Copilot stucks in infinite loop

[AJIXuMuK]

Describe the bug

If a prompt contains actions for the copilot, it stucks in the infinite loop.
Here are the examples:

copilot --allow-all-tools -p "create nl.json file with the next content: { name: 'abc', description: 'qwerty' }"
I'll create the nl.json file with the specified content.

✓ Create nl.json (+2)

I'll create the nl.json file with the specified content.

✗ Create nl.json
   Path already exists

I'll create the nl.json file with the specified content.

✗ Create nl.json
   Path already exists

I'll create the nl.json file with the specified content.

✗ Create nl.json
   Path already exists

^C

copilot --allow-all-tools -p "create nl.json file with the next content: { name: 'abc', description: 'qwerty' }. Replace the file if already exists"
I'll create the nl.json file with the specified content.

✓ Create nl.json file with specified content
   $ cat > nl.json << 'EOF'
   { "name": "abc", "description": "qwerty" }
   EOF

   ↪ 1 line...

I'll create the nl.json file with the specified content. Since the file already exists (as shown in the directory listing), I'll replace it.

✓ Create nl.json file with the specified content
   $ cat > nl.json << 'EOF'
   {
     "name": "abc",
     "description": "qwerty"
   }...
   ↪ 3 lines...

I'll create the nl.json file with the specified content in the current directory.

✓ Create nl.json file with specified content
   $ cat > "/Users/alexterentiev/Work/ms/Summary evals/ngsp-sx-docs/Seismic/nl.json" << 'EOF'
   { "name": "abc", "description": "qwerty" }
   EOF

   ↪ 1 line...

^C

It did work before.

Affected version

0.0.354 Commit: 076bd172b

Steps to reproduce the behavior

copilot --allow-all-tools -p "create nl.json file with the next content: { name: 'abc', description: 'qwerty' }"

Expected behavior

the copilot creates a file and stops the execution.

Additional context

No response

[EvanBoyle]

@AJIXuMuK what model are you using? I am unable to repro this with sonnet-4.5

[AJIXuMuK]

@EvanBoyle I using claude sonnet 4.5 in VS Code terminal on mac.
Node is v22

I will try to use fresh installation of node tomorrow to see if that helps

[EvanBoyle]

cc @friggeri we might have some divergence between CLI and SDK here.

[AJIXuMuK]

@EvanBoyle - OK, so performed a few more tests.

• On Mac it is stuck no matter what. Tried fresh Node install with v24, reboot, etc.
• It is working on Windows (tried in PowerShell terminal but I don't think it will matter)
• Prompts from Copilot view in VS Code work well. So it's not the underlying model issue seems like

[AJIXuMuK]

@EvanBoyle - seems like it is starting after some number of requests.
So I was able to run my automation on Windows that was like ~2k requests.
Then, when I tried to do that again, it stuck.
Same was on Mac. First it worked but after huge number of requests it stuck too.

[JuanCS-Dev]

🔬 Technical Investigation & Analysis

I've attempted to reproduce and analyze this bug. While I couldn't trigger the infinite loop on Linux + v0.0.358, I have some important findings and hypotheses.

Environment Tested

• OS: Linux Mint / Ubuntu 24.04
• Copilot CLI: 0.0.358 (Commit: f5a8b1e76)
• Node.js: v22.20.0
• Model: Claude Sonnet 4.5

---

Reproduction Attempts

✅ Test 1: Fresh State (Works)

copilot --allow-all-tools -p 'create nl.json with content: {"name": "abc"}'

Result: File created successfully

❌ Test 2: File Already Exists (Cannot Reproduce Loop)

# File already exists from Test 1
copilot --allow-all-tools -p 'create nl.json with content: {"name": "xyz"}'

Expected: Infinite loop (based on issue description)
Actual: LLM adapted intelligently:

1. Tried create → Got "Path already exists" error
2. Read the file to check contents
3. Used edit tool instead to change content
4. Completed successfully

❌ Test 3: Insistent Prompt (Still No Loop)

copilot --allow-all-tools -p 'You MUST use create tool. If it fails, try again.'

Result: LLM deleted existing file first, then created new one. No loop.

