[BUG] Context Loss Bug in v2.1.3/2.1.4

Found 2 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/17453
2. https://github.com/anthropics/claude-code/issues/15076

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

This is NOT a duplicate!

vs #17453:

• #17453: Terminal UI becomes slow/laggy during typing and execution
• This issue: Complete memory corruption - ALL context data shows 0 tokens
• Different failure modes entirely

vs #15076: API returns 500 errors after cumulative context threshold - this is a SERVER-side error

This issue: LOCAL memory corruption where Claude Code loses all context data in memory BEFORE making API calls

Evidence:

• #15076: API error 500, request_id: null (server fails)
• This: /context shows 0 tokens locally (client memory loss)
• Different symptoms, different failure points, different versions

#15076 is API/server issue. This is Claude Code client memory bug.

This bug is unique:
Complete context data loss in Claude Code's memory management. /context shows ALL files with 0 tokens:

• All CLAUDE.md files: 0 tokens
• All agents: 0 tokens
• All rules: 0 tokens

This is memory corruption in version 2.1.3/2.1.4, NOT performance degradation or API errors.

Please keep open - this needs separate investigation.

Hey @konian71
The same is still in 2.1.5. There is small detail that I never overflow the context -> its not needed to be compacted because I am not even reaching 150k/200k when my issues appears. The rest looks similar so I would keep your issue as a root one.

My personal input for the bug context:

What's Wrong?

During normal usage, after sending a message (especially with higher token counts ~70-80k+), Claude Code displays a "connection failed" error suggesting internet issues (internet is working fine). After pressing Esc to cancel, running /context shows the token count dropped from ~79k to ~2k - the entire conversation context is lost.

Debug logs show:

1. Error streaming, falling back to non-streaming mode: Connection error. occurs
2. Suspicious Summarizing all X messages (~0 tokens) entries appear - suggesting context is already empty when summarization triggers
3. After user aborts with Esc: Error in non-streaming fallback: Request was aborted.
4. Context is wiped instead of preserved

What Should Happen?

When a connection error occurs and the user cancels, the existing conversation context should be preserved. The error recovery path should not wipe the context.

Error Messages/Logs

Example from debug log (ffd1311b-d886-4026-9784-427e6076a6b9.txt):
  2026-01-11T17:18:59.766Z [DEBUG] Summarizing all 5 messages (~0 tokens)
  2026-01-11T17:19:00.439Z [ERROR] Error streaming, falling back to non-streaming mode: Connection error.
  2026-01-11T17:19:02.262Z [DEBUG] Summarizing all 6 messages (~0 tokens)
  2026-01-11T17:19:28.064Z [ERROR] Error streaming, falling back to non-streaming mode: Connection error.
  2026-01-11T17:19:42.879Z [DEBUG] Streaming aborted by user: Request was aborted.
  2026-01-11T17:19:42.879Z [ERROR] Error in non-streaming fallback: Request was aborted.

❯ /context 
  ⎿   Context Usage
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   claude-opus-4-5-20251101 · 86k/200k tokens (43%)
     ⛀ ⛀ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   ⛁ System prompt: 2.9k tokens (1.4%)
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   ⛁ System tools: 16.5k tokens (8.3%)
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   ⛁ MCP tools: 2.1k tokens (1.1%)
     ⛁ ⛁ ⛁ ⛀ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ Custom agents: 911 tokens (0.5%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ Memory files: 373 tokens (0.2%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ Skills: 1.9k tokens (1.0%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛝ ⛝ ⛝   ⛁ Messages: 61.3k tokens (30.7%)
     ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝   ⛶ Free space: 69k (34.5%)
     ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝   ⛝ Autocompact buffer: 45.0k tokens (22.5%)

❯ HERE I had a message for 35k tokens. (200 > 35 + 86) ! 



❯ /context 
  ⎿   Context Usage
     ⛁ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   claude-opus-4-5-20251101 · 2k/200k tokens (1%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ Skills: 1.9k tokens (1.0%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛶ Free space: 153k (76.5%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛝ Autocompact buffer: 45.0k tokens (22.5%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛝ ⛝ ⛝
     ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝
     ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝

     MCP tools · /mcp
     └ mcp__voicemode__converse: 0 tokens
     └ mcp__voicemode__service: 0 tokens
     └ mcp__hetzner__exec: 0 tokens
     └ mcp__hetzner__sudo-exec: 0 tokens

     Custom agents · /agents

     User
     └ docsmaster: 0 tokens
     └ session-developer: 0 tokens

     Memory files · /memory
     └ ~/.claude/CLAUDE.md: 0 tokens


❯ hey, you here ?                                                                                                                                                                                                  
● Yes, I'm here. Context was just compacted.                                                                                                                                                    
  
  Looking at that deployment log you shared - it actually succeeded. The log shows:                                                                                                               - Build completed successfully
  - All containers started                                                                                                    
  - "New container started" message at the end 

  The "Deployment is Failed" header might be stale UI or a different deployment. The es0g0o00oco8cg0w8oo848cw deployment we triggered earlier finished successfully.

  What would you like to do next?

❯ what are we doing?                                                                                                                                                                              ⎿  Request timed out. Check your internet connection and proxy settings
     Retrying in 9 seconds… (attempt 5/10)                                                                                                                                                      

Steps to Reproduce

1. Have a conversation until context reaches ~70-80k tokens
2. Send another message (just a normal flow, nothing specific or complex - normal every day prompts.)
3. Wait for "connection failed" error to appear
4. Press Esc to cancel
5. Run /context - observe token count dropped to ~2k

Note: Intermittent - seems related to WSL2 networking or API timeouts

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

I Don't know when EXACTLY this started. 1 month ago is the safest bet, but maybe even in January

Claude Code Version

2.1.5

Platform

Anthropic API

Operating System

Other Linux

Terminal/Shell

WSL (Windows Subsystem for Linux)

Additional Information

• Issue has been occurring for several days
• Autocompact never triggers normally before this happens
• The "0 tokens" summarization attempts suggest context is cleared during error recovery before user sees it

Hi @KindEmily

Thanks for the detailed debug logs and for confirming this is still happening in 2.1.5!

Your observation about the "Summarizing all X messages (~0 tokens)" entries is really interesting - it proves the context is already wiped BEFORE the summarization even runs. That's a key detail for the devs.

I downgraded to 2.0.58 and have been running stable ever since - no more context loss, no more random connection errors. Not ideal, but at least a working workaround until Anthropic fixes this.

Which version are you currently using? Did you also downgrade or are you still on 2.1.5?

Update: v2.1.12 seems to have fixed this issue. Can confirm it's working stable now.

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.