CRITICAL: "Interrupted by user" immediately on ANY input - macOS 15.6.1 - Makes Claude Code 100% unusable

Additional Technical Details for Issue #7261

Version Information (successfully retrieved):

• Claude Code: 1.0.108 (Native installer)
• Node: v20.19.4
• npm: 10.8.2
• macOS: 15.6.1 (Build 24G90)
• Architecture: Apple Silicon M-series (showing x86_64 in terminal, likely Rosetta 2)
• Shell: /bin/zsh

Complete Process Cleanup Performed:

# All Claude-related processes killed
pkill -f claude
pkill -f pavisus
killall node

# All LaunchAgents disabled and moved
launchctl unload ~/Library/LaunchAgents/com.claude.*
launchctl unload ~/Library/LaunchAgents/com.pavisus.*
mv ~/Library/LaunchAgents/*.claude.*.plist ~/Library/LaunchAgents/DISABLED/

Confirmed Clean State:

• ps aux | grep claude returns nothing
• launchctl list | grep claude shows no active agents
• No automation scripts running
• Fresh terminal sessions tested

Additional Testing Performed:

1. Different shells: zsh, bash, bash --noprofile --norc
2. Different terminals: Terminal.app, VS Code integrated terminal, iTerm2
3. Different directories: /tmp, ~/, project directories
4. With/without flags: Both claude and claude --dangerously-skip-permissions
5. Other REPL tools work fine: python3, node, irb all accept input normally

Timing Characteristics:

• Interruption occurs within ~50-100ms of pressing Enter
• No processing happens at all before interruption
• Even single character input like "a" triggers it
• Built-in commands like /help, /status, /model all fail identically

What This Is NOT:

• ❌ Not an authentication issue (no auth prompt appears)
• ❌ Not a permission issue (trust dialog works fine)
• ❌ Not a PATH issue (binary found at ~/.local/bin/claude)
• ❌ Not a shell configuration issue (tested in clean bash)
• ❌ Not a LaunchAgent conflict (all disabled)
• ❌ Not a Node.js issue (other Node apps work)

Suspected Root Cause:

This appears to be a signal handling or input stream issue specific to macOS 15.6.1 (Sequoia) on Apple Silicon, possibly related to:

• TTY/PTY handling differences in macOS 15.6.1
• Signal mask inheritance issues
• Input buffer race condition
• Rosetta 2 translation layer conflict (if Claude Code has x86 dependencies)

Impact:

This makes Claude Code 100% non-functional. I cannot execute any command, not even /help or /exit. This is blocking critical business operations.

Available for Testing:

I'm immediately available to:

• Run any diagnostic commands
• Test beta/debug builds
• Provide system logs
• Test proposed fixes
• Share screen if needed

Working with Anthropic support for 2+ hours exhausted all standard troubleshooting steps. This requires engineering investigation.

Please let me know if you need any specific diagnostic output or system information.

Same:

<img width="1102" height="430" alt="Image" src="https://github.com/user-attachments/assets/053de9b0-68f2-49f8-bb77-70a75c535020" />

FWIW, this also happened to me today. Same symptom, no amount of reinstalling etc. made it any better.

However, I ended up trying to remove MCP servers from my project MCP file (.mcp.json), and after removing a particular one, CC worked again (after exiting CC and starting a new session).

I have been using this exact MCP setup for months without issues, so it's odd. But I tried this because a "/doctor" showed me that I had a lot of MCP tools (over 50k tokens worth). However, removing other MCP servers that brought me well below that threshold didn't help - only removing that single MCP server made a difference (Unity-MCP). I don't think that MCP server is to blame though, since I've been using it daily in CC for months without any problems.

I'm suspecting that I was running an MCP tool from that MCP server right before my CC ended up in this broken state, and then interrupted that tool call (pressing "esc") - somehow causing corrupted state somewhere I haven't been able to clean up by deleting the ~/.claude.json file or the ~/.claude folder and uninstalling CC.

Note: I'm running CC in WSL2 (default Ubuntu distro) on Windows 11.

same issue on macOS 15.5 . Also can confirm that removing the MCP servers recover Claude functionalities.

