tool_use content block silently dropped despite stop_reason=tool_use

Summary

In a Claude Code session via cc-connect, the API returned a response with stop_reason: tool_use but the content array contained only a text block — the tool_use block was missing. The model consumed 581 output tokens but only ~70 chars of text were present in the content. The remaining ~500 tokens (likely the tool_use block) were silently dropped.

This caused the agent to appear to "announce intent" without acting, when in fact it had tried to call a tool.

Evidence

• API message ID: msg_01T93GF4xZQickPQqfSKN67D
• Model: claude-opus-4-6
• Timestamp: 2026-03-27T13:57:44Z
• Session ID: 7bbd674c-4adf-40dd-b570-30e89d25c0b1

Response metadata

{
  "stop_reason": "tool_use",
  "content": [{"type": "text", "text": "Let me orient first — pulling context on this channel and recent work."}],
  "usage": {
    "input_tokens": 3,
    "cache_creation_input_tokens": 22156,
    "cache_read_input_tokens": 11204,
    "output_tokens": 581
  }
}

Note: stop_reason: tool_use + content has zero tool_use blocks. 581 output tokens but text content is only ~15 tokens.

Context

• The session had been dormant for ~2h35m before being resumed with a user message
• The user message included a large system-reminder with config change diffs (~2000 chars)
• The thinking block (23 tokens) was delivered correctly in a prior streaming chunk

Frequency

Audited 20,855 responses with stop_reason: tool_use across all sessions — this is the only instance where the tool_use content block was missing. Rate: 1 in 20,855.

Expected behavior

When stop_reason: tool_use, the content array should always contain at least one tool_use block.

Impact

The agent appeared to produce a text-only response and "end its turn" without acting. The user had to manually re-prompt. We've implemented a detection guard (cc-connect auto-retries zero-tool-call first turns on triggered sessions), but the underlying API/SDK issue should be investigated.

Related

Linear issue: ULY-495