[BUG] Session becomes permanently unrecoverable after empty text content block during streaming

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?

A Claude Code session becomes permanently unrecoverable when the model streams an empty text content block alongside thinking blocks. After the API rejects the malformed response, every subsequent operation fails because the erroneous assistant message remains stuck in the conversation history. The error recovery code does not discard or repair it.

When this happens:

• Retry fails with the same error
• /compact fails (needs API call with corrupted history)
• /debug fails (same reason)
• Session is permanently bricked — only manual JSONL surgery recovers it

What Should Happen?

When the API rejects an assistant message, Claude Code should:

1. Discard the malformed assistant message from conversation history
2. Allow the user to retry cleanly
3. At minimum, /compact should recover by dropping the bad message

Error Messages/Logs

First error (from the empty text block):
messages: text content blocks must be non-empty

All subsequent errors (from re-serialized thinking blocks):
messages.1.content.1: thinking or redacted_thinking blocks in the latest assistant message cannot be modified. The thinking and redacted_thinking fields must contain their original values from the API response.

Error cascade repeats every retry with 85+ failed attempts recorded in transcript.

Steps to Reproduce

1. Run a long conversation (~5500 JSONL lines) using Opus 4.6 with extended thinking enabled
2. Fill context window until auto-compaction triggers
3. Send a message after compaction fires
4. Observe that response comes from Haiku instead of Opus (post-compaction model downgrade)
5. Response contains empty text content block between two thinking blocks
6. API rejects with "text content blocks must be non-empty"
7. User attempts retry, /compact, or /debug
8. Every operation re-sends corrupted message → cascading "thinking blocks cannot be modified" errors
9. Session is permanently unrecoverable — no command can fix it

Non-deterministic streaming issue, but happens reliably under these conditions:

• Extended thinking enabled
• Post-compaction turn (after model downgrade to Haiku)
• Large cached context (~67K tokens)
• Model producing thinking → text → thinking → text → tool_use pattern

Claude Model

Not sure / Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.37

Claude Code Version

2.1.38 (Claude Code)

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Other

Additional Information

Root Cause: The model (Haiku 4.5 - claude-haiku-4-5-20251001) streamed an empty text content block between two thinking blocks during post-compaction response:

• Content block 0: thinking block (346 words)
• - Content block 1: text block with empty string "" ← the poison
• - - Content block 2: thinking block (156 words)
• - - - Content block 3: text block (actual response)
• - - - - Content block 4: tool_use

The empty text block is invalid when re-sent to the API, but it was the streaming response itself that produced it.

Workaround: Sessions can be manually recovered by:

1. Locate the corrupted assistant message (contains empty text block)
2. 2. Truncate the JSONL transcript at the last clean assistant turn
3. 3. Resume with claude --continue

This workaround has been validated as working, though it requires manual file editing.

Related Issues: Similar symptoms reported in #23463 (context overflow), #12743 (thinking block API errors), and #10343 (unrecoverable state). This issue is more specific: empty text blocks during streaming at post-compaction, with evidence of model downgrade from Opus → Haiku as a contributing factor.