rtk

Imported from servitola/dotfiles (rtk/AGENTS.md). Install upstream with npx skills add servitola/dotfiles --skill rtk. Copyright stays with the author.

rtk/ — Rust Token Killer config

RTK is a CLI proxy that filters/compacts command output before it reaches the LLM context (60–90% token savings on dev ops). Installed via Homebrew.

What lives here (version-controlled)

• config.toml — global RTK config (mirrors built-in defaults; tune here).
• filters.toml — custom per-command output filters (python3, hammerspoon, launchctl, docker-compose, git ls-files, …). Add new noisy commands here.

Both are symlinked by the Makefile into ~/Library/Application Support/rtk/.

What is NOT tracked (runtime data, lives only in Application Support)

• history.db — per-command savings analytics (large, ephemeral).
• tee/ — captured stdout/stderr of failed commands.
• .DS_Store, .hook_warn_last.

How RTK is wired

• Auto-rewrite: a Claude Code PreToolUse hook (rtk hook claude, see claude-code/settings.json) rewrites Bash commands transparently (git status → rtk git status), including &&-chains. No manual prefixing.
• Docs for the AI: claude-code/RTK.md (imported into the global CLAUDE.md).

Adding a filter

Edit filters.toml, then rtk config to confirm it loads. A filter matches a command by regex (match_command) and strips/caps its output. Find candidates with rtk discover (missed savings) — the top "unhandled" commands are the ones worth a filter.

Useful meta commands

• rtk gain — token savings so far (global).
• rtk discover — commands still bypassing RTK + est. savings.
• rtk learn — recurring CLI corrections from history (see note below).

Note on rtk learn

For this user it currently yields 0 recurring rules — corrections are overwhelmingly unique one-off heredocs, not repeated mistakes. The real lever is turning recurring inline heredocs into named scripts in zsh/bin/ (see docs/commands.md), so they stop being opaque one-liners RTK can't compact.