[BUG] Claude Desktop Drops/Reverts Long Multi-Step Tasks Without Warning

I am seeing this behavior as well. I am also using Desktop Commander MCP and in this case Shopify Dev MCP. Desktop commander worked fine by itself previously, but last night this kept happening to me.

Also affected by this, also with Desktop Commander, prompt starts fine ... but after some time the whole prompt and its processing is vanished as never existed.

I am seeing this behavior when I have asked Claude to summarize a 14mb very large PDF for me - 594 pages of text. It begins the work and goes for a while, and then the sudden reversion as if I never hit enter on the prompt attached and no history at all - totally gone. This is using the Claud Code desktop app with Opus 4.5.

ONE NOTE ON BUG FIX: I ran the same prompt again but then had a new prompt in the box to remind it to save its work as it went along but never sent it. However, the second time running the exact same prompt it completed. Wondering if having something ready in the prompt box somehow prevented it from reverting since it would have had to erase what was there??? _Is this simple thing the workaround for this bug?_

I noticed this problem when testing some of my own MCP server(s). It works fine at first and then eventually nukes the whole chat history. I don't have this issue with VsCode with Github Copilot hooked up the same MCP server(s).

This happens very frequently with ANY MCP server use. Unfortunately Opus 4.5 is only available on Desktop with MCP so it's the only place I can use it for practical tasks, and it crashes out like this about 1/3 of turns that are complex.

I've tried with text in the Reply window and it doesn't seem to make any difference (Ref above comment).

This is extremely impactful and VERY EXPENSIVE to lose 60+ turns of Opus doing something complex and then just dump out to the last turn. It's particularly devastating on the FIRST turn as claude is particularly enthusiastic about context gathering in this span.

Claude Code does NOT have this issue, but does not have all models available or I would simply never use the Desktop client again, it's just too buggy.

It not only costs money, it also drives me crazy... it's always so irritating when I come back to Claude Desktop and think: Oh no, I forgot to send that command prompt, I was pretty sure I had sent it, okay, I'll do it now and get back to my other tasks meanwhile ... repeat this ;-)

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.

im still having this issue

Adding data from Linux (CachyOS/Arch) with Claude Desktop 0.14.10 and Desktop Commander MCP 0.2.28.

Log analysis from ~/.config/Claude/logs/claude.ai-web.log:

| Error | Count |
|-------|-------|
| Failed to invalidate conversation tree | 13 |
| Request timeout | 18 |
| Tool result is too large. Maximum size is 1MB | 5 |
| Internal server error | 4 |

The Failed to invalidate conversation tree error appears to be the smoking gun - it occurs when the app fails to sync conversation state after certain errors, causing the silent rollback.

Triggers observed:

• Large MCP tool responses (approaching 1MB limit)
• Request timeouts during response generation
• Extended conversations with many MCP tool calls

Workarounds that reduce frequency (but do not eliminate):

• Keep Claude Desktop in focus during generation
• Avoid large file reads / command outputs
• Start fresh conversations before context gets too large

This issue has also been reported in community discussions for 6+ months.

Cross-platform confirmation: Windows (OP), macOS (other comments), and now Linux all affected.

Related issue I filed before finding this one: #16937

Additional Diagnostic Data from Windows Claude Desktop App

I'm experiencing the same issue on Windows 11 with Claude Desktop v1.1.381. After investigating the logs, I found the root causes:

Log Evidence

Location: C:\Users\<user>\AppData\Roaming\Claude\logs\

1. Tool Result Size Limit (claude.ai-web.log)

2026-01-18 11:00:04 [error] [REACT_QUERY_CLIENT] QueryClient error: Error: Tool result is too large. Maximum size is 1MB.

2. Request Timeout + Message Store Sync Loss (claude.ai-web.log)

2026-01-18 19:33:02 [error] [COMPLETION] Request failed Error: Request timeout
2026-01-18 19:33:02 [error] [COMPLETION] Not retryable error, throwing
2026-01-18 19:33:02 [error] Uncaught (in promise) Error: Request timeout
2026-01-18 19:33:02 [warn] [COMPLETION] message_store_sync_loss [object Object]
2026-01-18 19:33:05 [error] [COMPLETION] Failed all attempts to invalidate conversation tree with consistency=eventual [object Object]

3. MCP Server Unexpected Closes (main.log)

2026-01-18 11:23:15 [info] [EventLogging] Sending event: desktop_mcp_unexpected_close

Multiple instances of this around the same time as the resets.

Reproduction Pattern

The reset occurs when:

1. MCP tool returns result > 1MB (e.g., verbose pytest output, large file reads)
2. Long-running operations (like running 100+ async client tests) cause server timeout
3. The message_store_sync_loss event triggers UI revert to previous state

Environment

• Windows 11 x64
• Claude Desktop v1.1.381
• Multiple MCP servers: Desktop Commander, Filesystem, memory, GitHub, etc.
• Triggers: pytest with large output, file operations on large files

Workaround

• Run long operations in terminal directly instead of through Claude
• Use -v instead of -vv for pytest
• Limit file read sizes with offset/length parameters

The key insight is that message_store_sync_loss is what causes the typed input to reappear - the app reverts to the last synced state when the completion request fails.

Additional data point from Windows user

Environment:

• Platform: Claude Desktop (Windows 11)
• MCP: Desktop Commander
• Use case: Long coding sessions (WordPress plugin development)

Log Analysis (%APPDATA%\Claude\logs\claude.ai-web.log):

| Error Type | Occurrences |
|------------|-------------|
| Request timeout | 1,050 |
| message_store_sync_loss | 822 |
| Failed to invalidate conversation tree | 611 |

Confirmed error cascade pattern:

1. [COMPLETION] Request timeout / Internal server error
2. [COMPLETION] message_store_sync_loss
3. Failed to invalidate conversation tree
4. Session rollback (5-50 messages lost)

Impact:

• Happens multiple times per day during active development
• Loses 5-50 messages per incident (not just current turn)
• More frequent during US evening hours (server load?)
• Desktop Commander MCP makes it worse (more tool calls = higher chance)

Workaround that helps:
Writing progress to .md file after EVERY tool call as "breadcrumbs" — so when rollback happens, can recover from file state. Essentially manual checkpointing.

