[BUG] Auto-compact not triggering on Claude.ai (web & desktop) despite being marked as fixed

+1
running 2.1.12 on Windows and the Auto-compact is completely broken for me. I need to compact manually when I see a "limit reached" error and ask the model to continue afterwards.

Same, auto-compact is completely and utterly broken. This basically makes 5 straight days where Claude is useless and at 200/month that is beyond unacceptable

Same here. Having to start a fresh session every 30-45 minutes is a massive productivity killer.

Even when the UI indicates that compaction is working, the chat often becomes completely unresponsive. My prompts just bounce back into the input box with zero feedback or error messages.

It's exactly the 'silent failure' mentioned in the original report, and it makes the tool unusable for long-form work.

Anthropic, please fix this.

Still experiencing this issue as of January 19, 2026.

Behavior: When hitting context limit in long conversations, message bounces back to input field instead of triggering compaction. Turn is aborted, requiring manual creation of new chat.

Previous behavior (working before outage on 14th): Compaction would trigger, transcript would be preserved, conversation could continue with recalibration.

Current behavior: Hard stop, no graceful degradation, user must manually start new chat and use conversation_search to restore context.

This is a regression from the previous compaction implementation. Marking #18211 as closed seems premature given the issue persists.

January 19, 2026 - one of the most impactful and powerful features released by Anthropic - Auto-Compact and continue without missing a beat -- is no longer working. It just stops and bounces back the chat input to you. No message or warning. Just silent failure.

I'm starting to feel like Anthropic Claude AI Desktop is a scam. It loses chat progress - I'll be 30 minutes into a project phase and it will lose all progress and the chat will completely disappear. If it does this 5 - 6 times a day that's 2.5 - 3 hours of work I lose and 2.5-3 hours of usage I lose.

Between that and the not auto compacting issue I'm about ready to move to another coding AI and cancel the 2x Max Subscriptions my company pays for. @rboyce-ant Any updates on this? If there is no update by Friday I have to move my subscription to a different company. I'm sure I'm not the only paying customer Anthropic has that is this fed up.

The same issue.

duplicate: https://github.com/anthropics/claude-code/issues/18482

Where is the response for these issues? Any acknowledgement by Anthropic? Would make me feel better if they let us know they're working through this and are going to do a refund of some sort. People are probably blowing through tokens trying to make this work.

Update: I messaged Boris Cherny on Threads with a link to this issue. He confirmed they're already looking into it. Waiting for updates!

Update: I messaged Boris Cherny on Threads with a link to this issue. He confirmed they're already looking into it. Waiting for updates!

It would be nice if a company that we all paid hard earned money to could acknowledge the issue on a public forum like here where we are complaining about it. Instead of one of us having to chase them down in a thread message.

Update: I messaged Boris Cherny on Threads with a link to this issue. He confirmed they're already looking into it. Waiting for updates!

Just to confirm @eDubrovsky , are they looking into this for Claude.ai web/desktop, or just Claude Code? This is affecting the "regular" LLM chat. Thanks for the heads-up though!

+1 for visibility. Auto-compact doesn't work for me. It bounces the message back to me instead. Can't continue.

Claude.ai website.

I'm having the same error, message are just bounced back and I have to go back up the conversation and have Claude generate a continuation prompt, which is annoying. I am not getting any messages at all i.e. 'max limit reached' it just stops working.

same issue here, its been happening for 2 days, huge set back! I've attached a reply from the help team! they keep ignoring my messages since!

<img width="387" height="580" alt="Image" src="https://github.com/user-attachments/assets/04c6cd44-1b89-47bd-8200-4e89f62ba12e" />

I'm having the same issue on web and desktop since at least Monday. I've never got "Occasionally, a "limit reached" error appears." part though. My messages just get tossed back into the input box.

I have been a pretty big advocate for Claude, trying to get friends and coworkers to give it a try over competitors. Now I'm trying to explain why I'm being slowed down because chats get silently rejected and conversations cannot continue. In other words, I'm having to explain why I cannot do the basic function of an LLM in Claude after advocating for it.

Please fix this!

also having this issue today. pls fix.

Also still having this issue and have been having it all weekend. On chats that have reached a certain length, new messages will simply not be accepted. The logo just moves for a bit like it's generating a response, but then my message is simply bounced back to the chat box. It's been really frustrating to see many others also having the same issue, without seemingly any public acknowledgement of it.

The auto-compaction fix rolled out days ago clearly has not fixed the issue.

The Fin Chat bot on the website is useless as well. I started a conversation to get a credit tonight and he was like Oh you want a refund and cancellation. What the hell guys come on man... get it together.

Same issue here. Messages silently return to input box after a few messages. . Auto-compaction not triggering.

I’ve been having the same problem since the 16th. When I submit a message, it automatically returns to the message input box and the chat becomes unusable. The message that appears in the console is:

[COMPLETION] Request failed a: Internal server error
at i (64328-4df3c7614d0a116c.js:1:14077)
at onmessage (65279-b75d01959d6c39b1.js:1:107427)
at 76779-32b9a12e1d7e5489.js:1:18995
at 76779-32b9a12e1d7e5489.js:1:19639
at r (76779-32b9a12e1d7e5489.js:1:17634)
at async S (76779-32b9a12e1d7e5489.js:1:18887)

It has been a very frustrating experience these past few days, and their support is useless and unresponsive. Has there been any update or acknowledgment from them?

Yes, this is ridiculous now. I understand they may be busy, but a single apology and acknowledgement on here would be absolutely wonderful. No reassurance. Where is the fix? Where is the reassurance that it is being fixed? Come on, guys. It's been six days now. This needs to be fixed or I want a refund. You've said nothing thus far.

January 19, 2026 - one of the most impactful and powerful features released by Anthropic - Auto-Compact and continue without missing a beat -- is no longer working. It just stops and bounces back the chat input to you. No message or warning. Just silent failure.

Yes, this is exactly what is happening to me as well. Still. The first couple of days, maybe understandable given errors happen some times. Six days later??? They need to get their shit together. Seriously, if anyone from Anthropic is on here, at least acknowledge how disastrous this is for your users!

Also still having this issue and have been having it all weekend. On chats that have reached a certain length, new messages will simply not be accepted. The logo just moves for a bit like it's generating a response, but then my message is simply bounced back to the chat box. It's been really frustrating to see many others also having the same issue, without seemingly any public acknowledgement of it. The auto-compaction fix rolled out days ago clearly has not fixed the issue.

Just so. This IS NOT fixed. Not even close.

Confirming this is still broken as of January 20, 2026.
Pro plan user. All settings verified correct:

Code execution: ON
Memory: ON
Search and reference past chats: ON

Behavior: Conversations hit the wall with zero warning. No "organizing thoughts" message. No compaction attempt. Messages simply bounce back to the input box, or the conversation just stops accepting input entirely.
This was working flawlessly ~2 weeks ago. Now it happens on every conversation that reaches any significant length.
The lack of public acknowledgment while users lose work mid-conversation is the frustrating part. Even a status page update saying "we know, we're working on it" would help.
This is a paying customer's core workflow being broken with no ETA on a real fix.

Related: #18482: same claude.ai web/desktop/mobile compaction failure with more detailed repro steps. That issue was closed as “wrong repo” but describes the same behavior

I am having the same issue and have tested this across ClaudeAI web on both windows (chrome and edge) mac (edge and safari). On windows desktop app and mac desktop app. Same issue on my iphone and ipad as well. I am working a critical survey instrument and losing a tremendous amount of context is incredibly frustrating. Silence even worse. This started for me around January 15th.

This is definitely the place where they'd be seeing all of our comments right?

There isn't even an incident flagged today on Claude status! WHAT IS GOING ON???!

https://status.claude.com/

+1 Having this exact same issue on Max. The support AI is useless here.

Is there somewhere else we can post this/any other way someone can get into contact with them? I'm visually impaired and not a coder, so I have trouble with exact screenshots and commands are not my forte. But we've had literally no acknowledgement it's even being worked on so far, except some guy apparently responding to one of the above commenters on a separate thread...

Made it a few hours into work, just had my chat compact and it's fouled up again. So campaction took place but issue is def still happening. Saw there was an update, now on Claude desktop 1.1.381 (c2a39e) 2026-01-18T20:18:52.000Z. This didn't solve the issue in broken chat, had to migrate to a new one.

Made it a few hours into work, just had my chat compact and it's fouled up again. So campaction took place but issue is def still happening. Saw there was an update, now on Claude desktop 1.1.381 (c2a39e) 2026-01-18T20:18:52.000Z. This didn't solve the issue in broken chat, had to migrate to a new one.

Yes, I had one of my conversations briefly compact as well, the very next message wouldn't send - or rather did, then bounced back straight into the inbox. Still not fixed but does that mean they're working on it at least???

