token-saver

token-saver

This skill guides and performs the installation and configuration of tools optimized to reduce token consumption (input/output) and context compression in coding LLMs (Claude Code, Codex, OpenCode, Agy).

The configurable tools are:

1. rtk (Rust Token Killer): CLI command proxy that filters redundant outputs from builds, lints, tests, and git, generating 60-90% savings in local operations.
2. caveman: Ultra-compressed communication mode that reduces ~75% of output token consumption by speaking in a telegraphic style (without linking verbs or formalities) without losing technical precision.
3. graphify: Codebase mapping into a structured knowledge graph. Allows the LLM to query relationships between files and god nodes without linearly scanning or cross-reading files.
4. context-mode: MCP/Plugin that intercepts giant outputs from tools (APIs, long logs, dumps) and stores them in a local FTS5 database, returning only compressed summaries to the active context window.

---

How to Calculate Token Savings for Each Tool

1. rtk:

   • Command: rtk gain
   • What it shows: Report of tokens saved in intercepted commands.
   • Command with history: rtk gain --history

2. caveman:

   • Calculation: Compare character/token counts generated in normal mode vs caveman mode.
   • Example: A 400-token response drops to ~100 tokens. Typical saving is ~75% on output tokens.

3. graphify:

   • Saving: Reduces input token consumption when searching the codebase.
   • Instead of reading 10 related files via grep/search (e.g., 20k-50k tokens), the LLM queries the graph in graphify-out/ directly, consuming less than 1k tokens.

4. context-mode:

   • Saving: Prevents context overflow from console dumps/logs.
   • A 50k-token build log is summarized to 1k tokens in the active window, while keeping the full log searchable via the local SQLite database. Typical saving of 90-98% on long logs and extensive outputs.

---

Skill Execution Flow

When this skill is activated:

1. Ask User (Iterative): Ask the user for each tool if they want to install/configure it, and the scope of the installation:

   • rtk: Install? [y/n] -> Scope: [g]lobal or [p]roject?
   • caveman: Install/Activate? [y/n] -> Scope: [g]lobal or [p]roject?
   • graphify: Install/Configure? [y/n] -> Scope: [g]lobal or [p]roject?
   • context-mode: Install/Configure? [y/n] -> Scope: [g]lobal or [p]roject?

2. Run Setup Script: Run the helper script skills/token-saver/scripts/setup-token-saving.sh (with BypassSandbox when necessary) to perform actions based on the user's selection.

3. Confirm and Teach: Display a summary of what was configured and quick usage instructions (command cheat-sheet) for each activated tool.