Read tool sends requests exceeding 20MB API limit without pre-checking payload size

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/13823
2. https://github.com/anthropics/claude-code/issues/22699
3. https://github.com/anthropics/claude-code/issues/8092

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

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

🤖 Generated with Claude Code

i also have this error.

me too

this is happening to me over time as well. the "file size" accumulates with a bunch of smaller images returned over an MCP and Claude Code neglects to prune them and / or intelligently decide what goes in

Issue happens on Windows too. Not only MacOS

did anyone find any fix for this ??

i am on linux mint os , and have the same issue !
● Request too large (max 20MB). Double press esc to go back and try with a smaller file.

I'm getting this every day now with the 1M context window

I'm getting this every time 😭 with small size files

Issues happening consistently. I'm on max plan. I urge everyone to downgrade when possible and look at switching providers

Same issue here. Can't continue my session anymore. Whenever I send a new message I just get the same error.

Joining this party since I can't really get much work done without hitting this very stupid error. Idk what happened but claude today is not the claude I was working with yesterday. Drastically dumbed down.

this shit is genuinely pathetic man i cant fucking stand anthropic @bcherny get your fucking head out of your ass you clown

@claude fix this bug plz, make no mistakes

@claude this is pretty annoying

Request too large (max 20MB). Double press esc to go back and try with a smaller file. hey claude can u please fix this

Request too large (max 20MB). Double press esc to go back and try with a smaller file.

Please fix this !

I'm getting the same over here. Please fix. Even while sending a short message.

same here

I get this pretty often while using the playwright mcp

Max plan user ($200/mo) — this is a daily blocker

I'm hitting this bug multiple times per day as a Max plan subscriber. The irony is brutal: the users paying the most are the ones most affected, because we're the ones running longer, heavier sessions — which is exactly what we're paying for.

The real damage isn't the error — it's the aftermath

The error itself is annoying, but the actual cost is what happens next:

1. Conversation bricks — every subsequent message fails with the same error. No recovery possible.
2. Forced to /clear — which nukes all conversation context.
3. The AI gets significantly dumber after clearing — it loses all accumulated understanding: decisions made, approaches rejected, preferences learned during the session, the mental model of how things connect. Compaction summaries are lossy. /clear is a full lobotomy.
4. Work slows to a crawl — I spend time re-explaining context that was already established, re-correcting mistakes the AI had already learned not to make.

The trigger is deceptive

The error message blames the current file/image, but the real cause is accumulated conversation payload silently approaching the limit. A 99KB PNG or a short text message tips it over. Users can't predict or prevent it because there's no visibility into how close you are to the limit.

What needs to happen

• Pre-validate payload size before sending to the API — don't brick the conversation
• Graceful degradation — if approaching limits, auto-compact or warn the user before it's too late
• Transaction rollback — if the API rejects, roll back to the previous state instead of corrupting the session
• Accurate error messaging — "Context too large, try /clear or /compact" instead of blaming a small file

This is a paying customer telling you: this single bug is the #1 thing degrading my experience with Claude Code. Not missing features, not model quality — this. Please prioritize it.

This is so neglected that the notifications of new comments in the GitHub issue are now as annoying as the issue itself

Can everyone please tag @bcherny on X regarding this issue? This is quite urgent. As a paying user on the $100 plan, I’m experiencing this repeatedly, and it’s impacting usage significantly. Would really appreciate

The best is if they made the agent harness open source so we could fix these ourselves (there are some already)

Otherwise we will need to build our own harness (not to mention our own open weights models)

OK it is, and you can create you own harness. Coming from a frequent rage poster, this is manageable by a well made harness, optimizing MCPs, controlling context bloat, and just restarting the session after it creates a comprehensive hand-off. Have it audit, have another agent audit it against the conversation and move on. You can only expect so much from this awesome tool. After doing those things and restarting my sessions more frequently, 0 instances of this issue. Not to say it does not still exist but it is entirely manageable while they sort it out.

I would still take this over gemini's hack and slash harness.

Interesting, I’ll keep this in mind. Thanks

On Wed, Mar 25, 2026 at 12:02 PM LewAdvSol-Ian @.***>
wrote:

LewAdvSol-Ian left a comment (anthropics/claude-code#26018) <https://github.com/anthropics/claude-code/issues/26018#issuecomment-4127801943> OK it is, and you can create you own harness. Coming from a frequent rage poster, this is manageable by a well made harness, optimizing MCPs, controlling context bloat, and just restarting the session after it creates a comprehensive hand-off. Have it audit, have another agent audit it against the conversation and move on. You can only expect so much from this awesome tool. After doing those things and restarting my sessions more frequently, 0 instances of this issue. Not to say it does not still exist but it is entirely manageable while they sort it out. I would still take this over gemini's hack and slash harness. — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/26018?email_source=notifications&email_token=AJXYKRUOOEXTRSZJZW7E5T34SP7JTA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJSG44DAMJZGQZ2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4127801943>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJXYKRUPGR622QORI32XAQT4SP7JTAVCNFSM6AAAAACVHQLZASVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCMRXHAYDCOJUGM> . You are receiving this because you commented.Message ID: @.***>

I was able to skip this problem using /compact.

You can get around this issue by:

1. /rewind to the latest message before the error
2. /compact instead of whatever you were asking Claude
3. Now ask again that thing from Claude

The slow response from @anthropics should make you pause, though, and consider open source alternatives:

• https://github.com/nousresearch/hermes-agent
• https://github.com/badlogic/pi-mono/tree/main/packages/coding-agent
• https://github.com/krzyzanowskim/claude-better

(Not to mention open weights models)

Their models has been degraded. Opus 4.6 among others have been degraded due to financial benefits from the company. I will switch to using a fake credit card. I urge you to do the same

Claude Code full code has been leaked, so now you can fix this on your own if you'd like (just ask Claude to do it)

Facing the same issue when working with PDF and big files as a source.