Request:
Please prioritize this. The silent data loss with no error message makes Claude Desktop unreliable for any serious development work.

+1 experiencing the same issue

I'm seeing this exact behavior on Claude Desktop with MCP tool calls. During a multi-step MCP session, Claude was actively working on a project.

It was making good progress , but when I came back to check on it, the entire turn reverted without warning. When I returned to the conversation, Claude had no memory of any of that work happening. From its perspective, those messages were never sent.

Key details:

Platform: Claude Desktop (macOS / Windows) ← edit to match your OS
Task involved: ~20+ MCP tool calls across multiple steps
The project work itself was saved (my MCP persists to disk), but all conversation context was lost
Screenshot attached showing the work that was reverted

This isn't a context window issue — the conversation was relatively short. Something is causing mid-turn state loss during complex tool call chains.

<img width="775" height="1339" alt="Image" src="https://github.com/user-attachments/assets/fbffdc1c-6fd4-4fb6-bfae-4b17e424df2f" />

Same issue

Same issue with Claude Desktop when using Opus 4.6, to the point that I canceled my Claude Pro subscription.

What I did to solve the problem, when it happens, have Claude read the
problematic session EXHAUSTIVELY to continue where he left off and divide
the work into smaller chunks and have him make a plan then save plan as a
MD(markdown) file. Remember the name and have him tell you where it is.
When its done have him spin up multiple agents to do the tasks or you can
have him do it one at a time depending on the subscription you have. Beware
multiple agents consume tokens way faster.

Le dim. 15 févr. 2026, 14 h 57, Baljak @.***> a écrit :

Baljak left a comment (anthropics/claude-code#10065) <https://github.com/anthropics/claude-code/issues/10065#issuecomment-3905082186> Same issue with Claude Desktop when using Opus 4.6, to the point that I canceled my Claude Pro subscription. — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/10065#issuecomment-3905082186>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/B46HU2NFW3DT2EC6WYVWIJL4MDFSJAVCNFSM6AAAAACJ2KPWCOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTSMBVGA4DEMJYGY> . You are receiving this because you commented.Message ID: @.***>

Still happening as of March 2026. Claude Code 2.1.63, macOS (Parallels), also reproduces on claude.ai and Claude Desktop.

This needs engineering attention. The bug is significantly worse in tool-heavy agentic sessions (bash, git, file creation, web search). The response completes, side effects execute (commits, file writes), then the response silently rolls back. No error. No recovery.

The dangerous part: git commits and file changes persist after rollback, but the user has no record of what was done. This leaves orphaned commits, unpushed changes, and files in unknown states. Each incident takes ~20 minutes to audit and recover from.

Frequency: 100+ times per week across sustained daily use. Workarounds (shorter sessions, more frequent restarts) reduce but do not eliminate it. Trigger appears to be high context usage near the compaction threshold.

This has been reported across #14225, #16937, #18676, and community/discussions#155163. Issues keep getting closed. Can we get confirmation this is tracked and an ETA on a fix?

Still happens on March 15, 2026. Very annoying. I have to click "continue" multiple times for one complete response. I have a feeling this also consumes tokens each time I tell it to continue. This only happened to me today, never had this issue before. I am on MAX subscription and it's frustrating that such an expensive subscription has these huge limitations.

Note that the response does not cancel out mid-sentence for me. Claude seems to pick up where it left off. But it is very tedious nonetheless.

Work, time and money is lost - and the trust in this great tool impaired - please please please escalate this bug, working with the claude is usually a pleasure!

!Image
I am seeing the exact behaviour as already noted by other user:
"In tool-heavy agentic sessions (bash, git, file creation, web search). The response completes, side effects execute (commits, file writes), then the response silently rolls back. No error. No recovery.

The dangerous part: git commits and file changes persist after rollback, but the user has no record of what was done. This leaves orphaned commits, unpushed changes, and files in unknown states.

Frequency: 100+ times per week across sustained daily use. Workarounds (shorter sessions, more frequent restarts) reduce but do not eliminate it. Trigger appears to be high context usage near the compaction threshold."

I'm am on Mac, on Max plan, doing app MVP development with distributed responsibilities/roles between Claude.ai and claude code with fairly detailed Claude.md (SOLID, TDD, DDD, Clean Code, development workflow) handoff documents and heavy conceptual work.

why is this issue inactive? this is a real problem and a real major bug

Confirming still a major issue.

I am running the latest public build - Claude 1.2278.0 (e5213f) 2026-04-13T17:49:21.000Z

And it is consistently dropping and rolling back chats, especially when MPC use is part of the task. It has rendered Claude effectively unusable in this mode. Max plan subscriber, not asking huge tasks, just basic MCP usage (Shopify MCP, Paper MCP).

If mythos is going to change the world why is all your software a buggy mess?

Confirming this major issue still exists.

During a Cowork session, with several days worth of work and tokens used, I went to settings to update my payment method. When I came back to the chat, it had lost all context and gave this as the output for a question I asked.

<img width="726" height="326" alt="Image" src="https://github.com/user-attachments/assets/43930bd7-ca58-4d9e-9f85-32e29828e993" />

Closing for now — inactive for too long. Please open a new issue if this is still relevant.

Still relevant as of this morning.