---RESPONSE-END-${nonce}---`;
-    return `Message\n\nWhen you finish responding, print this exact line:\n${endMarker}\n${response}\n${endMarker}`;
+    return `Some scrollback\n${instruction}\n${response}\n${endMarker}`;
   }
 
-  it('includes end marker in sent message', async () => {
+  it('includes end marker instruction in sent message (not literal marker)', async () => {
     const tmux = createMockTmux();
     const ui = createMockUI();
 
     // Return complete response immediately
     tmux.capture = () => {
       const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        return mockResponse(endMatch[1], 'Response');
+      // Extract nonce from instruction (looks for RESPONSE-END-xxxx pattern)
+      const nonceMatch = sent.match(/RESPONSE-END-([a-f0-9]+)/);
+      if (nonceMatch) {
+        return mockResponse(nonceMatch[1], 'Response');
       }
       return '';
     };
@@ -1249,21 +1250,24 @@ describe('cmdTalk - end marker detection', () => {
     await cmdTalk(ctx, 'claude', 'Test message');
 
     const sent = tmux.sends[0].message;
-    expect(sent).toMatch(/---RESPONSE-END-[a-f0-9]+---/);
-    expect(sent).toContain('When you finish responding, print this exact line:');
+    // New protocol: instruction describes marker verbally, doesn't contain literal marker
+    expect(sent).not.toMatch(/---RESPONSE-END-[a-f0-9]+---/);
+    expect(sent).toContain('output a completion marker on its own line');
+    expect(sent).toContain('three dashes, RESPONSE-END-');
   });
 
-  it('extracts response between two end markers', async () => {
+  it('extracts response before end marker', async () => {
     const tmux = createMockTmux();
     const ui = createMockUI();
 
     tmux.capture = () => {
       const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        const endMarker = `---RESPONSE-END-${endMatch[1]}---`;
-        // Simulate scrollback with old content, instruction, response, and agent's end marker
-        return `Old garbage\nMore old stuff\nMessage\n\nWhen you finish responding, print this exact line:\n${endMarker}\nThis is the actual response\n\n${endMarker}\nContent after marker`;
+      // Extract nonce from instruction
+      const nonceMatch = sent.match(/RESPONSE-END-([a-f0-9]+)/);
+      if (nonceMatch) {
+        const endMarker = `---RESPONSE-END-${nonceMatch[1]}---`;
+        // Simulate scrollback with old content, then agent's response with marker
+        return `Old garbage\nMore old stuff\nThis is the actual response\n\n${endMarker}\nContent after marker`;
       }
       return 'Old garbage\nMore old stuff';
     };
@@ -1294,9 +1298,10 @@ Line 4 final`;
 
     tmux.capture = () => {
       const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        return mockResponse(endMatch[1], multilineResponse);
+      // Extract nonce from instruction
+      const nonceMatch = sent.match(/RESPONSE-END-([a-f0-9]+)/);
+      if (nonceMatch) {
+        return mockResponse(nonceMatch[1], multilineResponse);
       }
       return '';
     };
@@ -1316,17 +1321,18 @@ Line 4 final`;
     expect(output.response).toContain('Line 4 final');
   });
 
-  it('handles empty response between markers', async () => {
+  it('handles empty response before marker', async () => {
     const tmux = createMockTmux();
     const ui = createMockUI();
 
     tmux.capture = () => {
       const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        const endMarker = `---RESPONSE-END-${endMatch[1]}---`;
-        // Agent printed end marker immediately with no content
-        return `Message here\n\nWhen you finish responding, print this exact line:\n${endMarker}\n${endMarker}`;
+      // Extract nonce from instruction
+      const nonceMatch = sent.match(/RESPONSE-END-([a-f0-9]+)/);
+      if (nonceMatch) {
+        const endMarker = `---RESPONSE-END-${nonceMatch[1]}---`;
+        // Agent printed end marker immediately with no content before it
+        return `${endMarker}`;
       }
       return '';
     };
@@ -1346,7 +1352,7 @@ Line 4 final`;
     expect(typeof output.response).toBe('string');
   });
 
-  it('waits until second marker appears (not triggered by instruction alone)', async () => {
+  it('waits until marker appears (not triggered while agent is thinking)', async () => {
     const tmux = createMockTmux();
     const ui = createMockUI();
 
