Stop hook breaks -p (print/pipe) mode — output is silently empty

Bug Description

Any Stop hook configured in ~/.claude/settings.json causes claude -p (print/pipe mode) to return empty output. The model processes the request and consumes tokens, but no text reaches stdout.

Reproduction Steps

1. Add a Stop hook to ~/.claude/settings.json:

{
  "hooks": {
    "Stop": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "say 'Done'"
          }
        ]
      }
    ]
  }
}

1. Run: echo "say hello" | claude -p
2. Expected: Text output (e.g., "Hello! How can I help you today?")
3. Actual: No output at all

Key Findings

• The model does generate a response — confirmed via --verbose --output-format stream-json, which shows the assistant message with valid text content (e.g. {"type":"text","text":"HELLO WORLD"})
• However, the final result field in the JSON output is an empty string ""
• Tokens are consumed and billed normally
• Even a no-op Stop hook ("command": "true") triggers the bug — it's not about what the hook does, but the mere presence of a Stop hook entry
• The Notification hook does not cause this issue — only Stop
• Setting "Stop": [] (empty array) works around the issue
• Affects both opus and sonnet models
• Reproducible from any working directory (not related to CLAUDE.md or project config)

Environment

• Claude Code version: 2.1.83
• Platform: macOS (Darwin 24.6.0)
• Shell: zsh

Workaround

Remove or empty the Stop hook array in settings:

"Stop": []

This restores -p mode output but loses the Stop hook functionality (e.g., audio notification when interactive sessions complete).

Impact

This silently breaks any scripts or automation using claude -p when the user also has Stop hooks configured for interactive use. There's no way to conditionally apply hooks for interactive vs pipe mode, and no error is surfaced — output is just empty.