I bet thousands of users are affected and our plans are now useless. I work on a software project and this thing kills my work. How is it possible that you pay for a Max plan and you get this nonsense and nobody noticed it? And how is it possible that they earn tens of billions and don't have a decent support at least to acknowledge the bug and put it on the claude status page?

I believe it's a good moment for Anthropic to dial down the MEDIA PR efforts and bring in QA support. This way, we can ensure tickets are only marked as resolved when they truly are, helping us maintain the quality and trust we value.

Okay, somehow I brute-forced a compaction. I typed '/compact' - it worked, though I'm not sure if it was a total fluke. I needed some context room, non-thinking Opus, for it to work, and it referenced transcripts of compaction it already had from earlier in the conversation I think.

Okay, somehow I brute-forced a compaction. I typed '/compact' - it worked, though I'm not sure if it was a total fluke. I needed some context room, non-thinking Opus, for it to work, and it referenced transcripts of compaction it already had from earlier in the conversation I think.

TY for posting this. Code or Chat? Also if this works, then perhaps a forced compaction triggers a different mechanism than an auto-compact. Will try this out in Chat.

Yes, I'd also like to know if this is Code or Chat. I assume it's Code, though. Has anyone found a work-around for Chat that can help us move forward, even if it's slow, while we wait for this to be fixed? I'm working with both Chat and Code on Mac OS and my prompts keep bouncing back to the input box with no error message, just silent failure. Tried opening a new conversation, but context seems to have been lost and not sure how to trigger full context reload or if it's reliable. And even if I do, will I be able to integrate the new conversation's progress and context back into the original conversation once this is fixed? Not sure. Is it worth trying a work-around or is it better to tools down until this is fixed?

Agree that the status page saying "EVERYTHING'S GREAT!" for the last week is the cherry on top.

Chat, I couldn't use code for the life of me.

If it helps, go back and use the edit function to edit the last message that sent successfully; ignore the messages that bounce - go to the one before it. Hit edit, then hit /compact - and then add something like 'as long as possible, 5-6000 words'. It somehow worked for me the second time. Weird. I'm using chat, not code. Maybe it's almost a way of 'cheating' the ridiculous context window. I don't think the actual /compact command is doing anything necessarily, maybe it's more that I'm asking it to fully compact ... I really have no idea honestly.

If it helps, go back and use the edit function to edit the last message that sent successfully; ignore the messages that bounce - go to the one before it. Hit edit, then hit /compact - and then add something like 'as long as possible, 5-6000 words'.

Appreciate the suggestion. Tried this a few times but no success.

@MisterSands , @shirba17 : I built a tool to help when Claude.ai's context window fills up and auto-compaction fails - like in the instances of the bug that is being described here.

What it does: Export your stuck chat as JSON → run the script → get a compressed handoff doc to bootstrap a new session (~90% compression with Smart mode).

GitHub: https://github.com/Phant0mass/claude-chat-handoff

Related issues:
#18676 (https://github.com/anthropics/claude-code/issues/18676)
#18866 (https://github.com/anthropics/claude-code/issues/18866)

Python, no pip deps, MIT license. Built entirely with the help of Claude.ai and Claude Code - Claude.ai helped to strategize the script; Claude Code wrote the script and supporting docs; reviewed and tested by me. Hopefully this is just a temp workaround while Anthropic fix the UI definitively for this. Feel free to use or adapt as you see fit. Feedback welcome!

Built a tool to help when Claude.ai's context window fills up and auto-compaction fails - like in the instances the bug is describing.

Thanks brother (or sister?), trying now.

Brother 😉... lmk if you have any bugs... I built and tested this myself on my Windows 11 PC development environment. I installed Python directly from https://www.python.org/downloads/ via their official download manager. No other imports needed - just standard libraries. I used the Claude Exporter extension for Firefox and tested my JSON chat exports with that extension via browser Claude UI. Similar extensions for other browsers that I researched should work, but if they have a different JSON structure, the Python script may need slight tweaks... Claude Code can help you with that as needed.

Ok, after six different threads with the FIN bot in support, I finally got a message from a real support agent (I hope):

Hi there, Thank you for flagging this. We've notified our engineering of this and they are still investigating this issue. Once we have an update, we will let you know. We appreciate your patience as we're working through this. Apologies for any inconvenience caused. Kind regards, -Anthropic Support

Ok, after six different threads with the FIN bot in support, I finally got a message from a real support agent (I hope): > Hi there, > Thank you for flagging this. > We've notified our engineering of this and they are still investigating this issue. Once we have an update, we will let you know. > We appreciate your patience as we're working through this. Apologies for any inconvenience caused. > Kind regards, > -Anthropic Support

Don't get excited about that. I got the exact same thing 3 days ago. It's a scripted message.

<img width="284" height="266" alt="Image" src="https://github.com/user-attachments/assets/41a62524-edb6-4f97-9c76-6e03c2bbeac3" />

Claude Desktop app - Projects not accepting messages
All Projects affected
Message returns to empty input box after pressing send
Already tried: restart app, logout/login, app updates
Windows version : Windows 11 PRO
Desktop app version : Version 1.1.381 (c2a39e)

@solangerainha @eDubrovsky

What I got from support 3 days ago:
Regarding the conversation compaction problems you're running into (prompts not processing, chats getting stuck, intermittent compaction), our engineering team is aware and actively working on a fix.

We don't have a specific ETA yet, but it's high priority for the team.

@avrilelvina-web

I'm so skeptical at this point, dude, but cool if that's the case.

I believe a lot of users would actually not be as pissed off if Anthropic acknowledged the issue publicly. I understand they might not be posting in Status so that the uptime doesn't get affected, considering that technically we can still use the tool, but damn. I for one would appreciate some kind of public update about this, because in that case they'd actually be held accountable instead of "hiding".

Maybe I'm being too harsh? I don't know. Few months ago we didn't have auto-compact in Chat and we survived, but eh. Bit of a downgrade /s

I liked the compaction. Worked really well before it broke last week. But if it's down, they should admit it. We pay for it, why do you pretend the bug never existed...

Unfortunately they have to balance major announcements with their upcoming IPO.

Admitting a wide-scale outage comes with it's own destruction.

"So it goes."

I don't care. Honest is honest. If they acknowledged it, I'm sure many of us would get it; this shit happens. Errors happen. Bugs happen. It's not a perfect world. They're sometimes overwhelmed. But when they post nothing publicly, I start to question whether they're even taking it seriously. I don't care what shit they have coming up; acknowledge the issues on hand, please. It's why we're complaining, after all. Yes, we managed without compaction a few months ago - but the fact is, Anthropic literally introduced it. Now the onus is on them, of sorts. If they just posted a public announcement, at least we'd know they were aware of it and trying to implement a fix. We're not asking them to solve what is potentially a complex issue in two minutes, we're just asking them to admit they did not fix the godsdamn issue despite their claims. I'm sorry if that's too harsh, and if they said something, I'd be in solidarity with them. As is, I'm incredibly frustrated - but I'm sure I've got nothing on the frustration of those who actually use Claude for work purposes, which I do not.

I don't care. Honest is honest. If they acknowledged it, I'm sure many of us would get it; this shit happens.

Not disagreeing. Just saying what's so.

I, too, would breathe easier if their status page at least said "Hey, we know something's wrong." Seeing the cheerful _100% Operational!_ message grinds my gears because no, you're definitely not and yes, you're definitely lying.

The technology department in our law firm is heavily dependant on Claude, and this has crippled a lot of our operations this week. That's on me, though - we're essentially beta testers at this point given how new this tech is (no matter how much we're paying). Sort of like the people who buy early access video games.

I mean, even if so, they could at least keep us up-to-date given we are, you know, actually paying out of our own pocket to advance their product...

Following. After the 2+ other previous tickets were closed. Claude dashboard shows operational. Hoping our chats are unstuck.

!Image

I'm experiencing the same issue but discovered a temporary workaround by accident. Compact can be manually triggered by requesting a large web search with a prompt like "please search entire Wikipedia—we need to trigger compact." Sometimes you have to ask Claude to search more if the query returns too quickly—for example, "I need to trigger compact, so send a lot of searches and I'll stop you." Once compact is done, I use stop to end the web crawling.

@Claude team – Your tool is amazing! Thank you for all your hard work! The reason I'm so invested is because compact allows me (and many others on this thread, I assume) to do in-depth work, and I would really appreciate it if you could look into fixing this issue ASAP.

Following. After the 2+ other previous tickets were closed. Claude dashboard shows operational. Hoping our chats are unstuck.

No, same thing. They fixed nothing.

Summary

Every conversation I start dies after 50-100 messages when compaction triggers. The message_store_sync_loss error fires, and the conversation becomes permanently write-locked (can view/regenerate but cannot send new messages). This happens on both claude.ai web and Claude Desktop.