@@ -1354,15 +1360,16 @@ Line 4 final`;
     tmux.capture = () => {
       captureCount++;
       const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        const endMarker = `---RESPONSE-END-${endMatch[1]}---`;
+      // Extract nonce from instruction
+      const nonceMatch = sent.match(/RESPONSE-END-([a-f0-9]+)/);
+      if (nonceMatch) {
+        const endMarker = `---RESPONSE-END-${nonceMatch[1]}---`;
         if (captureCount < 3) {
-          // Only ONE end marker (in instruction) - should keep waiting
-          return `Message\n\nWhen you finish responding, print this exact line:\n${endMarker}\nAgent is still thinking...`;
+          // No marker yet - agent is still thinking
+          return `Agent is still thinking...`;
         }
-        // Finally, agent prints second marker
-        return `Message\n\nWhen you finish responding, print this exact line:\n${endMarker}\nActual response\n${endMarker}`;
+        // Finally, agent prints marker
+        return `Actual response\n${endMarker}`;
       }
       return '';
     };
@@ -1384,7 +1391,7 @@ Line 4 final`;
     expect(output.response).toContain('Actual response');
   });
 
-  it('handles large scrollback with markers at edges', async () => {
+  it('handles large scrollback with marker at end', async () => {
     const tmux = createMockTmux();
     const ui = createMockUI();
 
@@ -1393,11 +1400,12 @@ Line 4 final`;
 
     tmux.capture = () => {
       const sent = tmux.sends[0]?.message || '';
-      const endMatch = sent.match(END_MARKER_REGEX);
-      if (endMatch) {
-        const endMarker = `---RESPONSE-END-${endMatch[1]}---`;
-        // TWO end markers: one in instruction, one from "agent" response
-        return `${lotsOfContent}\nMessage\n\nWhen you finish responding, print this exact line:\n${endMarker}\n\nThe actual response\n\n${endMarker}`;
+      // Extract nonce from instruction
+      const nonceMatch = sent.match(/RESPONSE-END-([a-f0-9]+)/);
+      if (nonceMatch) {
+        const endMarker = `---RESPONSE-END-${nonceMatch[1]}---`;
+        // ONE marker only - from agent response
+        return `${lotsOfContent}\nThe actual response\n\n${endMarker}`;
       }
       return lotsOfContent;
     };
@@ -1415,6 +1423,5 @@ Line 4 final`;
     const output = ui.jsonOutput[0] as Record<string, unknown>;
     expect(output.status).toBe('completed');
     expect(output.response).toContain('actual response');
-    expect(output.response).not.toContain('Line 0');
   });
 });

package/src/commands/talk.ts

@@ -49,6 +49,26 @@ function makeEndMarker(nonce: string): string {
   return `---RESPONSE-END-${nonce}---`;
 }
 
+/**
+ * Build a regex to match the end marker case-insensitively.
+ * This handles agents that might print the marker in different case.
+ */
+function makeEndMarkerRegex(nonce: string): RegExp {
+  return new RegExp(`---response-end-${nonce}---`, 'i');
+}
+
+/**
+ * Build the end marker instruction WITHOUT embedding the literal marker string.
+ * This prevents false-positive detection when the instruction is still visible
+ * in scrollback but the agent hasn't responded yet.
+ *
+ * The instruction describes how to construct the marker verbally, so the literal
+ * marker string can ONLY appear if the agent actually prints it.
+ */
+function makeEndMarkerInstruction(nonce: string): string {
+  return `When you finish responding, output a completion marker on its own line: three dashes, RESPONSE-END-${nonce}, three dashes (no spaces).`;
+}
+
 function renderWaitLine(agent: string, elapsedSeconds: number): string {
   const s = Math.max(0, Math.floor(elapsedSeconds));
   return `⏳ Waiting for ${agent}... (${s}s)`;
@@ -57,7 +77,10 @@ function renderWaitLine(agent: string, elapsedSeconds: number): string {
 /**
  * Extract partial response from output when end marker is not found.
  * Used to capture whatever the agent wrote before timeout.
- * Looks for first occurrence of end marker (in our instruction) and extracts content after it.
+ *
+ * With the new protocol, the instruction doesn't contain the literal marker.
+ * We look for the instruction line (contains "RESPONSE-END-<nonce>" without dashes)
+ * and extract content after it. Falls back to last N lines if instruction not found.
  */
 function extractPartialResponse(
   output: string,
@@ -65,12 +88,30 @@ function extractPartialResponse(
   maxLines: number
 ): string | null {
   const lines = output.split('\n');
-  const firstMarkerLineIndex = lines.findIndex((line) => line.includes(endMarker));
 
