No visible notice when auto mode ignores a skill's `model:` override

Summary

In auto mode, a skill or command whose model: names a model that auto mode doesn't support (for example model: haiku on the Anthropic API) runs on the session model, and nothing tells the user. The fallback itself is intended: it dates from 2.1.259 and the skills frontmatter reference describes it. This asks only for it to be visible. Today the only trace is a debug-log line, so a skill author who picked a model for cost or speed can't tell it was ignored without checking modelUsage by hand.

Environment

• Claude Code 2.1.283, macOS
• Anthropic API, signed in with a Claude account; session model Claude Opus 5.5
• Seen in print mode (-p) and in an interactive session

Steps to reproduce

1. In a scratch project, add .claude/commands/probe.md:

```markdown
---
description: Replies with a marker
model: haiku
---

Reply with exactly PROBE-OK and nothing else.
```

1. Run it in manual mode, then in auto mode:

``bash
claude -p "/probe" --output-format json --permission-mode manual
claude -p "/probe" --output-format json --permission-mode auto --debug-file auto.log
``

1. Compare the modelUsage keys in the two results, and search auto.log for not supported in auto mode.

Expected

The auto-mode run says, in its output or the session transcript, that the command's model wasn't used and the session model ran instead.

Actual

• Manual mode: modelUsage shows claude-haiku-4-5-20251001.
• Auto mode, print: modelUsage shows only the session model, and neither the output nor the transcript mentions the switch.
• Auto mode, interactive: the same fallback, and nothing on screen says so. The footer shows only auto mode on.
• The same happens when the model invokes a skill with model: haiku through the Skill tool.
• In each auto-mode case the debug log has:

``
Skill/command model "claude-haiku-4-5-20251001" is not supported in auto mode; keeping the session model
``

Suggestions

1. Show that line as a visible notice. There's a precedent: 2.1.223 added a warning when a forked skill's or slash command's requested subagent model is restricted and the parent model runs instead.
2. Consider flagging it once when skills load, since on the Anthropic API a skill with model: haiku can never run on Haiku in auto mode. The notice could point to context: fork with background: false, which does run the skill on its declared model in auto mode, as the frontmatter reference says.

Related

#85658 and #84262 report skill model: being ignored in other circumstances, not auto mode.