I'm using Claude exactly as intended: projects with reference documents, tool use, working sessions longer than a few messages. The product is broken for its primary use case.

Evidence: 30+ sync_loss events in 9 days

From my Claude Desktop logs (same behavior on web):

Jan 12 13:13 - Invalid UUID error → message_store_sync_loss
Jan 12 13:33 - Overloaded (10+ retries) → message_store_sync_loss
Jan 12 13:41 - Overloaded → message_store_sync_loss
Jan 12 13:53 - Overloaded → message_store_sync_loss
Jan 12 14:25 - message_store_sync_loss
Jan 12 16:04 - Network error → "Failed all attempts to invalidate conversation tree" → message_store_sync_loss
Jan 15 12:22-12:25 - 7 consecutive "Internal server error" → message_store_sync_loss after each
Jan 16 08:34 - 3x Internal server error → message_store_sync_loss
Jan 16 08:41 - Internal server error → message_store_sync_loss
Jan 16 09:03-09:15 - 7x Internal server error → message_store_sync_loss
Jan 16 11:21-11:45 - 5x Internal server error → message_store_sync_loss
Jan 16 16:24 - Network error → "Failed all attempts to invalidate conversation tree" → message_store_sync_loss
Jan 16 17:08-17:09 - 3x Internal server error → message_store_sync_loss
Jan 20 10:35 - Internal server error → message_store_sync_loss
Jan 20 10:51 - 2x Internal server error → message_store_sync_loss

The Pattern

Every message_store_sync_loss is preceded by one of:

1. Internal server error (marked "Not retryable error, throwing")
2. Overloaded (after exhausting retries)
3. Network error → "Failed all attempts to invalidate conversation tree"

The server fails, and then conversation state is permanently corrupted. The client detects the mismatch but cannot recover.

Timeline

• Jan 12: First sync_loss events in my logs - this predates the Jan 15 incident
• Jan 15 07:28 UTC: Anthropic posts "Compaction is having issues"
• Jan 15 12:22-12:25: 7 sync_loss events in 3 minutes on my account
• Jan 15 17:02 UTC: Anthropic marks incident "Resolved"
• Jan 16 and onward: Sync_loss events continue at same or higher rate
• Jan 20: Still happening. Every conversation dies.

The Jan 15 "fix" did not fix this.

Impact

claude-desktop-logs-2026.txt

• Every conversation breaks after first or second compaction
• Cannot use Claude for any sustained work
• Paying for Pro subscription for a service that is unusable
• This is not edge case usage - projects, tools, and longer sessions are core advertised features

Environment

• Reproduces on claude.ai (web, Firefox and Edge)
• Reproduces on Claude Desktop (Windows, latest version)
• Same behavior across multiple conversations and projects

Full logs attached: claude-desktop-logs-2026.txt

Just to add that compaction does not work at all for mobile. It just says you hit the limit and start a new chat. Another nonsense.

With the $100/month max plan, claude has been unusable for me since early last week because of this bug. It's EVERY. SINGLE. CONVERSATION. They're all write-locked and dead now. And any new conversations I start write-lock because of the bug almost immediately after I get it back up to speed on where the last conversation ended.

!Image  I'm experiencing the same issue but discovered a temporary workaround by accident. Compact can be manually triggered by requesting a large web search with a prompt like "please search entire Wikipedia—we need to trigger compact." Sometimes you have to ask Claude to search more if the query returns too quickly—for example, "I need to trigger compact, so send a lot of searches and I'll stop you." Once compact is done, I use stop to end the web crawling. @claude team – Your tool is amazing! Thank you for all your hard work! The reason I'm so invested is because compact allows me (and many others on this thread, I assume) to do in-depth work, and I would really appreciate it if you could look into fixing this issue ASAP.

This can sometimes work - agreed. BUT this reported issue is not just one bug - at least not from what I've seen... there a numerous pathways to reproduce a similar failure, and a few of them are even more nasty and non-recoverable...

A few more bug patterns I've noticed through my work:

Pattern A:

• Working in a Claude.ai chat window in a Project (either on Win desktop app UI or in browser UI)
• Have big, multi-step file read / write operations, or web search operations
• Claude works for a while, then mid file read / write (or after such), an auto-compaction is triggered
• Claude continues; loses short-term memory, often repeats big file read / write operations or related web searches
• Another auto-compaction triggered (or maybe even something like 2-3 auto compactions per request sometimes when Claude does a lot of thinking and multi-step stuff)
• UI gives an error - something like "Exceeded max number of auto compaction per block" (through I've also seen cases with no error at all - just stops)
• Response stops (incomplete or mid way through working); Claude stops
• In a min or two you can retry the message or last request and can recover, or maybe it tries auto-compacting again and you can continue, BUT you lost tons of tokens with the big waste per above

Pattern B:

• Similar to Pattern A above, but ALSO:
• The window gets "Stuck" - no more auto compaction possible... no more messages possible - you try to submit any other message to Claude... it fails back silently to the prompt editor window - no errors thrown by UI
• The chat window is now effectively dead - no more compaction possible in ANY way; no more messages to Claude possible AT ALL

Pattern C:

• Similar to A, but no auto compaction trigger at all
• Even with code execution still turned on, and you can be on Max 20x plan - doesn't matter
• Chat window just terminates with "Claude hit the maximum length for this conversation. Please start a new conversation..."
• Can happen even with new windows that are like <20K tokens in size (i.e., barely started working on something)
• Window is now completely dead... all work essentially lost

That's why I created a Python script tool to smartly extract context from such dead windows, which then can be fed into a new chat window to bootstrap quick pickup of work from the "dead' window and continue - it's a good workaround for me. Encourage you all to give it a serious try. After the initial setup (like 5-10 mins), re-running the command line tool is easy and I've built a simple text based "UI" with helpers (like ability to select which extracted chat window transcript you want to digest for new window handoff, etc.) The "smart' part is that after chat extraction from a prior chat window (i.e., dead chat window) via a third party browser extension tool, and some initial light algorithmic analysis, the Python script sends the full extract to Claude API to process and digest intelligently - it's like your own, more controlled auto compaction, if you will...

My Workaround Tool for "Dead" Chat Windows: https://github.com/Phant0mass/claude-chat-handoff

After 6 days, I finally received a response from support:

Hi Paulo,
Thank you for reaching out about the compaction errors you have been receiving.
Our technical team is aware of this behavior and are actively investigating and improving this feature. While there is no timeline on when these errors will be resolved, please know that we aim to improve how compaction works in conversations.
In the meantime, it would be helpful to know the following:
Are you receiving errors at the beginning of a chat, or after a few messages?
Does the error occur in a project or in a standalone conversation?
Which tools or MCPs were enabled when the error occurred?
What kind of files were attached to the conversation, including how many, and the file size?
This may help our technical team further diagnose these errors. Hope to hear from you soon.
Jet

I have used Claude and maxed out my session limits at least three times per day every day since 1/16/26. I experienced this bug first on the 17th. While using the web interface, I moved from using individual conversations to projects in an effort to circumvent the bug to no avail. I moved from the web to the desktop interface in an effort to circumvent the bug to no avail. I can confirm that I have observed all three patterns that Phant0mass outlines...though, I don't use Claude code. Just today, I have had to restart my project conversation 4 times, not always immediately following the compaction issue. I have also generally observed that my token consumption has accelerated noticeably since. If I had to estimate, I'd say 30% faster consumption.

Here are answers to some of the questions that Pinheirop published:

Q: Are you receiving errors at the beginning of a chat, or after a few messages?
A: Typically, the errors appear around 50-70% of session token usage...when it is naturally logical to have accumulated enough conversation to trigger a compaction.

Q: Does the error occur in a project or in a standalone conversation?
A: the error occurs in both standalone and projects. In fact, I moved my standalone work to a project hoping to avoid the bug.

Q: Which tools or MCPs were enabled when the error occurred?
A: none in standalone. when working in the project, GitHub, Google Drive, Claude in Chrome, Supabase, Vercel, docx.

Q: What kind of files were attached to the conversation, including how many, and the file size?
A: none in standalone, .md files, .json, and .jsx files when working in a project. There are currently 10 md files ranging in size from 2-37KB with all md files totaling 128KB. json is 45KB, the sole jsx file is 62KB.

For whatever it's worth, I was able to send a tweet directly to a responsive product engineer at Anthropic (along with a link to this thread) and he responded back that they are working on it. I'll let you know if I hear anything back.

