[BUG] "Request too large (max 20MB)" error on small files (99KB PNG)

[biznmo]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

After extended multi-day session with compaction, Claude Code triggers
"Request too large (max 20MB)" on operations involving small files. Observed
when writing a 160-line (~8KB) markdown file and when reading small PNG
screenshots (99KB-224KB). Files are confirmed small via ls -la. The error
appears to be a false positive from degraded context state, not actual file
size. Closing and reopening the session is the only workaround.

What Should Happen?

Claude should read and write small files without size errors regardless of
session length. If the session context has grown too large from extended
multi-day use or many file operations/screenshots, Claude should either
auto-compact proactively before hitting limits, or surface a helpful message
like "Session context is full — compacting now" instead of the misleading
"Request too large (max 20MB)" error, which implies the file itself is too
large when it isn't.

Error Messages/Logs

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

Steps to Reproduce

1. Keep a Claude Code session open for extended period (multiple days, with
   automatic compaction occurring)
2. Continue working — writing files, reading files, normal operations
3. Eventually any file operation (Write, Read, drag-and-drop) triggers the
   error

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

2.1.76

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Other

Additional Information

Bug correlates with long-running sessions (multi-day with compaction). Fresh
sessions work fine. Not file-type specific — happens with both .md and .png
files. Only workaround is closing and restarting the session.

[github-actions]

Found 3 possible duplicate issues:

1. [BUG] This error kills the active session: API Error: 413 {"error":{"type":"request_too_large","message":"Request exceeds the maximum size"}} #8092
2. Read tool sends requests exceeding 20MB API limit without pre-checking payload size #26018
3. [BUG] Accumulated PDF/image Read tool results permanently kill conversations — unrecoverable rate limit loop #30542

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

[yurukusa]

The error is misleading — it's not the individual file that's too large. It's the total API request payload (conversation history + compacted summaries + all tool outputs) that's approaching the 20MB limit. After multiple compaction cycles in a multi-day session, residual context accumulates.

What helps:

1. /clear instead of restarting. Resets conversation history while keeping CLAUDE.md loaded. Faster than a full restart and immediately resolves the size issue.

2. Avoid drag-and-drop images in long sessions. Each image gets base64-encoded into conversation history, which inflates context size much faster than text. Use Read tool with file paths instead — the output gets compacted more efficiently.

3. Monitor context with /cost. When you see usage climbing high, proactively /clear before you hit the wall. Beats the "it was working and suddenly stopped" experience.

4. For multi-day work, start fresh sessions per logical task. Each compaction cycle leaves some residual overhead. After several compactions, the accumulated overhead can push you near the limit even without large files.

The error message itself should probably say "Context too large, try /clear" rather than "Request too large (max 20MB)" — that would save a lot of confusion.

[biznmo]

Thanks for the quick and thorough breakdown! The distinction between individual file size vs total API payload context was exactly the "aha" moment — the error message really had me thinking it was the file itself. Already implemented your suggestions: - Added context management rules to my CLAUDE.md (avoid drag-drop images, use file paths + Read tool, proactive /clear, handoff docs before clearing) - Starting fresh sessions per logical task instead of multi-day marathons The error message suggestion is spot on — "Context too large, try /clear" would save a lot of confusion. Hope the team picks that up. Really appreciate the fast response!

…

On Sun, Mar 15, 2026 at 6:02 PM ゆる草 ***@***.***> wrote: *yurukusa* left a comment (anthropics/claude-code#34751) <#34751 (comment)> The error is misleading — it's not the individual file that's too large. It's the total API request payload (conversation history + compacted summaries + all tool outputs) that's approaching the 20MB limit. After multiple compaction cycles in a multi-day session, residual context accumulates. What helps: 1. */clear instead of restarting.* Resets conversation history while keeping CLAUDE.md loaded. Faster than a full restart and immediately resolves the size issue. 2. *Avoid drag-and-drop images in long sessions.* Each image gets base64-encoded into conversation history, which inflates context size much faster than text. Use Read tool with file paths instead — the output gets compacted more efficiently. 3. *Monitor context with /cost.* When you see usage climbing high, proactively /clear before you hit the wall. Beats the "it was working and suddenly stopped" experience. 4. *For multi-day work, start fresh sessions per logical task.* Each compaction cycle leaves some residual overhead. After several compactions, the accumulated overhead can push you near the limit even without large files. The error message itself should probably say "Context too large, try /clear" rather than "Request too large (max 20MB)" — that would save a lot of confusion. — Reply to this email directly, view it on GitHub <#34751 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AZGCR4LZL3TSV6RW4CK4NUL4Q5AB3AVCNFSM6AAAAACWSVS2FSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DANRUGIZDONZZGM> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[Djay77]

I have the same problem with MCP Chrome-Devtool. After only 1 screenshot !!
/clear do nothing. If I restart claude, the same

⏺ chrome-devtool - navigate_page (MCP)(type: "url", url: "http://localhost:3000")
⎿  Successfully navigated to http://localhost:3000.
## Pages
1: http://localhost:3000 [selected]

⏺ chrome-devtool - take_screenshot (MCP)
⎿  Took a screenshot of the current page's viewport.
⎿  [Image]

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

[AmazingAng]

have the same issue

[rockstarr896]

Same herer

[gabdevbr]

Same here

[LewAdvSol-Ian]

The best thing can do is optimize, and clear/compact. I have 5 monitors that drawns the request limit, after forcing clause not to make the request directly causing him to lock up...smh we finaly got this optimization. I have not have the error since, I suspect the cause will be different depending on your workflow.

Problem: The windows-mcp MCP server encodes screenshots as uncompressed PNG at 1920x1080
resolution, which on a complex multi-monitor desktop produces images of 8-10MB+. When
base64-encoded for MCP transport, this exceeds Claude Code's 20MB API request size limit (a known
bug tracked in GitHub issues #34751, #26018, #26019, #36469), causing Claude Code to either error
with "Request too large (max 20MB)" or hang indefinitely with no recovery path. Fix: Three changes
 applied to windows_mcp/tools/_snapshot_helpers.py in the uv cache — (1) reduced MAX_IMAGE_WIDTH,
MAX_IMAGE_HEIGHT from 1920x1080 to 1280x720, (2) changed the image encoding from format="PNG" to
format="JPEG", quality=70, optimize=True with an RGBA→RGB conversion since JPEG doesn't support
alpha channels, (3) updated the MCP response format tag from format='png' to format='jpeg', and
additionally set MAX_MCP_OUTPUT_TOKENS=50000 in ~/.claude/settings.json to raise Claude Code's
default 25,000-token MCP response limit.