Why Claude Code Keeps Stopping [Fix]

Claude Code is not crashing. It is doing what you asked—working on your code. But every file read, grep search, and command output dumps tokens into your context. After 30-50 tool calls, the window fills up, and Claude has to stop. The tokens are not going to your prompts or Claude's responses. They are going to timestamps, blank lines, repeated file paths, and decorative formatting that Claude reads but never uses.

What it looks like when you hit the limit

• ✓Claude Code stops mid-sentence and says "I can't continue"
• ✓Responses become very slow or time out
• ✓You get a message about usage limits or rate limiting
• ✓The session becomes unresponsive for 30+ seconds between messages
• ✓Claude cannot read files or run commands, even simple ones
• ✓You have to start a completely new session to continue working

Why does this happen so fast?

A single file read of a 500-line file returns 12,000-15,000 tokens. A grep with 30 matches returns 8,000-12,000 tokens. A build log with errors returns 10,000-20,000 tokens. Do that 5-10 times and you have consumed 50,000-150,000 tokens—and 80% of it is noise: line numbers, file paths, timestamps, blank lines, and status messages that Claude never needed.

Your conversation history (your prompts + Claude's responses) might only be 10,000-20,000 tokens. But the tool outputs are 50,000-100,000 tokens. The ratio of signal to noise is roughly 1:3 to 1:5. And that is on a 1 million token context window. You hit the limit not because you wrote that much code, but because tool outputs are bloated.

Fix 1: Token Limits proxy (the permanent fix)

Install Token Limits to strip noise before tool outputs reach your context. This is the permanent fix—one install, compression happens on every request, you stop hitting limits.

1. npm install -g token-limits
2. token-limits start (or add to startup scripts)
3. Claude Code: cmd+comma → Tools → API URL → http://localhost:4800
4. Done. Compression happens automatically on all future tool calls.

Fix 2: /compact command (immediate relief)

If you are already mid-session and Claude Code is about to stop, type /compact in your chat. This command rewrites your conversation history in compressed form, shrinking it by 40-60% and freeing up space to keep working.

/compact is a one-time fix for the current session. It works, but you will hit the limit again in another 1-3 hours unless you install the proxy.

Fix 3: Start new sessions earlier

If you can't install the proxy yet, start a new Claude Code chat every 60-90 minutes instead of waiting for it to crash. The new session gets a fresh context window. This is not ideal (you lose previous context), but it keeps you working.

Real example: Why stopping happens

You start a session with 1M token context window. You paste a file (3k tokens), ask Claude to refactor it (2k tokens), Claude responds (5k tokens). You are at 10k. Then you ask Claude to search the codebase (15k tokens for results), review a second file (12k tokens), run a build (20k tokens), grep for a pattern (18k tokens), review git history (16k tokens). You are now at 108k tokens consumed, but your actual conversation is maybe 20k of meaningful discussion. The other 88k is tool output noise. At 150k, Claude starts slowing down. At 200k+, it becomes unresponsive. You have to start over.

With Token Limits, that same session compresses the tool outputs from 88k down to 10k. Now you can do 8-10x more work before hitting the limit.