@ebbe-brandstrup

Yes, it related with the mcp parameters definition. For example: one of my mcp tool define the parameters with complex schema like:

/**
 * JSON-serializable value schema
 */
const JsonifiableSchema: z.ZodType<any> = z.lazy(() =>
  z.union([
    z.null(),
    z.boolean(),
    z.number(),
    z.string(),
    z.array(JsonifiableSchema),
    z.record(z.string(), JsonifiableSchema),
  ])
);

After I simplify the `args` definition, this issue fixed. 

/**
 * Execute command input schema
 */
export const ExecuteCommandInputSchema = z.object({
  command: z.string().describe('VSCode command to execute (e.g., \'vscode.open\', \'editor.action.formatDocument\')'),
// this param use recursion and union
  args: z.array(JsonifiableSchema).optional().describe('Optional arguments to pass to the command'),
  saveAllEditors: z.boolean().optional().default(true).describe('Save all dirty editors after executing the command (default: true)'),
}).strict();

check: https://github.com/tjx666/vscode-mcp/commit/c7ae01f2d432ce4298227eb8cce23b4a9cb96987#diff-e46bb510c06569a4274304e8626df23eb9b19475331dcccbeb62412ac6136e18

I ran into the same issue today with new MCP server setup (using the import command from Claude Desktop to Claude Code). Will uninstall some now and hope this fixes the "Interrupted by user" issue.

I suspect it is directly related to the large context size (call /doctor):

Context Usage Warnings
└ ⚠ Large MCP tools context (~146.656 tokens > 25.000)
└ MCP servers:
└ github: 90 tools (~45.641 tokens)
└ mcp-atlassian: 42 tools (~28.764 tokens)
└ atlassian: 24 tools (~16.945 tokens)
└ billing-cost-management-mcp-server: 13 tools (~15.104 tokens)
└ memory: 28 tools (~13.636 tokens)
└ (6 more servers)

PS: Removing a couple and the duplicate "atlassian" and "map-atlassian" MCP servers did the job - I can work again.

PPS: Untill it did not - two prompts later, same issue. Will follow up once I know more.

Update: Works again without any further changes. Hard to reproduce.

✅ SOLVED - Root Cause Found & Fixed!

Thank you all for the incredible detective work on this issue, especially @ebbe-brandstrup whose insights about MCP servers and configuration bloat pointed me in the right direction!

### The Problem
My ~/.claude.json had grown to 4.5MB with an estimated 1.2 million tokens (48x over the safe threshold). The file was storing massive amounts of conversation history and pasted content from 240+ sessions across multiple
projects.

### The Solution
Created two maintenance scripts that completely resolved the issue:

1. Health Monitor - Checks configuration health and warns when approaching limits
2. Deep Clean Tool - Aggressively cleans the configuration while preserving essential settings

Results:

• Config reduced from 4.5MB → 0.04MB (98.5% reduction)
• Token count: 1,190,868 → 10,803
• Claude Code now works perfectly!

### Key Findings

• Safe threshold appears to be <25,000 tokens in the config file
• The issue builds up gradually through normal use (history accumulation)
• Projects with 50+ history entries start contributing to the problem
• Pasted content is stored indefinitely and never cleaned up

### Preventive Measures
Set up an automated LaunchAgent (macOS) that runs hourly health checks to prevent this from happening again. The monitor alerts when approaching dangerous thresholds.

### Recommendations for Others
If you're experiencing this issue:

1. Check your config size: ls -lh ~/.claude.json
2. Estimate tokens: echo "Tokens: $(($(wc -c ~/.claude.json | awk '{print $1}')/4))"
3. If over 25k tokens, backup and clean the file
4. Keep only recent history entries (10-20 per project)
5. Remove all pasted content sections

Thank you again to everyone who contributed to identifying this issue! Special thanks to @tjx666 for identifying the complex schema problem, @lexxx233 for confirming the MCP connection, and @mahdimanesh for documenting the token
threshold warnings.

This was a team effort and your contributions helped solve a critical blocker. The issue isn't with MCP servers themselves but rather with how conversation history accumulates over time. Regular maintenance is the key!

