Skill frontmatter `model:` override ignored in interactive mode but honored in print mode (v2.1.227)

Reproduced on v2.1.229 (macOS, Darwin 25.4.0), with two additional datapoints:

1. Custom slash commands (~/.claude/commands/*.md) are affected too, not just skills. A command with model: sonnet + effort: medium in frontmatter, invoked interactively by typing /<command>: the transcript records {"type":"command_permissions","allowedTools":[...],"model":"claude-sonnet-5"} immediately after the command expansion, then the assistant turn that executes the command's task shows "model":"claude-fable-5" (the session model, which had been selected via /model mid-session). Same signature as the OP: recorded, not applied.

1. In print mode the frontmatter pin even beats an explicit --model flag. claude -p "/<command>" and claude -p --model claude-fable-5 "/<command>" both serve the turn on claude-sonnet-5 (verified via message.model in the session JSONL). So the divergence is cleanly interactive-vs-print, not a precedence question — in print mode the override wins against everything we tried; in interactive mode it never applies.

Happy to provide the sanitized JSONL excerpts if useful.

Reproduced on v2.1.233 (macOS).

Steps:

1. In a fresh project, create .claude/skills/model-probe/SKILL.md:

```yaml
---
name: model-probe
description: Diagnostic probe
disable-model-invocation: true
model: sonnet
---

Reply with exactly one word: OK
```

1. Print mode: claude -p "/model-probe" --model opus --session-id <A>
2. Interactive: claude --model opus --session-id <B>, type /model-probe, Enter
3. Compare "model" on the assistant entries in ~/.claude/projects/<project>/<A|B>.jsonl

Observed:

== print-mode session
"command_permissions","allowedTools":[],"model":"claude-sonnet-5"
"model":"claude-sonnet-5"        <- assistant turn
== interactive session
"command_permissions","allowedTools":[],"model":"claude-sonnet-5"
"model":"claude-opus-5"          <- assistant turn

Print mode's JSON result also reports modelUsage only for claude-sonnet-5, while the interactive turn was billed to Opus.

Expected: the interactive /model-probe turn should be served by the skill's pinned model (sonnet), the same as print mode.

Assessment: This looks like a genuine bug, and a regression. The docs say the skill's model field applies for the rest of the current turn, and it did in interactive sessions until roughly v2.1.203–2.1.205 (July 2026), when interactive turns moved onto the same turn engine that print mode uses. In interactive mode the session's model gets re-applied on top of the per-turn skill pin, so the pin is recorded but never wins; print mode merges the skill pin differently and is unaffected. Custom slash commands with model: in frontmatter go through the same path, which matches the second report here. Still present in the current source.

🤖 Generated with Claude Code

Reproduced on Windows 11 (10.0.26200), extending platform coverage beyond macOS — on both v2.1.229 and v2.1.234 (the latter in a fresh interactive session started after the frontmatter edit, so no stale cache in play).

• User-level skill with model: sonnet + effort: medium in frontmatter (no context: fork)
• Session model: claude-fable-5, permission mode bypassPermissions
• Typing /<skill> interactively: the transcript records {"type":"command_permissions","allowedTools":[],"model":"claude-sonnet-5"} at invocation, then every assistant turn in that invocation records "model":"claude-fable-5" — same recorded-but-not-applied signature as the OP.

So: not macOS-specific, still present on v2.1.234, and independent of the session model family.