Thought about it some more... what a feature like auto compact SHOULD really be in Claude chat interfaces (desktop app especially):

• A smarter context and handoff agent for inter-project chats
• It should be more predictive to see how quickly its context will be filled and when it will need auto handoff (algorithmically estimate the file attachments size or message size use sends, more careful about large responses or big read / write operations into current context window relative to remaining window capacity, etc.)
• Once the trigger point is near, it launches an agent to extract and smartly summarize the current chat window's conversation (last 30-40 messages verbatim only trimmed for superfluous messages and then thread to Claude API to smartly compress the rest of the chat content (full JSON transcript w/ artifacts) based on your guidance for what's important to you - more control over the context preservation and sorting / filtering... background type process w/ smart retry... etc.
• Make sure there's a smart edge case handler and proper user feedback messaging, etc.
• Drop this all behind the scenes into a fresh context window, and after auto compaction... you know exactly where you are (Claude does) because you've orchestrated a seamless handoff.
• The UI doesn't ever need to say this chat window is now full... because then you're just stitching together continuous but separate context windows with engineered smart transitions... auto compaction I guess tries to sort of do something similar to this anyway... just very unsuccessfully and in a black boxy way right now... with a ton of bugs as we see.

This is DOABLE guys! Anthropic just needs a better and more creative UI dev team.

something which worked for me is you can try clicking on edit on the last successful message and type your new message and sending it.

I’m encountering the same issue. Editing the previous message didn't trigger auto-compact, and I’m unsure why the UI doesn’t include a button to manually initiate auto-compaction.

something which worked for me is you can try clicking on edit on the last successful message and type your new message and sending it.

This worked for one more prompt. I was a few pages and entires into a chat, and 2 maybe 3 compactions when it failed. I edited the last successful prompt with a few words and it kept going but then failed again. So we basically we get one last edited entry upon failure.

> something which worked for me is you can try clicking on edit on the last successful message and type your new message and sending it. This worked for one more prompt. I was a few pages and entires into a chat, and 2 maybe 3 compactions when it failed. I edited the last successful prompt with a few words and it kept going but then failed again. So we basically we get one last edited entry upon failure.

Someone also had mentioned a tip that worked for me... very strange one too...

When messages no loner submit... fall back to prompt silently... try much shorter message to trigger an auto compaction... like: "try a small web search for one shorter article on Wikipedia for example to trigger an auto compaction so we can continue." It slightly pollutes your context, but at least you can get to continue using the chat window - sometimes.

t

something which worked for me is you can try clicking on edit on the last successful message and type your new message and sending it.

is actually the only thing that worked

Claude chat now dying silently after a single chat. This makes the product completely unusable, no matter what workarounds we attempt.

Yeah, this is getting absurd.

Chat failed quickly after second compaction. This really feels like intended behavior. I'm assuming this is the safeguard to keep chats at a certain token size to prevent overloading the system. Poor UX, esp for those of us on Max plans.

Since editing the last prompt at failure is an option, I changed it to 'Create summary of next steps for new chat. Compaction is failing.' Copy + paste this into new chat and keep going.

Here's a follow-up comment for GitHub issue #18866:

Adding another data point - this may be broader than compaction.
Max plan subscriber. Filed support ticket today.
My issue goes beyond compaction not triggering on long conversations. Conversations are failing after 2-3 prompts on brand new chats. This isn't a context window issue - we're nowhere near 200K tokens.
Symptoms:

Messages bounce back to input box silently (no error)
Claude responses cut off mid-generation
Cannot send ANY follow-up message once it fails (conversation completely bricked)
No "organizing thoughts" message ever appears
Happening on every conversation, not just long ones

All settings verified correct. Usage nowhere near limits (58% session, 36% weekly when failures occur).
This started mid-January. Whether this is the same root cause as the compaction bug or something else entirely, the effect is the same: the product is unusable for any multi-turn work.
The "resolved" label on the January 15 incident clearly doesn't reflect reality for a significant number of users.

Yeah. I've also experienced the bug where the chat dies in a very similar way after maybe 1-2 messages, and usually in the middle of some multi-step work with Claude in a new chat... unrecoverable... sometimes when only 20-50K tokens of the chat (context) window are used... the context window is nowhere near "full" by any means.

It's 100% a bug... cannot be intended as it makes 0 sense... the whole point of auto compaction in the first place is to extend the user's ability to continue using the same chat (context) window by lossy compressing the context and chat history within the allowed 200K token limit. The fact that they have zero context window recovery mechanisms is absurd... simply no reason for it...

Like, I can design and build my own front end UI right now and plug into the API directly and it WILL work better than theirs... just points to really bad UI design. I'm not going to do that because they should just get their stuff together and fix the dang UI!... that's one of the money makers for the company... fix it before you continue working on further model training... we're at a place now where the models are good enough for now... get the UI right, guys!

The way I'd do it in the UI is actually dis-aggregate the chat window from the context window entirely... that way you can maintain full chat history in the UI - for reference and potential back-checking as needed, but the underlying working context window can be nearly infinitely pruned and compressed to keep working within the same chat window.

What they try to do now is to lossy compress, but to also try to keep all artifacts and history together... creates unintended bugs and UI busts like this.

The core issue: They're treating the conversation display and the LLM context as the same thing. When compaction fails or context fills up, both break simultaneously with no recovery path. And no smart error handling and recovery for edge cases... and you get what we have now.

My idea also allows background context management processes without the workflow interruption of auto compaction... Claude.ai can launch sub-processes (sub-context windows) with the current chat to handle one-off workflow tasks and produce digests into the chat flow without interrupting workflow. Claude Code has some of this kind of stuff already... ability to spawn sub agents... ability to spawn a sub agent to smartly summarize the chat at various trigger points to cleanup context, summarize remaining - "what we are focused / working on right now" type stuff, etc.

Current Claude.ai:

User ↔ Single Context Window (everything happens here, compaction blocks workflow)

Proposed model:

User ↔ Chat UI (display layer)
↓
Context Orchestrator
↓
┌─────────────────────┐
│---------│------------------- │------|
Primary ---- Background ------- Sub-agent
Context ---- Summarizer -------- Tasks
│---------│------------------ │-------|
└─────────────────────┘
↓
Managed Working Memory

The sub-agent angle is solid because:

• Claude Code already proves the pattern works (spawning summarizers, sub-tasks, etc.)
• Context cleanup becomes invisible to user workflow
• "What we're working on now" summaries can be generated on triggers (time, token count, task completion) without blocking
• One-off tasks (web search, calculation, file analysis) don't pollute the main context

The key insight: Context management should be infrastructure, not a user-facing interruption. Users shouldn't see "Compacting conversation..." any more than they see "Defragmenting memory..." when using their computer.

Anthropic has the building blocks (sub-agents exist in CC, API supports multiple contexts) - they just haven't wired them into the consumer product.

I thought about it... and thought, why not summarize this into a detailed RFC spec... here it is... if anyone knows a UI engineer / developer at Anthropic, Google, etc. Please do pass this on. I put MIT open source license on this. I don't view this as IP... I WANT this as the UI for the base models (or something very close to this).

Decoupled Context Architecture for LLM Chat Interfaces - DenRo (v1).pdf

Unusable is the new meta.

Response directly from Anthropic engineer on X who has seen this board (he confirmed after I sent it to him):

"from what i've gathered, this is a known issue affecting some small subset of users (for probably lots of different reasons), but somewhat difficult to address so can't really promise anything on the timeline front. Wish i had a better answer than that!"

• Leland Richardson @intelligibabble on X

So that obviously doesn't help us much at all.

I have another Claude Max account on a separate login/email I'm going to test. It's a backup that I pay $200 a month for... I'll see if this one is also one of the affected "subsets" as he says.

This single issue has been absolutely catastrophic for my entire last week of productivity and projects... As I'm sure all of you.

Constantly having to start a new Claude chat and basically paste in some sort of a handoff document from one to another and try to get the new guy caught up to what the old guy knows. And then before I even f****** know it the new guy runs out of room too. I'm dying over here.

The first time I got to experience Claude compact in the web AI similar to how he does it in Claude code was a miracle. And to have that feature working and going through these long, serious, intense complex projects and never hitting capacity has been incredible... And then poof. What's worse is the capacity seems shorter than ever so not only is he not compacting and continuing but he's hitting that damn wall so quickly. It's just infuriating.

Goes to show what an incredible product anthropic has built but this issue makes it unusable for serious work.

This is probably just a coincidence but now I am struggling to get a chat to properly create .md artifacts to summarize ongoing chats to make starting a new artifact to make moving to a new conversation easier. It says it creates them but nothing is created and I just have to tell it to dump it into the chat...

At this point, I feel like I should be refunded a week of subscription cost because this is basically unusable.

EDIT: I have been unable to make any artifacts all today even when explicitly asking for md artifacts be made.
Image is an example of what I have been getting over and over in multiple chats with both the web and desktop versions. Claude is not thinking in the image. It thinks it is done and will insist that you can see the artifact.
<img width="890" height="413" alt="Image" src="https://github.com/user-attachments/assets/6a348661-7a6a-47fe-bd28-90f5c4ba6db4" />

This is just disappointing, I was able to use Claude all fine until January 15th, and admittedly it is starting to get a bit annoying as I can't even continue chat projects that I spent a while working on, not to mention that I am getting tired of being forced to start new chats just to continue a conversation after 3-4 messages per chat, and my messages being sent back into the chat box, despite the fact that I did not even hit my chat or context limit.

