Restore model-invocability of the built-in /code-review (drop disable-model-invocation), or provide a first-class programmatic equivalent

Problem

The built-in code-review skill is marked disable-model-invocation, so a Claude Code session can no longer run /code-review max on its own — only a human typing the slash command can. Until 2026-07-25 the model could invoke it via Skill(skill="code-review"); we have 99 such autonomous invocations in our logs between 2026-06-30 and 2026-07-25, as the mandatory closing step of every implementation session.

For agentic development workflows this restriction inverts its own safety intent:

• The review it blocks is the safety step. Our methodology requires a fresh-context multi-agent review at the end of every implementation block, run by the implementing session itself before delivery. In one documented case, treating the step as "impossible" (right after the restriction landed) nearly shipped a release-blocking regression that the review then caught.
• It pushes users onto fragile workarounds. The only remaining in-session path is spawning a nested headless CLI (claude -p "/code-review max"), which we now wrap in a CI script. That path just cost us three days of debugging a headless-only silent deadlock (filed separately: #83270) — a failure mode that cannot even occur in the supported interactive path.
• The restriction is trivially bypassed anyway (any session can shell out to claude -p), so it does not actually prevent autonomous review runs — it only removes the supported, observable way of doing them.

Request

Any of the following would resolve it (in order of preference):

1. Remove disable-model-invocation from the built-in code-review skill (or make it configurable, e.g. a settings key or --allow-model-invocation code-review), so a session can run its own close-review again.
2. Expose a SlashCommand-style tool (or SDK API) that lets a session run built-in commands like /code-review max with the same rendering/routing as interactive use.
3. Document claude -p "/code-review …" as a supported invocation path — with output routed to the print-mode result reliably (see the linked bug), so wrappers don't have to reverse-engineer its behavior.

Context

Solo developer running an AI-first ERP codebase with a strict block-close ritual; /code-review max is a mandatory, self-run step in it. Happy to share workflow details or logs if useful.