---

Root Cause Hypotheses

H1: Model Version Difference ⭐ (Most Likely)

The current Sonnet 4.5 (as of Nov 16, 2025) may be smarter than the version from Nov 11:

• Anthropic frequently updates models server-side
• Newer version exhibits better error recovery
• Older version gets stuck in retry loop without adapting

Evidence: Your issue from Nov 11 vs my tests today (Nov 16)

---

H2: State Accumulation Bug ⭐⭐ (Critical Finding)

I found evidence of session state persistence in ~/.copilot/session-state/:

$ ls -lh ~/.copilot/session-state/
-rw------- 1 user user 960K Nov 16 635d43de...jsonl  # 539 lines!
-rw------- 1 user user 252K Nov 16 890b9093...jsonl

These .jsonl files are event logs of tool calls and responses. Key observations:

1. Sessions can grow to 960KB+ with hundreds of events
2. Format: JSON Lines with tool execution history
3. After ~2k requests, context window may be polluted with repeated errors
4. Hypothesis: Model loses ability to reason about alternatives when context is full of error messages

This aligns with your observation: "seems like it is starting after some number of requests. So I was able to run my automation on Windows that was like ~2k requests. Then, when I tried to do that again, it stuck."

---

H3: CLI vs SDK Divergence

@EvanBoyle mentioned this. Possible issues:

1. CLI has implicit retry logic that SDK doesn't
2. Error feedback format differs between implementations
3. Tool call wrapping in CLI adds retry behavior
4. Context management differs (CLI may not prune old messages)

---

H4: Platform-Specific (macOS)

You mentioned: "On Mac it is stuck no matter what"

Possible causes:

• macOS-specific file system error handling
• Different error message format from filesystem
• VS Code terminal integration differences
• Session state handling differs on macOS

---

Critical Questions for Maintainers

1. Is there a retry mechanism for failed tool calls in the CLI?
2. How is error feedback formatted when sent back to the model?
3. Is there a limit on conversation history length before pruning?
4. Has Sonnet 4.5 been updated server-side between Nov 11-16?
5. Are there macOS-specific code paths for file operations?
6. Does the CLI differ from SDK in error handling?

---

Proposed Fixes

Fix 1: Add Retry Limit for Same Tool

const MAX_SAME_TOOL_RETRIES = 3;

if (failedAttempts[toolName] >= MAX_SAME_TOOL_RETRIES) {
  return terminateWithError(
    "Tool '${toolName}' failed ${MAX_SAME_TOOL_RETRIES} times. " +
    "Please try a different approach."
  );
}

Fix 2: Enhanced Error Messages to Model

if (error.code === "PATH_EXISTS") {
  return {
    error: "File already exists. Cannot create.",
    suggestion: "Use 'edit' tool to modify existing file, or 'bash(rm <file>)' to delete first.",
    available_alternatives: ["edit", "view", "bash"]
  };
}

Fix 3: Context Window Management

if (conversationHistory.length > MAX_CONTEXT_MESSAGES) {
  // Keep system prompt + recent 50 messages
  conversationHistory = [
    systemPrompt,
    ...conversationHistory.slice(-50)
  ];
}

Fix 4: Detect Loop Pattern

function detectToolLoop(history: ToolCall[], windowSize = 5) {
  const recent = history.slice(-windowSize);
  const uniqueTools = new Set(recent.map(c => c.toolName));
  
  // If same tool called 5 times in a row with errors
  if (uniqueTools.size === 1 && recent.every(c => c.failed)) {
    throw new LoopDetectedError(
      "Detected infinite loop: same tool failing repeatedly"
    );
  }
}

---

Test Plan for Verification

If maintainers can provide:

1. Source code access or debug build with logging
2. Reproduction environment details (exact macOS version, VS Code terminal config)
3. Session state from a stuck instance

I can:

1. Analyze retry logic in source
2. Test with state pollution (100+ consecutive creates)
3. Monitor session-state growth
4. Identify exact triggering condition

---

Impact

This is a high-severity bug because:

• ✅ Affects production automation (user's 2k request workflow)
• ✅ Platform-specific (macOS always triggers)
• ✅ State-dependent (appears after extended use)
• ✅ Blocks users indefinitely (requires Ctrl+C)

Workaround for affected users:

1. Clear session state: rm -rf ~/.copilot/session-state/*
2. Restart Copilot after every 100-200 requests
3. Use --model gpt-5 as alternative (if doesn't exhibit bug)

---

Let me know if you need additional testing or analysis! I'm happy to dig deeper with maintainer guidance.