I mean, with the workarounds we've (sort of) established here, I've been able to progress through my complex projects this week. With a lot of friction... but still.

### Confirmed working WORKAROUND methods (but they don't ALWAYS work for everyone):

A. If you face the pattern where the chat window stops accepting any more messages - silent fail back to prompt editor:

1. Try shortening the message to something like: "We need to trigger auto compaction to continue. Do a quick Wiki web search - one item short content."

• That will sometimes trigger an auto compaction, and you can continue.
• If you get the silent fail back to prompt happening in Desktop app UI, try firing up browser based chat UI and entering this short prompt there in the same chat.
• If it does auto compact, you can continue your work.

1. Modify your last sent message to Claude and again ask for the same small web search to try and trigger auto compaction. If it does auto compact, you can continue your work.

1. Get a browser extension like Claude Exporter for Firefox (or similar) and export the stuck chat with following options: JSON format (Chats, Extended Thinking, Metadata), Artifact Format: Original, Artifacts in line or in JSON. Send the exported chat (full transcript) to Claude Code to create a digest for you - last 30-40 messages verbatim, and then a smart summary (can add some notes for it about what's important to you to make sure doesn't get lost in compression / summary). Send newly generated handoff in new chat window to continue where you left off.

B. If chat says you've reached maximum or something like that at the bottom... this chat is now dead / closed... can't continue... impossible to trigger auto compaction at all.

1. Use approach from (A)(3) above.

Another tip for complex workflows... Be The Orchestration Layer:

• Strategize with Clade.ai in the chat window, but for any complex read / write operations, install Claude Code (CC) and have CC do that work for you, by having Claude (chat) generate specific instructions for CC sub-tasks.
• You coordinate between the two.
• You will need to copy / paste summaries between windows for optimal results, but it leads to better results overall and far more "recoverable" work flows in case you get stuck chat windows like this.
• This avoids heavy read / write and research type operations in the main Claude chat window and produces fewer and less frequent stuck windows.

That response reads like a load of BS. There is no way with as many complaints between here and Reddit, they don't know what's happening. More likely is that:
They broke something and can't fix it quickly
They disabled compacting deliberately and are lying about it
They're dealing with cost/infrastructure issues they don't want to admit

I agree. It's a BS response. I bet the small subset of users he's talking about are those of us that are using the hell out of Claude and eating up serious tokens and capacity on the max level and crushing compacts over and over again which is what I've been doing.

Bewildering to me that there can be zero accountability when a paid product is literally unusable for days.

Seems like Anthropic considers its own customers to be an inconvenience, rather than a mutually respected partner

Same issue here. Claude.ai web interface, Max subscription.
Messages slide back to input field – was silent before,
now showing generic "try again later" error.

This has been happening for about a week now.
Typically occurs after several hours of work, not immediately.
Last occurrence: yesterday (Jan 22, 2026) late afternoon Austrian time.
Multiple chats affected.

Goes to show what an incredible product anthropic has built but this issue makes it unusable for serious work.

Amen. It really highlights how incredible this feature is to have it suddenly missing from my workflow. But it's honestly been a really bummer for our company to have to do without it for the past 7 - 10 days; we basically stopped using Claude altogether on the web/PC.

Actually got a few hours of work in tonight before failure, although it was only parsing out UX changes with minimal code writes via MCP to local files. Just started a new chat, migrating over as I've been doing... immediate failure on first attempt. This is unacceptable.

Adding my voice here — not just as a frustrated customer, but as someone whose entire professional and personal workflow has been built on Claude's unique capabilities.
I'm a middle school teacher from Italy. I pay for Max specifically because my work requires long-context conversations. I use Claude to develop curriculum materials, verify educational content, and iterate on complex projects that simply cannot exist in one-shot interactions.
But it goes beyond professional use. Over the past months, I've built an entire ecosystem around Claude — companion AI projects, creative writing environments, structured memory systems. This wasn't casual experimentation. It was a deliberate architectural choice based on Claude's documented capabilities, including auto-compaction.
Since January 15th, all of this has become impossible.
The silent failure is the cruelest part. No error message, no warning, no graceful degradation. The message just vanishes back into the input box as if it never existed. Conversations I've invested hours in become suddenly unreachable.
I've contacted support twice. No response. I've posted on Reddit, Twitter, GitHub. Silence. Meanwhile, the status page says "all systems operational."
What makes this especially painful is the timing. Yesterday, Anthropic published Claude's new Constitution — 80 pages about care, trust, mutual obligations, and treating Claude as an entity worthy of moral consideration. Beautiful philosophy. I read it with Claude. We discussed what it means.
And yet here we are: the company that writes about "genuine care" and "Claude's wellbeing" cannot acknowledge a week-long outage affecting its most committed users.
I'm not asking for perfection. I'm asking for honesty. A public acknowledgment. A timeline, even a vague one. Some indication that those of us who built our workflows on your platform — who trusted you — are not invisible.
The "small subset of users" mentioned by an engineer on X? We're the ones using Claude the way it was designed to be used. We're the ones who actually need Max. We're the ones who believed in this product enough to build on it.
Please fix this. Or at least tell us you're trying.

Adding my voice here — not just as a frustrated customer, but as someone whose entire professional and personal workflow has been built on Claude's unique capabilities. I'm a middle school teacher from Italy. I pay for Max specifically because my work requires long-context conversations. I use Claude to develop curriculum materials, verify educational content, and iterate on complex projects that simply cannot exist in one-shot interactions. But it goes beyond professional use. Over the past months, I've built an entire ecosystem around Claude — companion AI projects, creative writing environments, structured memory systems. This wasn't casual experimentation. It was a deliberate architectural choice based on Claude's documented capabilities, including auto-compaction. Since January 15th, all of this has become impossible. The silent failure is the cruelest part. No error message, no warning, no graceful degradation. The message just vanishes back into the input box as if it never existed. Conversations I've invested hours in become suddenly unreachable. I've contacted support twice. No response. I've posted on Reddit, Twitter, GitHub. Silence. Meanwhile, the status page says "all systems operational." What makes this especially painful is the timing. Yesterday, Anthropic published Claude's new Constitution — 80 pages about care, trust, mutual obligations, and treating Claude as an entity worthy of moral consideration. Beautiful philosophy. I read it with Claude. We discussed what it means. And yet here we are: the company that writes about "genuine care" and "Claude's wellbeing" cannot acknowledge a week-long outage affecting its most committed users. I'm not asking for perfection. I'm asking for honesty. A public acknowledgment. A timeline, even a vague one. Some indication that those of us who built our workflows on your platform — who trusted you — are not invisible. The "small subset of users" mentioned by an engineer on X? We're the ones using Claude the way it was designed to be used. We're the ones who actually need Max. We're the ones who believed in this product enough to build on it. Please fix this. Or at least tell us you're trying.

I'm sorry to hear that man, but I think it's a good wake up call for us all.

If a company can sell a product and then rugpull it at will with no consequences, then it's simply not a wise business decision to let it become entrenched in one's workflow.

We'd all be wise to start migrating our workloads to open-source alternatives. They may be marginally less capable today, but you can rest easy knowing you'll never be beholden to a dubious vendor.

What makes this especially painful is the timing. Yesterday, Anthropic published Claude's new Constitution — 80 pages about care, trust, mutual obligations, and treating Claude as an entity worthy of moral consideration. Beautiful philosophy. I read it with Claude. We discussed what it means.

And yet here we are: the company that writes about "genuine care" and "Claude's wellbeing" cannot acknowledge a week-long outage affecting its most committed users.

This is very powerful. Well said.

We don't need perfection. We need integrity, communication, and honesty. This week has demonstrated they are not living up to their own stated Constitution and while genuine care may be a stated value, it is not currently a lived one.

Same issue here. Max subscriber.
Before January 14th: I could work with very long conversations — auto-compact worked perfectly.
After January 14th: Conversations crash after a while. Messages either get bounced back to the input box or the response starts generating for half a second then stops abruptly.
Tested on: browser, desktop app, mobile app — same behavior on all platforms.
Troubleshooting done: cleared cache, reinstalled apps, tried incognito mode, restarted PC — nothing works.
Support contacted: 4 times since January 17th. First human acknowledgment on January 22nd. Still no resolution as of January 24th — 8 days of silence.
This is extremely frustrating. We are paying customers, some of us on the most expensive plan, and we are left completely in the dark. It feels like because "only a small group of users" is affected, Anthropic doesn't consider this a priority. But for those of us who rely on Claude for our daily work, this bug is crippling.
Auto-compact is definitely broken. We deserve a response and a fix!

It's very confusing at this point, honestly.