-  if (firstMarkerLineIndex === -1) return null;
+  // Extract nonce from endMarker (format: ---RESPONSE-END-xxxx---)
+  // Use case-insensitive match to be flexible with nonce format changes
+  const nonceMatch = endMarker.match(/RESPONSE-END-([a-f0-9]+)/i);
+  if (!nonceMatch) return null;
+  const nonce = nonceMatch[1];
+
+  // Find the instruction line (contains "RESPONSE-END-<nonce>" but not the full marker)
+  // Case-insensitive to handle potential format variations
+  const instructionLineIndex = lines.findIndex(
+    (line) =>
+      line.toLowerCase().includes(`response-end-${nonce.toLowerCase()}`) &&
+      !line.includes(endMarker)
+  );
+
+  let responseLines: string[];
+  if (instructionLineIndex !== -1) {
+    // Extract lines after instruction
+    responseLines = lines.slice(instructionLineIndex + 1);
+  } else {
+    // Fallback: just take the output (no instruction found in view)
+    responseLines = lines;
+  }
 
-  // Get lines after our instruction's end marker
-  const responseLines = lines.slice(firstMarkerLineIndex + 1);
   const limitedLines = responseLines.slice(-maxLines); // Take last N lines
 
   const partial = limitedLines.join('\n').trim();
@@ -92,6 +133,11 @@ interface AgentWaitState {
   partialResponse?: string | null;
   error?: string;
   elapsedMs?: number;
+  // Per-agent timing
+  startedAtMs: number;
+  // Debounce tracking (per-agent)
+  lastOutput: string;
+  lastOutputChangeAt: number;
 }
 
 interface BroadcastWaitResult {
@@ -260,8 +306,9 @@ export async function cmdTalk(ctx: Context, target: string, message: string): Pr
   const endMarker = makeEndMarker(nonce);
 
   // Build message with preamble and end marker instruction
+  // Note: instruction doesn't contain literal marker to prevent false-positive detection
   const messageWithPreamble = buildMessage(message, target, ctx);
-  const fullMessage = `${messageWithPreamble}\n\nWhen you finish responding, print this exact line:\n${endMarker}`;
+  const fullMessage = `${messageWithPreamble}\n\n${makeEndMarkerInstruction(nonce)}`;
 
   // Best-effort cleanup and soft-lock warning
   const state = cleanupState(ctx.paths, 60 * 60); // 1 hour TTL
@@ -279,8 +326,10 @@ export async function cmdTalk(ctx: Context, target: string, message: string): Pr
   const isTTY = process.stdout.isTTY && !flags.json;
 
   // Debounce detection: wait for output to stabilize
-  const MIN_WAIT_MS = 3000; // Wait at least 3 seconds before detecting completion
-  const IDLE_THRESHOLD_MS = 3000; // Content unchanged for 3 seconds = complete
+  // Adaptive: for very short timeouts (testing), reduce debounce thresholds
+  const timeoutMs = timeoutSeconds * 1000;
+  const MIN_WAIT_MS = Math.min(3000, timeoutMs * 0.3); // Wait at least 3s or 30% of timeout
+  const IDLE_THRESHOLD_MS = Math.min(3000, timeoutMs * 0.3); // Stable for 3s or 30% of timeout
   let lastOutput = '';
   let lastOutputChangeAt = Date.now();
 
@@ -352,7 +401,8 @@ export async function cmdTalk(ctx: Context, target: string, message: string): Pr
       if (!flags.json) {
         if (isTTY) {
           process.stdout.write('\r' + renderWaitLine(target, elapsedSeconds));
-        } else {
+        } else if (flags.verbose || flags.debug) {
+          // Non-TTY progress logs only with --verbose or --debug
           const now = Date.now();
           if (now - lastNonTtyLogAt >= 30000) {
             lastNonTtyLogAt = now;
@@ -412,8 +462,9 @@ export async function cmdTalk(ctx: Context, target: string, message: string): Pr
       const elapsedMs = Date.now() - startedAt;
       const idleMs = Date.now() - lastOutputChangeAt;
 