---
macOS 15.6.1 | Claude Code 1.0.108 | Apple Silicon M-series

[like] Mahdi Manesh reacted to your message:
________________________________
From: Daniel Manfred Stoffel @.*>
Sent: Sunday, September 7, 2025 5:54:48 PM
To: anthropics/claude-code @.*>
Cc: Mahdi Manesh @.>; Comment @.>
Subject: Re: [anthropics/claude-code] CRITICAL: "Interrupted by user" immediately on ANY input - macOS 15.6.1 - Makes Claude Code 100% unusable (Issue #7261)

[https://avatars.githubusercontent.com/u/79516989?s=20&v=4]danielmstoffel left a comment (anthropics/claude-code#7261)<https://github.com/anthropics/claude-code/issues/7261#issuecomment-3263938781>
✅ SOLVED - Root Cause Found & Fixed!

Thank you all for the incredible detective work on this issue, especially @ebbe-brandstrup whose insights about MCP servers and configuration bloat pointed me in the right direction!

The Problem

My ~/.claude.json had grown to 4.5MB with an estimated 1.2 million tokens (48x over the safe threshold). The file was storing massive amounts of conversation history and pasted content from 240+ sessions across multiple
projects.

The Solution

Created two maintenance scripts that completely resolved the issue:

1. Health Monitor - Checks configuration health and warns when approaching limits
2. Deep Clean Tool - Aggressively cleans the configuration while preserving essential settings

Results:

• Config reduced from 4.5MB → 0.04MB (98.5% reduction)
• Token count: 1,190,868 → 10,803
• Claude Code now works perfectly!

Key Findings

• Safe threshold appears to be <25,000 tokens in the config file
• The issue builds up gradually through normal use (history accumulation)
• Projects with 50+ history entries start contributing to the problem
• Pasted content is stored indefinitely and never cleaned up

Preventive Measures

Set up an automated LaunchAgent (macOS) that runs hourly health checks to prevent this from happening again. The monitor alerts when approaching dangerous thresholds.

Recommendations for Others

If you're experiencing this issue:

1. Check your config size: ls -lh ~/.claude.json
2. Estimate tokens: echo "Tokens: $(($(wc -c ~/.claude.json | awk '{print $1}')/4))"
3. If over 25k tokens, backup and clean the file
4. Keep only recent history entries (10-20 per project)
5. Remove all pasted content sections

Thank you again to everyone who contributed to identifying this issue! Special thanks to @tjx666 for identifying the complex schema problem, @lexxx233 for confirming the MCP connection, and @mahdimanesh for documenting the token
threshold warnings.

This was a team effort and your contributions helped solve a critical blocker. The issue isn't with MCP servers themselves but rather with how conversation history accumulates over time. Regular maintenance is the key!

---
macOS 15.6.1 | Claude Code 1.0.108 | Apple Silicon M-series

—
Reply to this email directly, view it on GitHub<https://github.com/anthropics/claude-code/issues/7261#issuecomment-3263938781>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/A7B4ZFZRTRVYQ7FPIJ2SCND3RRWORAVCNFSM6AAAAACF2UO3RWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTENRTHEZTQNZYGE>.
You are receiving this because you commented.Message ID: @.***>

Doesnt work for me. I have cleaned the claude.json but still getting the same error.

Ok so for me the issue was Directus.io MCP server. Something is wrong with this MCP

Ok so for me the issue was Directus.io MCP server. Something is wrong with this MCP

This was the same for me. I ran the /mcp command and selected to restart MCP server. Directus failed to connect, but claude became functional again.

I had to remove 2 of the 3 MCPs I had, context7 and drawio. Then Claude started to work again.

Hey, very sorry about this. It looks like this is indeed a regression for MCP servers with really large contexts. I'll aim to get a fix in for the next release.

This should be fixed in the latest version.

@ollie-anthropic
Can you explain the fix please? Is the 25k limit no longer present?

@ollie-anthropic Hey. I encountered the issue two nights ago, and it stopped working for me altogether. After upgrading to the latest CC version today, the issue persisted. At first I thought it was related to the mcp configurations, after clearing all mcp settings, leaving them empty, the issue still persisted. Once I read this thread and looked at the size of the file (mine was 3.4 mb), I used a script to shrink its size and after that it all started working again. The upgrade allowed for one command to run successfully but on the second prompt, the user interrupted behavior started again until shrinking the file per the solution above. Just a heads up. Thanks!

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.