Skip to main content

​
Overview

The codebase patterns demo shows ICRL’s ability to learn and apply your team’s specific coding patterns, conventions, and architectural decisions. Vanilla LLMs use generic patterns; ICRL learns from successful trajectories and applies your APIResponse wrapper, custom exceptions, and service layer conventions automatically.

​
The Problem

Every time you ask a vanilla LLM to add a new API endpoint, it uses generic patterns:
  • Generic error handling (raise HTTPException(...))
  • Standard response formats (raw dicts)
  • Default logging approaches
You end up repeatedly correcting it: “No, we use our custom APIResponse wrapper!”

​
How ICRL Solves This

After successfully completing 2–3 similar tasks, ICRL:
  1. Stores successful trajectories with your corrections baked in
  2. Retrieves relevant examples when you ask for similar tasks
  3. Applies your patterns automatically without re-prompting

​
Demo Structure

examples/codebase_patterns_demo
README.md
setup_demo.py
run_demo.sh
validate_patterns.py
mock_codebase
app
main.py
core
response.py
exceptions.py
logging.py
models
routes
services
tests
_reference
pyproject.toml

​
Custom Patterns (What ICRL Learns)

  1. Response wrapper — All endpoints return APIResponse[T], not raw dicts
  2. Exception handling — Use NotFoundError, ValidationError from app.core.exceptions, not HTTPException
  3. Service layer — Routes delegate to services; business logic lives in services
  4. Structured logging — Use get_logger(__name__) and consistent context

​
Running the Demo

This demo uses the ICRL CLI (icrl chat) interactively: 
cd examples/codebase_patterns_demo
uv run python setup_demo.py

cd mock_codebase

# 1. First task: Add orders endpoint (ICRL explores and learns)
uv run icrl chat
# Ask: "Add a GET /orders endpoint that returns a list of orders. Follow the existing patterns."
# Say 'yes' to save the trajectory

# 2. Second task: Add categories endpoint (ICRL applies learned patterns)
# Ask: "Add a GET /categories endpoint with CRUD operations"
# ICRL retrieves the orders trajectory and applies patterns immediately

# 3. Ablation: Compare with and without examples
uv run icrl run "Add a GET /items endpoint" --ablate

​
What to Observe

  • First task — ICRL explores users.py, products.py, response.py, etc., and creates the endpoint following exact patterns
  • Second task — ICRL retrieves the orders trajectory and applies correct patterns from step 1: fewer exploration steps, faster completion
  • Ablation — --ablate runs with and without retrieved examples and prints a comparison

​
Prerequisites

  • OPENAI_API_KEY or ANTHROPIC_API_KEY set
  • ICRL CLI installed (uv run icrl chat from project root)