-      // Find end marker
-      const hasEndMarker = output.includes(endMarker);
+      // Find end marker (case-insensitive to handle agent variations)
+      const endMarkerRegex = makeEndMarkerRegex(nonce);
+      const hasEndMarker = endMarkerRegex.test(output);
 
       // Completion conditions:
       // 1. Must wait at least MIN_WAIT_MS
@@ -437,9 +488,10 @@ export async function cmdTalk(ctx: Context, target: string, message: string): Pr
       const lines = output.split('\n');
 
       // Find the line with the end marker (last occurrence = agent's marker)
+      // Find end marker line (case-insensitive)
       let endMarkerLineIndex = -1;
       for (let i = lines.length - 1; i >= 0; i--) {
-        if (lines[i].includes(endMarker)) {
+        if (endMarkerRegex.test(lines[i])) {
           endMarkerLineIndex = i;
           break;
         }
@@ -447,11 +499,25 @@ export async function cmdTalk(ctx: Context, target: string, message: string): Pr
 
       if (endMarkerLineIndex === -1) continue;
 
-      // Find where response starts (after instruction's end marker, if visible)
-      const firstMarkerLineIndex = lines.findIndex((line) => line.includes(endMarker));
-      let startLine = firstMarkerLineIndex + 1;
-      // Limit to N lines before end marker
-      startLine = Math.max(startLine, endMarkerLineIndex - responseLines);
+      // Protocol: instruction describes the marker verbally but doesn't contain the literal string.
+      // So any occurrence of the literal marker is definitively from the agent.
+      //
+      // Try to anchor extraction to the instruction line (cleaner output when visible).
+      // Fall back to N lines before marker if instruction scrolled off.
+      let startLine: number;
+      const instructionLineIndex = lines.findIndex(
+        (line) =>
+          line.toLowerCase().includes(`response-end-${nonce.toLowerCase()}`) &&
+          !endMarkerRegex.test(line)
+      );
+
+      if (instructionLineIndex !== -1 && instructionLineIndex < endMarkerLineIndex) {
+        // Instruction visible: extract from after instruction to marker
+        startLine = instructionLineIndex + 1;
+      } else {
+        // Instruction scrolled off: extract N lines before marker
+        startLine = Math.max(0, endMarkerLineIndex - responseLines);
+      }
 
       let response = lines.slice(startLine, endMarkerLineIndex).join('\n').trim();
       // Clean Gemini CLI UI artifacts
@@ -500,6 +566,12 @@ async function cmdTalkAllWait(
   const pollIntervalSeconds = Math.max(0.1, config.defaults.pollInterval);
   const captureLines = config.defaults.captureLines;
 
+  // Debounce detection constants (same as single-agent mode)
+  // Adaptive: for very short timeouts (testing), reduce debounce thresholds
+  const timeoutMs = timeoutSeconds * 1000;
+  const MIN_WAIT_MS = Math.min(3000, timeoutMs * 0.3); // Wait at least 3s or 30% of timeout
+  const IDLE_THRESHOLD_MS = Math.min(3000, timeoutMs * 0.3); // Stable for 3s or 30% of timeout
+
   // Best-effort state cleanup
   cleanupState(paths, 60 * 60);
 
@@ -519,17 +591,19 @@ async function cmdTalkAllWait(
     const endMarker = makeEndMarker(nonce);
 
     // Build and send message with end marker instruction
+    // Note: instruction doesn't contain literal marker to prevent false-positive detection
     const messageWithPreamble = buildMessage(message, name, ctx);
-    const fullMessage = `${messageWithPreamble}\n\nWhen you finish responding, print this exact line:\n${endMarker}`;
+    const fullMessage = `${messageWithPreamble}\n\n${makeEndMarkerInstruction(nonce)}`;
     const msg = name === 'gemini' ? fullMessage.replace(/!/g, '') : fullMessage;
 
