setup-rtk

Token Optimization Setup

Configure RTK (Rust Token Killer) and comprehensive token optimization for Claude Code sessions.

Steps

1. Check RTK Installation

# Check if RTK is installed
if command -v rtk &>/dev/null; then
  echo "RTK installed: $(rtk --version)"
  echo ""
  rtk gain 2>/dev/null || echo "No savings data yet"
else
  echo "RTK is NOT installed"
  echo ""
  echo "Install options (the curl|bash pattern is BLOCKED by Claude Craft hooks):"
  echo "  1. (Recommended) make install-rtk    # from claude-craft root"
  echo "  2. cargo install rtk-cli            # if you have Rust toolchain"
  echo "  3. Download release binary manually: https://github.com/rtk-ai/rtk/releases"
fi

2. Configure RTK Optimizations

If RTK is installed, apply these optimizations:

a) Enable ultra-compact mode

Check the hook at ~/.claude/hooks/rtk-rewrite.sh. The rewrite command should use --ultra-compact:

REWRITTEN=$(rtk rewrite --ultra-compact "$CMD" 2>/dev/null)

If it doesn't have --ultra-compact, update the hook file.

b) Optimize RTK limits

Check ~/.config/rtk/config.toml and recommend these limits:

[limits]
grep_max_results = 100
grep_max_per_file = 10
status_max_files = 10
status_max_untracked = 5
passthrough_max_chars = 1500

c) Add custom filters

Check ~/.config/rtk/filters.toml. If it only contains template comments, suggest filters based on the detected project stack:

• Docker projects: Add docker exec, compose, logs filters
• Node.js projects: Add npm/npx install filters
• PHP projects: Add composer filters
• Python projects: Add pip install filters

3. Configure Sub-Agent Model and Forked Subagents

Check if both env vars are set:

echo "CLAUDE_CODE_SUBAGENT_MODEL=${CLAUDE_CODE_SUBAGENT_MODEL:-NOT SET}"
echo "CLAUDE_CODE_FORK_SUBAGENT=${CLAUDE_CODE_FORK_SUBAGENT:-NOT SET}"

If not set, recommend adding to ~/.bashrc (or ~/.zshrc):

# Use Sonnet 5 for sub-agents (exploration, grep, file reading) instead of Opus
# → 40-60% cost reduction on sub-agent invocations
export CLAUDE_CODE_SUBAGENT_MODEL="claude-sonnet-5"

# Run sub-agents in isolated contexts (Claude Code 2.1.117+, see COMPATIBILITY.md)
# → Avoids polluting the main context window with sub-agent intermediate state
# → Compounds with context: fork on skills (~8-15K tokens saved per long session)
export CLAUDE_CODE_FORK_SUBAGENT=1

# Enable 1-hour prompt cache TTL (Claude Code 2.1.108+)
# → -40% cost on repetitive sessions (BMAD sprints, /team:* loops)
# → Same prompt cache key is reused for up to 1h instead of 5min default
export ENABLE_PROMPT_CACHING_1H=1

# Force 5-minute cache writes on every turn (Claude Code 2.1.108+)
# → Useful for short-burst dev loops that hit the cache repeatedly
# → Trade-off: small write overhead, large hit-rate gains on iterative work
export FORCE_PROMPT_CACHING_5M=1