[BUG] Skill frontmatter `model:` field documented but not working

Summary

The official documentation at https://code.claude.com/docs/en/skills lists model as a supported frontmatter field:

| model | No | Model to use when this skill is active. |

However, when a skill specifies model: opus in its frontmatter, the model is NOT switched when the skill is invoked. The skill executes with the session's current model instead.

Steps to Reproduce

1. Create a skill with model: opus in frontmatter:

---
name: test-opus-skill
description: Test skill that should use Opus
model: opus
---
This skill should execute with Opus model.

1. Start Claude Code with default model (Sonnet)
2. Invoke the skill with /test-opus-skill
3. Check debug logs - no model switch occurs

Expected Behavior

When skill with model: opus is invoked, Claude Code should:

1. Switch to the specified model (opus)
2. Execute the skill
3. (Optionally) return to previous model

Actual Behavior

The model: field is completely ignored. Skill runs with whatever model the session is using.

Evidence from Debug Logs

• Debug log shows NO claude-opus-* API calls when skill with model: opus is invoked
• Only haiku calls (for system tasks like summarization) are visible
• No "model changed" warning appears in logs
• The only time "model changed" appears is when user manually switches via Alt+P/Cmd+P

Severity: HIGH

This is a documentation/implementation mismatch that misleads users into believing they can control model selection per-skill. Users may be:

• Unknowingly using more expensive models than intended
• Not getting the reasoning capabilities they expected for complex skills
• Building skill libraries based on documented-but-non-functional features

Related Issues

• #7795 - Feature request for this exact functionality (suggests it was never implemented?)
• #19269 - Requests model routing, author notes "may not be implemented"

Environment

• Claude Code 2.1.23
• macOS Darwin 25.2.0

Question for Maintainers

Is model: in skill frontmatter:

1. A documented feature that's broken (bug)?
2. A planned feature that was documented prematurely (docs bug)?
3. Working but requires specific conditions we're missing?