Feature Request: Auto-apply skills (rtk, caveman)

Two related UX improvements that would make day-to-day usage much smoother.

1. Auto-apply rtk and caveman skills

Context

There are two very useful built-in skills that the user has to remember to invoke manually:

• rtk — RTK CLI wrapper that produces compact, token-efficient shell output (cuts ~75% of ls, tree, grep, git, npm, cargo, docker, kubectl, find, etc.). Currently triggered only when the user thinks to route a command through it.
• caveman — Ultra-compressed communication mode (lite / full / ultra / wenyan variants). Currently triggered only by explicit phrases like "caveman mode" / "use caveman" or slash command.

Both are high-signal triggers that are easy to forget, and both significantly reduce token / time cost when applied.

Proposed behavior

1. Context-aware auto-apply for rtk — When the model is about to run a shell command and the command matches the RTK-supported surface (file listing, search, git inspection, package manager reads, log greps, JSON inspection, env reads, dependency trees, container/k8s reads, etc.), auto-route through rtk (or its non-interactive equivalent) instead of the raw command. Show a one-line note in the response, e.g. ↳ rtk ls (auto-applied). Provide a way for the user to opt out per-command or globally.
2. Signal-aware auto-apply for caveman — Detect when the user is asking for brevity, hitting token-budget pressure, or sending prompts that match the existing trigger phrases, and switch the response style automatically. Show a one-line toggle indicator so the user knows the mode is active, with a single command to disable mid-session.

Why

• These two skills save a lot of tokens, but their value compounds only if they fire automatically. Forcing the user to remember to invoke them defeats the purpose.
• The signals for both are already well-defined in the existing skill descriptions, so detection is cheap and low-risk.
• Both should be opt-out-able per session so power users retain control.

Acceptance criteria

• A shell command that fits the RTK surface runs through rtk automatically, with a visible note. A --no-rtk per-command escape and a session-level rtk: off config both work.
• Saying "be brief" / "less tokens" / "caveman" (or any pattern currently in the skill description) flips the response style automatically and shows an indicator.
• No behavior change for users who never trigger these contexts.