     try {
+      const now = Date.now();
       tmux.send(data.pane, msg);
       setActiveRequest(paths, name, {
         id: requestId,
         nonce,
         pane: data.pane,
-        startedAtMs: Date.now(),
+        startedAtMs: now,
       });
       agentStates.push({
         agent: name,
@@ -538,11 +612,17 @@ async function cmdTalkAllWait(
         nonce,
         endMarker,
         status: 'pending',
+        // Per-agent timing
+        startedAtMs: now,
+        // Initialize debounce tracking
+        lastOutput: '',
+        lastOutputChangeAt: now,
       });
       if (!flags.json) {
         console.log(`  ${colors.green('→')} Sent to ${colors.cyan(name)} (${data.pane})`);
       }
     } catch {
+      const now = Date.now();
       agentStates.push({
         agent: name,
         pane: data.pane,
@@ -551,6 +631,9 @@ async function cmdTalkAllWait(
         endMarker,
         status: 'error',
         error: `Failed to send to pane ${data.pane}`,
+        startedAtMs: now,
+        lastOutput: '',
+        lastOutputChangeAt: now,
       });
       if (!flags.json) {
         ui.warn(`Failed to send to ${name}`);
@@ -568,7 +651,6 @@ async function cmdTalkAllWait(
     return;
   }
 
-  const startedAt = Date.now();
   let lastLogAt = 0;
   const isTTY = process.stdout.isTTY && !flags.json;
 
@@ -591,34 +673,33 @@ async function cmdTalkAllWait(
   try {
     // Phase 2: Poll all agents in parallel until all complete or timeout
     while (pendingAgents().length > 0) {
-      const elapsedSeconds = (Date.now() - startedAt) / 1000;
-
-      // Check timeout for each pending agent (#17)
+      // Check timeout for each pending agent using per-agent timing
       for (const state of pendingAgents()) {
-        if (elapsedSeconds >= timeoutSeconds) {
+        const agentElapsedMs = Date.now() - state.startedAtMs;
+        const agentElapsedSeconds = agentElapsedMs / 1000;
+
+        if (agentElapsedSeconds >= timeoutSeconds) {
           state.status = 'timeout';
-          state.error = `Timed out after ${Math.floor(timeoutSeconds)}s`;
-          state.elapsedMs = Math.floor(elapsedSeconds * 1000);
+          state.error = `Timed out after ${Math.floor(agentElapsedSeconds)}s`;
+          state.elapsedMs = agentElapsedMs;
 
           // Capture partial response on timeout
           const responseLines = flags.lines ?? 100;
           try {
             const output = tmux.capture(state.pane, captureLines);
-            console.log('debug>>', output);
             const extracted = extractPartialResponse(output, state.endMarker, responseLines);
             if (extracted) {
               state.partialResponse =
                 state.agent === 'gemini' ? cleanGeminiResponse(extracted) : extracted;
             }
-          } catch (err) {
-            console.error(err);
+          } catch {
             // Ignore capture errors on timeout
           }
 
           clearActiveRequest(paths, state.agent, state.requestId);
           if (!flags.json) {
             console.log(
-              `  ${colors.red('✗')} ${colors.cyan(state.agent)} timed out (${Math.floor(elapsedSeconds)}s)`
+              `  ${colors.red('✗')} ${colors.cyan(state.agent)} timed out (${Math.floor(agentElapsedSeconds)}s)`
             );
           }
         }
@@ -627,16 +708,18 @@ async function cmdTalkAllWait(
       // All done?
       if (pendingAgents().length === 0) break;
 
-      // Progress logging (non-TTY)
-      if (!flags.json && !isTTY) {
+      // Progress logging (non-TTY, only with --verbose or --debug)
+      if (!flags.json && !isTTY && (flags.verbose || flags.debug)) {
         const now = Date.now();
         if (now - lastLogAt >= 30000) {
           lastLogAt = now;
           const pending = pendingAgents()
             .map((s) => s.agent)
             .join(', ');
+          // Use the oldest pending agent's elapsed time for logging
+          const maxElapsed = Math.max(...pendingAgents().map((s) => now - s.startedAtMs));
           console.error(
-            `[tmux-team] Waiting for: ${pending} (${Math.floor(elapsedSeconds)}s elapsed)`
+            `[tmux-team] Waiting for: ${pending} (${Math.floor(maxElapsed / 1000)}s elapsed)`
           );
         }
       }
@@ -651,7 +734,7 @@ async function cmdTalkAllWait(
         } catch {
           state.status = 'error';
           state.error = `Failed to capture pane ${state.pane}`;
-          state.elapsedMs = Date.now() - startedAt;
+          state.elapsedMs = Date.now() - state.startedAtMs;
           clearActiveRequest(paths, state.agent, state.requestId);
           if (!flags.json) {
             ui.warn(`Failed to capture ${state.agent}`);
@@ -659,27 +742,37 @@ async function cmdTalkAllWait(
           continue;
         }
 
