claude-code - 💡(How to fix) Fix Advisor tool completes in UI but result never injected into model context (under Remote Control)

Summary

When using Claude Code with advisor mode enabled and Remote Control active, the advisor tool shows full completion in the status line (with multi-minute durations like "Baked for 57s" or "Cogitated for 4m 28s"), but the model's conversation receives zero advisor tool-result blocks. The model narrates intent to call the advisor but receives no response, causing it to proceed without the requested second opinion. This occurs consistently in long-running interactive sessions and appears specific to the Remote Control execution mode.

Environment

• Claude Code version: Latest (verify with claude --version)
• Advisor model: Opus 4.8
• Executor model: Opus 4.8
• Remote Control: Active (/remote-control initiated)
• Session type: Long-running interactive multi-step task with MCP tool use
• OS/Platform: macOS
• Reproducibility: 3/3 attempts in same session

Steps to Reproduce

1. Enable advisor mode in settings (advisor model set to Opus 4.8)
2. Initiate Remote Control (/remote-control)
3. Begin a long-running interactive session involving multiple steps with MCP tool calls
4. At some point during task execution, the model will narrate that it wants to call the advisor tool for a second opinion before proceeding
5. Observe the UI status line as the advisor call is issued
6. Watch the status line show the advisor completing with a duration (e.g., "Baked for 57s", "Cogitated for 4m 28s", "Brewed for 2m 9s")
7. Return focus to the chat transcript and check the model's next message

Expected Behavior

After the advisor tool completes (as shown by the status line), the model's conversation should include one or more advisor tool-result blocks containing the advisor's response. The model should reference this advice in its subsequent reasoning and decisions.

Actual Behavior

The status line correctly reports advisor completion with multi-minute durations, confirming the tool was triggered and finished. However, the model's conversation transcript contains no advisor tool-result blocks. The model's next message proceeds as if no advisor call was ever made or answered—it has no record of the advice and cannot reference it. The two halves (UI/status reporting vs. model context) disagree.

Frequency

Reproduced in 3 out of 3 attempts to call the advisor during a single long-running session under Remote Control.

Related Issues

• #63880: "Advisor narrates but then fails to make tool calls" — same prose-then-no-tool-call symptom; suspected relation to parameterless tools in long-context or batched-tool-result sessions
• #49994: Advisor sessions unrecoverable
• #55889: Context-injection channels dropped

Diagnostic Info to Attach

When filing this issue, please run and share the output of:

• claude --version (Claude Code version)
• /doctor (environment and configuration diagnostics)
• /context (current context state and token usage)
• /status (session status and advisor state)

Additionally, capture:

• The session transcript (or a redacted portion showing the advisor call attempt and model's next turn)
• A screenshot of the status line showing the completed advisor tool with duration

Additional Notes

Caveat: The reporter cannot inspect harness internals and therefore cannot definitively determine whether the advisor executes server-side and the result fails to inject back into model context, or whether the model never actually emits the call while the UI spinner shows activity anyway. However, the observable and reproducible facts are:

1. UI status line clearly shows advisor tool completion with significant elapsed time (multiple minutes)
2. Model conversation transcript receives zero advisor tool-result blocks
3. Model's reasoning on the next turn shows no awareness of or reference to any advisor response

This suggests a disconnect in the result-injection path under Remote Control specifically.