[BUG] API returns thinking-only response with no text block — tokens consumed but response lost

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?

After a subagent (search-specialist) returns results, the main assistant's relay response flashes briefly in the CLI then disappears. The JSONL session log records the assistant turn but it contains only an empty thinking block — no text block — despite output_tokens: 230 being consumed. The response is lost.

What Should Happen?

The assistant's text response should persist in the CLI and be written to the session JSONL with a text content block.

Error Messages/Logs

No error shown. The assistant turn in the JSONL has this structure:


{
  "type": "assistant",
  "message": {
    "content": [{"type": "thinking", "thinking": "", "signature": "..."}],
    "stop_reason": "end_turn",
    "usage": {
      "output_tokens": 230,
      "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0},
      "speed": "standard",
      "inference_geo": ""
    }
  }
}


Anomalous fields compared to every other assistant turn in the same session:

| Field | Normal turns (33 total) | Dropped turn |
|---|---|---|
| `context_management` | `{'applied_edits': []}` | **missing entirely** |
| `usage.server_tool_use` | absent | `{'web_search_requests': 0, 'web_fetch_requests': 0}` |
| `usage.speed` | absent | `"standard"` |
| `usage.inference_geo` | `"not_available"` | `""` (empty string) |

All other assistant turns in the session have consistent structure. Only the dropped turn has these anomalies, suggesting a different server-side code path.

Steps to Reproduce

1. Start a Claude Code session with Opus
2. Use the Agent tool to launch a subagent (e.g. search-specialist) that returns a substantial result (~800 words)
3. When the agent result comes back, observe the assistant's relay response
4. The response text flashes briefly then disappears
5. Check the session JSONL — the assistant turn has only a thinking block, no text block

Not reliably reproducible on demand — intermittent.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.101 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

• This is distinct from the client-side persistence bugs (#47657, #50594 ) where no assistant JSONL entry is written at all. Here the entry exists but is structurally incomplete.
• The server_tool_use, speed, and empty inference_geo fields are not present on any other turn in the session, indicating the response was routed through a different server-side path.
• 230 output tokens were consumed but only an empty thinking block was persisted — the text content was generated (user saw it flash) but not included in the final response object.
• No reprimand, profanity, or ethics trigger involved — this was a normal agent result relay.