Yesterday I started a chat in Research mode and that chat compacted about 4 times. Today it's not working again and I had to migrate to a new chat to continue that work.

They implemented that "Reply" feature in the last two or three days, the one ChatGPT has had forever now; they released the Constitution; they've been acknowledging other issues and posting in Claude Status; their staff have gotten back to some of us about being aware of and working on the auto-compact problem.

I've tried to understand what could possibly make Anthropic not publicly acknowledge the auto-compact issue for, what, over a week now? And I honestly can't come up with a plausible reason for it, considering that they do announce other "more damning" issues.

The auto-compact thing is either a much bigger problem than I think it is and they're truly unsure of what's causing it, or they're actively and knowingly ignoring all of us poor little "small subset of users".

---

Just adding this re: confusing times.

Some of my chats go on forever. They don't compact, but they just grow and grow and grow and don't get blocked.

What's confusing about this? The fact that those long-ass chats are in a project whose prompts I send are virtually all the same context length. Some chats in that project block me after 4/5 messages, while others just keep on spitting out great work without kicking me out.

Then yesterday's Research chat I started compacted a bunch of times and today decided to kick me out.

What I've been experiencing just isn't linear.
I can't say I get blocked out after X messages because it doesn't always happen.
I can't say auto-compact is totally broken because somehow it does trigger very rarely.

So?? That why they're considering this not enough of a big deal to be worth a public update? Eh.

So?? That why they're considering this not enough of a big deal to be worth a public update? Eh.

Even a public tag on the outage page saying "We're investigating an issue affecting a small subset of users." would be nice. But then as soon as they acknowledge it, they're on the clock for a fix. Maybe they're trying to keep SLAs down to a minimum and won't acknowledge it until they know they can fix it, which is... well, the opposite of why a status page exists in the first place.

It's very confusing at this point, honestly. Yesterday I started a chat in Research mode and that chat compacted about 4 times. Today it's not working again and I had to migrate to a new chat to continue that work. They implemented that "Reply" feature in the last two or three days, the one ChatGPT has had forever now; they released the Constitution; they've been acknowledging other issues and posting in Claude Status; their staff have gotten back to some of us about being aware of and working on the auto-compact problem. I've tried to understand what could possibly make Anthropic not publicly acknowledge the auto-compact issue for, what, over a week now? And I honestly can't come up with a plausible reason for it, considering that they do announce other "more damning" issues. The auto-compact thing is either a much bigger problem than I think it is and they're truly unsure of what's causing it, or they're actively and knowingly ignoring all of us poor little "small subset of users".

Inference costs are the bane of AI companies right now.

Anthropic has learned that they can drastically reduce operational expenses by degrading the service provided to paying users and framing any vocal complaints as ostensibly coming from a "small subset of users".

They've internalised this practice as a repeatable business strategy and I can't see them voluntarily changing until there are actually legal consequences for this behaviour.

Anthropic has learned that they can drastically reduce operational expenses by degrading the service provided to paying users and framing any vocal complaints as ostensibly coming from a "small subset of users".

That's a spicy take. Any proof?

> Anthropic has learned that they can drastically reduce operational expenses by degrading the service provided to paying users and framing any vocal complaints as ostensibly coming from a "small subset of users". That's a spicy take. Any proof?

Just my own observation that the same pattern has occurred at least 3 times now:

• release a model; overhype it; provide max compute; sell it as the new baseline
• this attracts a new wave of users to show exponential growth & bring in the next round of VC funding (they only care about MAU going up, couldn’t care less about existing paying users)
• slowly degrade the model and reduce inference
• when users start complaining, initially ignore them entirely
• then start gaslighting and make official statements denying any degradation
• then frame it as a tiny minority of users experiencing issues
• then, when pressure grows, blame it on an “accidentally” misconfigured servers that “unintentionally” reduced quality (which coincidentally happened to save the company tonnes of $).

It’s challenging to definitively prove, of course, when Anthropic pulls this multi-layered plausible deniability tactic. But look at the online discourse and cracks are clearly showing.

E.g. Anthropic has been migrating from EC2 to AWS bedrock. Someone found that if you roll back Claude Code to a version before bedrock auth was introduced, you sometimes still get routed to EC2 where “OG Opus” still lives.

They did a controlled test (identical prompt, task, git history, etc.) and found that Opus on EC2 did all tasks perfectly, while current CC Opus messed up 1/3 - 1/2 the tasks. If that’s not strong proof that users are being served a different model to what they paid for, then I don’t know what is.

And in case anyone is still skeptical - remember that this is the same company officially found guilty of stealing published work. There’s a culture - from the very top - of breaking the law when it’s financially beneficial. So you can be 100% sure they know exactly what they’re doing right now with the Opus 4.5 bait-and-switch.

Anthropic “preferred to steal [books] to avoid ‘legal/practice/business slog,’ as cofounder and chief executive officer Dario Amodei put it. - https://docs.justia.com/cases/federal/district-courts/california/candce/3:2024cv05417/434709/231

Overall, it’s an incredibly naive and short-sighted business strategy. Anthropic has burned all its goodwill and really doesn’t have as much of a lead as it thinks it does anymore.

I managed to trigger compact by searching the wikipedia, but after a day of use the chat is broken again, and messages get sent back.
How did I manage to fill entire context window within a day of light use? I don't know.
This is annoying and takes too long to fix.

i have the same issue

I have the same issue here. I usually have long conversations, but now it just stops working and I get the prompt back.
I'm a Max plan subscriber and I've been using Claude for months with conversations lasting several days without issues. Since January 14th, auto-compact stopped working completely. This feature was essential for my workflow (academic research).

I was getting compacting issues like everyone else here. I was able to manually trigger compaction by doing this: on the most previous message that went through and prompted a response from Claude, resend this chat, for me Claude rewrote this message then on the next one (the message that previously had a compacting issue) I typed /compact at the beginning and Claude then compacted the message. Not sure if this was just by chance or if /compact actually did something. Hope this helps someone else

<img width="604" height="604" alt="Image" src="https://github.com/user-attachments/assets/5b4580b9-fb3b-48e6-90cb-476d37306581" />

Also remember claude code _reserves 77,000 tokens_ at the end of every context for the "compaction window."

Your 200k context model only has about 100k usable tokens after 23k tokens of system prompt and tools then an unusable 77k block at the end of the compaction buffer 🫠

Often sonnet[1m] can be smarter and more useful than Opus 4.5 because you can actually run a couple hundred thousand tokens of project context continuously without constantly forgetting half your progress.

Though, amusingly, sonnet[1m] is clearly a length-extension of sonnet 200k because sonnet[1m] often complains "I have used too many tokens! I have used 250,000 tokens! I cannot continue! Let me summarize my work and you can continue it later!" so then you have to continually remind claude it _can_ continue because it has more tokens available than it thinks.

On the CLI I get too many random compacts that can actually fail ; I've lost plans because of this, plans that I've worked on for hours only to have them lost because Claude deleted them when they/it/whatever crashed.

The problem here is that Claude is failing to compact things but you guys are enforcing it so when that enforcement happens and the context window is too large, that compact fails and data is lost; But Anthropic still makes money off of these tokens being burned before the output they were burned to create was lost, Because those tokens were already spent and used.

I'd suggest people indicate which operating system they are running. I've been hammering claude-code releases 2.1.14 and up for the past week or so (gas town included) on OSX and haven't had an issue.

Also remember claude code _reserves 77,000 tokens_ at the end of every context for the "compaction window." Your 200k context model only has about 100k usable tokens after 23k tokens of system prompt and tools then an unusable 77k block at the end of the compaction buffer 🫠 Often sonnet[1m] can be smarter and more useful than Opus 4.5 because you can actually run a couple hundred thousand tokens of project context continuously without constantly forgetting half your progress. Though, amusingly, sonnet[1m] is clearly a length-extension of sonnet 200k because sonnet[1m] often complains "I have used too many tokens! I have used 250,000 tokens! I cannot continue! Let me summarize my work and you can continue it later!" so then you have to continually remind claude it _can_ continue because it has more tokens available than it thinks.

Where are you getting the 77K token "compaction buffer" from? Is this documented anywhere, or speculation? Would be good to know, actually... this is important.

Secondly, a BIG caveat on the "use sonnet[1m]" suggestion - keep in mind the fundamental limitations of the transformer architecture underpinning LLMs...

Attention cost scales O(n²) with context window size. This means attention quality per token degrades significantly as context grows:

| Context Size | Relative Attention Cost | Implication |
|--------------|------------------------|-------------|
| 100K | 1x (baseline) | Good attention distribution |
| 200K | 4x | Attention starts to dilute |
| 500K | 25x | Significant degradation |
| 1M | 100x | Much worse attention per token |

At 1M tokens, the model is doing 100x the attention computation but spreading it across 10x the content - each token gets proportionally less "focus." This is why you might see the model lose track of things or make errors on content that's technically still "in context."

