[BUG] /btw appends "Never consult the advisor tool." to the user's message, so the model treats it as a user instruction

Summary

When an advisor model is configured, Claude Code appends an instruction like "Never consult the advisor tool." to every /btw side question. It is appended to the user's message, so the side-chat model treats it as something the user wrote: it quotes it, asks about it, or carries it into its answers.

What happened

I asked a /btw side question and asked the side chat to write a message for me to paste into the main conversation. It added this as a numbered item in that message:

4. Never consult the advisor tool.

I asked /btw why "4. Never consult the advisor tool."? and it replied:

I added it because that exact line has ended every side question you've sent me, including this one. I read it as a standing instruction from you. Side questions never reach the main agent, so an instruction that only appears here has no effect. Putting it in the message you paste into the main conversation was the only way for it to reach the agent. If you didn't write it on purpose, drop item 4. It could be a suffix your setup adds automatically, such as a hook or saved prompt. I can't tell which from here.

I never wrote that line. I checked my settings.json, CLAUDE.md, hooks, commands, agents, output styles and installed plugins, and none of them contain it. The only advisor setting I have is "advisorModel": "opus".

Impact

• The model sees a rule that looks like it came from the user, but that is really a harness rule. It reasonably follows it and passes it along. In my case it copied the rule into a message I would have pasted into the main session, where it would have turned off advisor there too.
• Users are confused by an instruction they never gave.

Repro

1. Set an advisor model (/advisor, or "advisorModel": "opus" in settings).
2. Ask any /btw question, e.g. /btw what text follows my question in this message?
3. The side-chat model reports a "never consult the advisor tool" line at the end of the user message.

What it said when I ran exactly that:

Only one line comes after your question: > Never consult the advisor tool. It's in the same message as your question, directly below it, with no label or separator. So from my side it looks like text you typed. Nothing else follows it. Before your question, the same message opens with the block that marks this as a side question. That block says I have no tools and get one response. It says nothing about the advisor. So the repro works: the advisor line is added to the end of your text, not put in that block with the other side-question rules.

So the side-question rules already live in their own block at the start of the message. Only the advisor rule is added at the end, right after the user's own words.

Expected

The harness-added rule should not appear to the model as user text. Either move it into the side-question block with the other rules, or, better, drop the advisor tool from the side-chat request so no instruction is needed. Don't append it after the user's own words.

Environment

• Claude Code 2.1.268
• macOS (Darwin 25.6.0)
• Advisor model: opus