fix(coding-agent): recover from 413 request too large

[ogulcancelik]

Anthropic returns HTTP 413 request_too_large when the request payload exceeds 32 MB. I could reproduce this consistently in image-heavy sessions where context token usage was still relatively low (~30%), but the request body had grown too large because conversation history contained many base64-encoded images.

I couldn’t reproduce the same failure on OpenAI models, likely because their effective payload limits are higher, but 413 is a standard HTTP error for this class of problem.

Root cause

Compaction currently tracks token usage, not HTTP request size.

That works for normal text-heavy sessions, but images are a bad mismatch for token-only accounting:

• images are relatively cheap in tokens
• images are expensive in bytes once embedded as base64 in message history

So a session can stay far below the context window while still exceeding the provider’s request size limit.

Fix

When a request fails with HTTP 413 / request_too_large:

• detect it separately from token-based context overflow
• remove the failed assistant error from in-memory state
• strip the oldest half of image blocks from context
• replace stripped images with a text placeholder
• retry the request
• repeat up to 3 attempts

This recovery now works even when auto-compaction is disabled.

The stripping logic handles images in:

• user messages
• toolResult messages
• custom messages

Session history on disk is left untouched. Only the in-memory request context is modified for recovery.

Tests

Added tests for:

• isRequestTooLarge() detection
• image counting and oldest-first stripping
• repeated stripping across retries
• custom messages with images
• recovery from 413 even when auto-compaction is disabled

[ogulcancelik]

also should i have opened an issue first ? let me know if you do prefer issue first prs now, just wanted to provide a quick solution.

[badlogic]

currently refactoring agent session and fixing this will be part of that work.