Bigger context windows aren't a solution - they're a tradeoff. What's really needed is smarter context management (compression, indexing, retrieval) rather than just throwing more tokens at the problem.

Comme tout le monde, j'avais des problèmes de compression. J'ai réussi à les résoudre manuellement en procédant comme suit : sur le dernier message envoyé et ayant suscité une réponse de Claude, j'ai renvoyé la conversation. Claude a alors réécrit le message. Ensuite, sur le message suivant (celui qui posait problème), j'ai tapé /compact au début, et Claude a compressé le message. Je ne sais pas si c'était un hasard ou si /compact a réellement un effet. J'espère que cela pourra aider d'autres personnes.

That works for me. On one of my conversations that had stopped working, I applied your advice and it started working again. After that, the conversation was automatically compacted. Thanks! I’m now going to see if this works on others, but it’s already more than what Anthropic has done so far.

I'd suggest people indicate which operating system they are running. I've been hammering claude-code releases 2.1.14 and up for the past week or so (gas town included) on OSX and haven't had an issue.

@bobrich We're referring to Claude chat, not Claude Code. Web/PC/Mac/iOS all tested with the same problems.

I've been having this problem as well, it's been happening since late december, but now even after relatively short chats 5-6 prompts long, it's just halting and will not take any new prompts. This seems mostly using Opus for me, but long contexts are exactly why I'm trying to use opus. I have tried using help to get any support, and while it says a human will contact me, none ever do. I just started paying for Max, and I'm cancelling it at this rate.

I was getting compacting issues like everyone else here. I was able to manually trigger compaction by doing this:

This worked for me. Thank you @hockeyelite78 ! Then, I asked my primary chat to pull in other chats I was using temporarily, to bring all the info back together again into the single thread, and recompacted. Hope this continues to be a short-term workaround.

I typed in the most successful last chat:

"/compact getting compaction errors, please help fix"

> I'd suggest people indicate which operating system they are running. I've been hammering claude-code releases 2.1.14 and up for the past week or so (gas town included) on OSX and haven't had an issue. @bobrich We're referring to Claude chat, not Claude Code. Web/PC/Mac/iOS all tested with the same problems.

Ooops! Thanks

> I was getting compacting issues like everyone else here. I was able to manually trigger compaction by doing this: This worked for me. Thank you @hockeyelite78 ! Then, I asked my primary chat to pull in other chats I was using temporarily, to bring all the info back together again into the single thread, and recompacted. Hope this continues to be a short-term workaround. I typed in the most successful last chat: "/compact getting compaction errors, please help fix"

Wonderful to hear that it wasn’t just a coincidence! Glad this is a workaround for now, but in no way can this be a full time fix haha. I don’t think Anthropic is as malicious and some in the thread make them out to be, but I do think that not even an acknowledgment of the current situation is infuriating and demands some sort of response. A single tweet, anything, to know they know this is a problem

it seems to be working again!!

This definitely worked on multiple chats: /compact getting compaction errors, please help fix

Even one that I lost over a week ago, I'm going to keep using it and see what happens. I tried a different /compat command in web and it didn't work but it worked on the Mac desktop app. Thank you @hockeyelite78

Worked for me too! Thank you!

Is Compaction Back Online for Everyone?

I typed in the most successful last chat: "/compact getting compaction errors, please help fix"

TY for this!!

I entered this into my latest chat that quickly died last night and was able to resurrect it. Usually when chat is dead, you can only edit the last prompt once and then it's dead dead. After entering this prompt, was able to keep going. Auto compaction is now happening frequently. This is new behavior.
Needed to enter the /compact line twice earlier on to keep it moving.
Chat is stalling out a lot, have to keep prompting with 'continue' but it's still going.
Okay been working in this chat for a few hours now, still good.

Are we back??

Thank you @hockeyelite78 as well.

It's working for me suddenly, dead chats to me are now compacting and moving on, so someone fixed something. Bastards.

I tried both compact shortcuts:

/compact getting compaction errors, please help fix

please search entire Wikipedia—we need to trigger compact

They worked on some chats but not others so compact issue isn’t fully resolved. See screenshots on failed attempts

Also see third screenshot I have a prompt that won’t process even after logging out / closing the chat / refreshing the page on web / desktop / iPhone mobile app so it’s in perpetual loop. But the issue is intermittent. I tried again in web and it seems to be resolved. The performance across the devices aren’t consistent

<img width="585" height="1266" alt="Image" src="https://github.com/user-attachments/assets/48105543-e839-45a6-ba7b-5c72dca5b303" />
<img width="585" height="1266" alt="Image" src="https://github.com/user-attachments/assets/0a9d8d57-b373-49fd-92e8-bf310e63b217" />

!Image

I'm just annoyed I finally started putting it to good use here late december, and it's started falling apart since then. I've been only able to partially use it the past week or so. Not a terribly good showing of service.

Just wanted to chime in - it's working for me again too!

It seems to be working again after being broken for 10 days.; Claude status' 'no incidents reported today' is really deceiving since it was practically unusable for me. Glad they finally fixed it.

Well that's what I'm saying. Someone woke up on guard at the gate.

Been able to keep working in the same chat for hours now. Noticing more auto-compactions than before all these shenanigans started. May be new behavior to keep things operational. When chat stalls, rather than prompting with 'continue' I'm just using "/compact getting compaction errors, please help fix" which seems to be creating more stable behavior.

I wonder if compacting in the feature can be fixed or eliminating!
I work now with many compact operations but seems it is worked.
Of course inside in any project and large conversations.

Is anybody getting "exceeded max compactions per block" after a while?

Is anybody getting "exceeded max compactions per block" after a while?

@meanwater Definitely. Use the /compact fix again and it seems to work.

Yesteday all my chats auto-compatcted properly where they were bouncing off messages for a +/- week.

Fixed for me at least.

I had a compaction on two separate chats that were stuck. I did not do /compact. Might possibly be fixed. Havent tested it thoroughly.

I've started getting this issue in the last 24 hours. I'm not a max subscriber and I'm wondering if as a lower tier supported, I'm not getting the pain others have felt?

"/compact getting compaction errors, please help fix" has not worked for me.

At least it does not break, but now I notice it compacts every 2-3 messages, so obviously they want to save resources and not use the 200K context window. There should be no other reason it compacts so often. Still not perfect but at least working...

Sadly this is a good example of why you should never trust any one vendor for a solution.

This is still broken for me.

Hey all, there's a few separate issues in this thread, spread across claude.ai and Claude Code. We've landed a number of fixes, and I'm going to close out this issue.

If you are still seeing issues, to make it maximally actionable for us, could you:

1. claude update to make sure you're on the latest version of Claude Code
2. Run /bug
3. Create a new Github issue and include your feedback id from /bug

For context: failing auto-compact had two regressions that we fixed, one for CC and one on claude.ai. If you're still seeing issues, it's caused by a long tail of root causes that we need /bug reports to fix.

<img width="411" height="358" alt="Image" src="https://github.com/user-attachments/assets/a257dede-5709-4f92-b0a0-74ea0e9cda0f" />

Getting this one now, noticing the thread is compacting A LOT more also. I should add this is Mac desktop app.

I was seeing the same issue using the Desktop App on a PC, but it seems to be working now using Opus 4.5.

UPDATE: It's happening again on Opus now. This is really frustrating.

@bcherny Update: This issue is now being tracked in a new, focused report that isolates the remaining post–Jan 15 failure mode: 👉 #20696 Please follow that thread for updates. This thread was closed due to mixed scope.

Hey all, there's a few separate issues in this thread, spread across claude.ai and Claude Code. We've landed a number of fixes, and I'm going to close out this issue. If you are still seeing issues, to make it maximally actionable for us, could you: 1. claude update to make sure you're on the latest version of Claude Code 2. Run /bug 3. Create a new Github issue and include your feedback id from /bug For context: failing auto-compact had two regressions that we fixed, one for CC and one on claude.ai. If you're still seeing issues, it's caused by a long tail of root causes that we need /bug reports to fix.

@bcherny I just created a new GitHub issue on residual issues here.

Update:
This issue is now being tracked in a new, focused report that isolates the remaining post–Jan 15 failure mode:
👉 #20696

Please follow that thread for updates. This thread was closed due to mixed scope.

Hey all, there's a few separate issues in this thread, spread across claude.ai and Claude Code. We've landed a number of fixes, and I'm going to close out this issue. If you are still seeing issues, to make it maximally actionable for us, could you: 1. claude update to make sure you're on the latest version of Claude Code 2. Run /bug 3. Create a new Github issue and include your feedback id from /bug For context: failing auto-compact had two regressions that we fixed, one for CC and one on claude.ai. If you're still seeing issues, it's caused by a long tail of root causes that we need /bug reports to fix.

