[BUG] Thinking mode 400 error on FIRST tool call in brand new conversation (Opus 4.5)

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?

This is NOT a long-conversation or auto-compact issue. The thinking block error occurs on the very first tool call in a brand new conversation — making thinking mode completely unusable.

Error Message

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.0: If an assistant message contains any thinking blocks, the first block must be \`thinking\` or \`redacted_thinking\`. Found \`text\`."}}

Steps to Reproduce

1. Open Claude Code with Opus 4.5 model
2. Enable thinking mode (Tab key)
3. Send any message that triggers a tool call (e.g. "check my claude version")
4. Claude responds with thinking + text + tool_use (first turn works)
5. After the tool result returns, Claude tries to respond again → 400 error

This is 100% reproducible. Every single new conversation breaks after the first tool call when thinking is enabled.

Key Difference from Other Issues

Most reported thinking block issues (#12311, #20714, etc.) occur in long conversations or during auto-compact. This issue is different:

• Happens in a brand new conversation (not resume, not long session)
• Fails at messages.1 (the first assistant message) — not messages.33 or messages.125
• Triggers on the first tool call — no compaction involved
• 100% reproducible, not intermittent

This suggests the client is not properly preserving/ordering thinking blocks when constructing the second API request after a tool call, even in the simplest case.

What I've Tried

• Downgrading from v2.1.19 to v2.1.14 — same error
• Removing native install, using npm only — same error
• Running without environment variables (clean claude --dangerously-skip-permissions) — same error
• Disabling alwaysThinkingEnabled in settings — same error (Tab-toggled thinking still breaks)
• Starting fresh conversations with /exit — same error on every new conversation

Only workaround: Not enabling thinking mode at all. With thinking disabled, everything works fine.

Environment

• Claude Code version: 2.1.19 (also tested 2.1.14)
• Model: claude-opus-4-5-20251101 (Opus 4.5) via Claude Max
• OS: macOS 15.7.2 (arm64, Apple Silicon)
• Node.js: v25.2.1
• npm: 11.6.2
• Installation: npm global (/opt/homebrew/bin/claude)

Expected Behavior

Thinking mode should work across multi-turn conversations with tool calls without corrupting thinking block ordering.

Actual Behavior

The client fails to preserve thinking block ordering in assistant messages when constructing the follow-up API request after a tool result, causing a 400 error on every conversation where thinking is enabled and a tool is called.