TokenRoll Claude Code Plugin

llmdoc + SubAgent RAG: Solve the Context Floor Problem

English | 简体中文

---

The Problem: Context Floor

In serious production environments, AI Coding Agents face a fundamental challenge: they don't truly understand your codebase. They achieve understanding through CLAUDE.md + massive code file reading, which leads to:

• Endless tool calls before reaching sufficient context
• High token consumption with low information density
• Slow Time to Context Ready (TTCR)

We call the "minimum context richness required for an Agent to solve a task" the Context Floor.

Existing Solutions Fall Short

Approach	Tool Calls	Token Usage	Info Density	Effectiveness
LSP MCP	High	Medium	High	Good, but slow
ACE / RAG	Low	Low	Sparse	Poor correlation
Agentic RAG (Explorer)	Medium	Low	High	Good, but TTCR too slow

Our Solution: llmdoc + SubAgent RAG

Fast. High-density. Low main-agent token usage. Strongly correlated with tasks.

llmdoc

A documentation system designed from the ground up for AI to quickly acquire high-density information while remaining human-readable.

Based on Diataxis, optimized for LLM retrieval:

llmdoc/
├── index.md          # Entry point - always read first
├── overview/         # "What is this project?" - MUST read all
├── guides/           # "How do I do X?" - step-by-step instructions
├── architecture/     # "How does it work?" - LLM retrieval map
└── reference/        # "What are the specifics?" - API specs, conventions

Key Design Principles:

• Leverages Agent's ability to batch-read files quickly
• Documents retain critical file paths + module descriptions
• Project overview + architecture + topic-linked guides + references

Example: TokenRoll/minicc/llmdoc

SubAgent RAG

Two primary functions:

1. Investigation: Based on llmdoc + existing code, investigate decomposed tasks as prerequisites
2. Recording: After completing coding tasks, automatically maintain llmdoc

---

Quick Start

Step 1: Install Plugin

# Add TokenRoll plugin marketplace
/plugin marketplace add https://github.com/TokenRollAI/cc-plugin

# Install tr plugin
/plugin install tr@cc-plugin

Step 2: Configure System Prompt

Copy the contents of CLAUDE.example.md into your ~/.claude/CLAUDE.md file.

That's it. Once configured, all behaviors activate automatically:

• Agent will always read llmdoc first before any action
• Investigation uses documentation-first approach
• After coding tasks, Agent will ask if you want to update docs
• All skills trigger automatically based on context

Update Plugin

/plugin marketplace update https://github.com/TokenRollAI/cc-plugin

---

How It Works

Automatic Behaviors (No Commands Needed)

Once CLAUDE.example.md is configured, these behaviors are always active:

Behavior	What Happens
Documentation First	Agent reads llmdoc/ before any action
Smart Investigation	Uses investigator agent instead of generic exploration
Option-Based Coding	Never jumps to conclusions; presents choices via questions
Doc Maintenance Prompt	After coding, asks if you want to update documentation

Available Skills (Auto-Triggered)

These skills activate automatically based on your prompts:

Skill	Triggers	Description
/investigate	"what is", "how does X work", "analyze"	Quick codebase investigation
/commit	"commit", "save changes"	Generate commit message
/update-doc	"update docs", "sync documentation"	Update llmdoc
/read-doc	"understand project", "read the docs"	Read llmdoc overview

Commands (When You Need Control)

Command	Description
/tr:initDoc	Initialize llmdoc for a new project
/tr:withScout	Complex tasks: deep investigation first, then execute
/tr:what	Clarify vague requests with structured questions

---

Recommended Workflow

For New Projects

# Initialize documentation system
/tr:initDoc

For Daily Development

Just talk naturally. The system handles the rest:

"How does the auth system work?"
# -> Auto-triggers /investigate, reads llmdoc first

"Add a new API endpoint for user profiles"
# -> Reads llmdoc, investigates, implements, asks about doc update

"commit"
# -> Auto-triggers /commit with intelligent message

---

Cost & Effectiveness

Honest assessment: This approach costs approximately 1.5x more to achieve a jump from 85 to 90 points in task completion quality.

• Simple projects: Marginal benefit
• Complex projects: Significant benefit
• Production codebases (100k+ lines): Excellent results

In our production backend (100k lines of code):

• Task completion cost: $1-5 per feature
• Human intervention: Significantly reduced
• Output quality: Ready for review and minor adjustments

---

Internal Agents

Agent	Purpose
worker	Execute well-defined plans with precision
investigator	Rapid, stateless codebase analysis
recorder	Create and maintain llmdoc documentation
scout	Deep investigation for initDoc

---

Made with care by DJJ & Danniel for the TokenRoll team