-        // Find end marker
-        const firstEndMarkerIndex = output.indexOf(state.endMarker);
-        const lastEndMarkerIndex = output.lastIndexOf(state.endMarker);
-
-        if (firstEndMarkerIndex === -1) continue;
-
-        // Check if marker is from agent (not just in our instruction)
-        const afterMarker = output.slice(lastEndMarkerIndex + state.endMarker.length);
-        const followedByUI = afterMarker.includes('╭') || afterMarker.includes('context left');
-        const twoMarkers = firstEndMarkerIndex !== lastEndMarkerIndex;
+        // Track output changes for debounce detection (per-agent)
+        if (output !== state.lastOutput) {
+          state.lastOutput = output;
+          state.lastOutputChangeAt = Date.now();
+        }
 
-        if (!twoMarkers && !followedByUI) continue;
+        // Use per-agent timing for accurate elapsed calculation
+        const now = Date.now();
+        const elapsedMs = now - state.startedAtMs;
+        const idleMs = now - state.lastOutputChangeAt;
+
+        // Find end marker (case-insensitive to handle agent variations)
+        const endMarkerRegex = makeEndMarkerRegex(state.nonce);
+        const hasEndMarker = endMarkerRegex.test(output);
+
+        // Completion conditions (same as single-agent mode):
+        // 1. Must wait at least MIN_WAIT_MS
+        // 2. Must have end marker in output
+        // 3. Output must be stable for IDLE_THRESHOLD_MS (debounce)
+        if (elapsedMs < MIN_WAIT_MS || !hasEndMarker || idleMs < IDLE_THRESHOLD_MS) {
+          continue;
+        }
 
         // Extract response: get N lines before the agent's end marker
         const responseLines = flags.lines ?? 100;
         const lines = output.split('\n');
 
-        // Find the line with the agent's end marker (last occurrence)
+        // Find end marker line (case-insensitive)
         let endMarkerLineIndex = -1;
         for (let i = lines.length - 1; i >= 0; i--) {
-          if (lines[i].includes(state.endMarker)) {
+          if (endMarkerRegex.test(lines[i])) {
             endMarkerLineIndex = i;
             break;
           }
@@ -687,13 +780,25 @@ async function cmdTalkAllWait(
 
         if (endMarkerLineIndex === -1) continue;
 
-        // Determine where response starts
-        let startLine = 0;
-        if (twoMarkers) {
-          const firstMarkerLineIndex = lines.findIndex((line) => line.includes(state.endMarker));
-          startLine = firstMarkerLineIndex + 1;
+        // Protocol: instruction describes the marker verbally but doesn't contain the literal string.
+        // So any occurrence of the literal marker is definitively from the agent.
+        //
+        // Try to anchor extraction to the instruction line (cleaner output when visible).
+        // Fall back to N lines before marker if instruction scrolled off.
+        let startLine: number;
+        const instructionLineIndex = lines.findIndex(
+          (line) =>
+            line.toLowerCase().includes(`response-end-${state.nonce.toLowerCase()}`) &&
+            !endMarkerRegex.test(line)
+        );
+
+        if (instructionLineIndex !== -1 && instructionLineIndex < endMarkerLineIndex) {
+          // Instruction visible: extract from after instruction to marker
+          startLine = instructionLineIndex + 1;
+        } else {
+          // Instruction scrolled off: extract N lines before marker
+          startLine = Math.max(0, endMarkerLineIndex - responseLines);
         }
-        startLine = Math.max(startLine, endMarkerLineIndex - responseLines);
 
         let response = lines.slice(startLine, endMarkerLineIndex).join('\n').trim();
         // Clean Gemini CLI UI artifacts
@@ -702,7 +807,7 @@ async function cmdTalkAllWait(
         }
         state.response = response;
         state.status = 'completed';
-        state.elapsedMs = Date.now() - startedAt;
+        state.elapsedMs = elapsedMs;
         clearActiveRequest(paths, state.agent, state.requestId);
 
         if (!flags.json) {