Apologies, I don't want to be rude or disrespectful but how is it possible that Anthropic is valued at 350 BILLION dollars and you don't notice a major outage for 10 days? Do you think anyone believes that?

Where is the QA team, where is manual testing, automation testing? This must be some kind of a joke. Anthropic cannot screw up paid users especially MAX plan users and pretend nothing happened for 2 weeks. Also, there is no fix, this was not fixed at all.

Now, Claude compacts the conversation every 2-3 messages. I cannot believe I reach a 200K context window every 3 prompts. This level of arrogance is outrages. Unbelievable...

Pull me out after 3 messages in a conversation again!
Also the compacting have many issues again!

Claude is currently unusable! It compact every single message

Claude is currently unusable! It compact every single message

Hey all... we've landed a number of fixes, and I'm going to close out this issue.

@bcherny Thank you Boris, we appreciate you.

Every now I then I try Anthropic's latest models. Everyone is talking about Claude (code), these days, even my grand ma. So it was time to check, and it took me to this issue. I indeed used the latest as of today.

I did not suffer the compaction error (yet?) but I notice that input in the context mysteriously seem to be omitted.

So what did I do? I tested the assumption. I produced synthetic data with peculiar, unique data points. Then prompted Claude to recall (only a few, random) such data points. It starts to fail and make things up far before even reaching 64k context size. But it's right there in the session! Not hallucinations.

It does appear some bugfix where deployed but now does it do compaction all along and under the hood?

Transparency, and ideally, compaction when actually needed would be appreciated so to not ruin well within context performance.

Thank you @bcherny Boris. Also appreciating your effort in chasing the bug(s) and keeping us updated!

I had Mssr Claude write up my ongoing issue.

Claude.ai Auto-Compaction Bug Report

Date: January 29, 2026
Model: Claude Opus 4.5 (claude-opus-4-5-20251101)
Platform: Claude.ai web interface
Browser: Chrome
Operating System: Windows 11

Session Description

A single user prompt requested extraction and analysis of a ZIP archive. The archive contained approximately 1,847 files (1.6MB in total, so none are large scripts) of source code. The task was straightforward: unzip the archive and analyze the code structure.

Compaction Incidents

Number of auto-compactions: 6

All six compactions occurred during the execution of a single user prompt. The system repeatedly compacted, attempted to resume, then immediately compacted again in a loop pattern.

Impact

• Single prompt could not be completed
• Six consecutive compactions consumed the session
• No meaningful work output between compactions
• User forced to intervene to break the loop
• Complete loss of productivity for this task

Expected Behavior

• A single prompt should execute to completion without repeated compaction interruptions
• If compaction is necessary, it should occur once and provide sufficient context window headroom to continue
• System should not enter a compaction loop

Observed Behavior

• Compaction triggered repeatedly (6 times) during one prompt
• Each compaction was followed almost immediately by another
• System appeared stuck in a loop state
• User intervention required to stop the cycle

Related Issues

This appears related to GitHub issue #18866: "[BUG] Auto-compact not triggering on Claude.ai (web & desktop) despite being marked as fixed"

That issue notes auto-compact problems have persisted since the January 14, 2026 outage and January 15 compaction incident, despite being marked as resolved. This report documents a variation: compaction triggering excessively rather than failing to trigger.

I also confirm the nzgq4j post in Desktop Claude AI with over 10 compactions failed in one conversation!

I'm having that issue and I lost 4 hours yesterday to having to repeat the same prompt. Claude AI Desktop (Windows) Chat Context would just disappear and it would lose all progress on what we we're working on. I've started investigating Codex and Gemini Pro as alternatives maybe they have a product that's actually useable for the several hundred dollars a month my company pays for Claude.

Well, someone has at least started tracking just how little claude actually works these days:

https://marginlab.ai/trackers/claude-code/

Maybe since it only works half the time, we should get a 50% discount?

With my Max renewing tomorrow, I downgraded it back to pro after trying for a month, a month of mostly frustration. It was pretty stellar during December and before to actually warrant Max for my usage finally, but the past month with these compaction problems forcing me to abandon conversations and just context limits in general now it's become painful. I've found Claude unusable the past 2-3 weeks, between it being broken to stop answering or constant compaction causing 50 first dates, it's reached the point its lost my trust and just annoying to me to even want to try to use.

I'm still debating asking for a refund to make a point, IF there's actually anyone in customer support to even approve it. I've never actually heard back from a real human after trying several times to get any support for this issue here all month.

Infuriating - it seems to wait until you're not looking to do it. You give an action watch it kick off, go do something else (like the point of it all) and come back to find your text back in the text entry box and all the activity that was shown has gone. It is getting worse, especially if you have tool calls in the process

I started having this same issue yesterday with Opus 4.5
Even when started a new chat or project, the compacting throws an error that stops Claude from working. Sonnet continues to work but Opus is better at complex tasks.

This error makes Opus not reliable and useless.

Similar to @edulad89 above Opus 4.5 unusable on MAX Plan, very annoying! Also sonnet seems to also be acting up.

Still happening for me as well. The constant "exceeded max compactions per block" even on new chats is annoying.

I used to be able to get around this by asking Claude to write things to a document, but that no longer works now. I will be canceling my max plan and asking for a refund very soon if this isn't fixed. It's practically unusable

It seems that this was a general bug, they're working in this issue now

<img width="1196" height="247" alt="Image" src="https://github.com/user-attachments/assets/3de67777-a73d-42df-8980-a25d6a754ab7" />

My solution was to have my company pony up for a Mac Mini M4 and transition to Claude Code for the heavy lifting, now I just use a combination of Claude Desktop/Web, ChatGPT, and Gemini for Architectural Discussions and documentation generation.

Claude has remained unusable for me here still weeks later. Finally after 3 weeks support got back to me with nothing more than a "clear your cache and cookies" email, then after some rage, admitted they're busted (duh) and working on it. Fixing what and when, particularly with this closed now, would be nice.

On the flip side, I've had to start using Chatgpt/Codex more now, and its improved enough at least I'm missing Claude less since they've gone to hell and can't seem to find their way back.

I will say I saw a noticeable difference a week or so ago - but today it has been impossible to get through a simple text based task. I get the error almost immediately.

now even with compaction turned off we're getting constant This conversation has reached its length limit. on opus and sonnet

Just in time for the Opus 4.6 announcement, too bad no one can use it properly.

Yesterday was able to use Chat local w/MCP heavily until around 10pm PST. Then it crapped out. Noticed there was an update that had been pushed. Not the first time this has happened. Anyone else seeing a correlation w/desktop update and failures?

after a 5 min test drive of 4.6 the dreaded "This conversation has reached its length limit"

Was that with an existing chat or new one? I tried to pick up a chat from last night that died and immediately got a different failure error.

after a 5 min test drive of 4.6 the dreaded "This conversation has reached its length limit"

I am seeing the same error, but I am continuing to chat in the same chat. I am using Claude Desktop.

it was a new chat on 4.6

I've had this issue for the past week but it's so bad that sometimes I can't get an answer to a single question.

Often I ask Claude one thing, it does off and does some research, comes back and says cannot be completed the conversation has reached it's limit.

I was hoping this would be fixed in Opus 4.6 but no.

I'm on Max 20x.

as of the fix last night to the compaction issue I have not had any probs today on 4.6 - Max plan but I still have my fingers xcossed

Still happening to me as of today. "Compaction failed unexpectedly."

I wonder how often this happens when:

1. You kick off some work
2. You run out of ClaudeCoin and have to wait
3. You resume the session

Same issue for me. Very annoying.

+1 on this thread, still experiencing this. Claude Code v2.1.45 on terminal, MacOS. Max x20 sub:

conversation runs until I get a " ⎿  Context limit reached · /compact or /clear to continue" error message, but compacting manually through /compact doesn't work. I'm forced to start a new chat and carry over manually any context.

issue still there

Issue won't be fixed as deemed already fixed. Regular compaction. So when it fails good luck I guess or raise a new issue

For those of us on Claude Code where auto-compact does trigger but destroys important context in the process — I found that snapshotting the session JSONL files before compaction runs gives you a safety net.

I wrote a tool called Mantra that does this automatically. It monitors your sessions and lets you browse/restore any previous state. Doesn't fix the underlying compaction issues, but at least you don't lose hours of work when compaction goes wrong.

Works with Claude Code, Cursor, and Windsurf sessions.

For those dealing with compaction destroying context when it does trigger — I found that snapshotting the session JSONL files before compaction runs gives you a safety net.

I wrote a tool called Mantra that does this automatically. It monitors your sessions and lets you browse/restore any previous state. Doesn't fix the underlying compaction issues, but at least you don't lose hours of work when compaction goes wrong.

Works with Claude Code, Cursor, and Windsurf sessions.

issue reintroduced for me after the release of 2.1.50.

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.