[BUG] Session resume fails with 503 when JSONL contains accumulated PDF/file attachments (Bedrock provider)

Preflight Checklist

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

What's Wrong?

Claude Code persists every PDF/file attachment to the session JSONL in two places: as base64 in document/image content blocks, AND as base64 in toolUseResult.file.base64 cache entries. Both get included when the session is replayed on --resume, sending the cumulative bytes from every prior turn in a single request to the provider. On the Bedrock provider (CLAUDE_CODE_USE_BEDROCK=1) this routinely exceeds backend limits and the entire session becomes unresumable.

In my case a session of 309 lines with ~5 PDFs attached over its lifetime grew to 36.8 MB on disk. After stripping the base64 payloads (replacing with text placeholders), the same session was 0.61 MB and resumed instantly.

/compact is not a workaround: it runs as a request against the bloated history, so it fails with the same 503.

What Should Happen?

A session that successfully created or read N attachments should remain resumable regardless of N. Concretely, one or more of:

Don't persist full base64 in the JSONL; persist a reference (path/hash) and re-load on demand
Cap individual attachment size in JSONL; persist a reference for anything above the cap
Provide a built-in command (claude session prune <id> or similar) that strips attachment payloads while preserving conversation structure
Make /compact resilient to oversized history (e.g., truncate attachment payloads before sending its own request)
Any of these would prevent unrecoverable session loss from accumulated attachments.

Error Messages/Logs

In the broken session on resume, after the first prompt:

503 {"detail":{"error":"{\"message\":\"Bedrock is unable to process your request.\"}"}}
Retrying in 0s · attempt 1/10
... [identical retries through attempt 10/10]
The 0-second retry interval also seems wrong — it can prolong throttling on the upstream gateway.

Steps to Reproduce

1. macOS 15.x, Claude Code 2.1.170 (SF-managed distribution)
2. Provider: Bedrock via gateway

CLAUDE_CODE_USE_BEDROCK=1
ANTHROPIC_BEDROCK_BASE_URL=<gateway>/bedrock

1. Model: us.anthropic.claude-opus-4-7[1m]
2. In a fresh session, attach 4–5 PDFs of ~5–10 MB each via the Read tool over the course of normal conversation. Continue working with the assistant for ~30 turns.
3. Exit Claude Code.
4. Inspect the session JSONL at ~/.claude/projects/<project-hash>/<session-id>.jsonl — observe size in tens of MB.
5. Resume the session: claude --resume <session-id>
6. Send any prompt (e.g., "test").

Expected: prompt is processed normally.
Actual: request fails with the 503 above, retries 10x, session is unrecoverable.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.153 (Claude Code)

Platform

AWS Bedrock

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

2.1.153 (Claude Code) is the latest version I get when I run claude --update in my organization