Skip to content

Runnable examples

Starter scripts live in examples/ in the repo. Each one writes to its own SQLite file so you can keep them side-by-side and compare context growth patterns.

Setup

uv add ctxrot python-dotenv
cp .env.example .env  # then fill in your credentials

For rlm_reasoning.py: also install Deno.

All examples load from a .env file in the project root:

MODEL=openai/gpt-4.1-mini   # any litellm model string
API_KEY=sk-...              # optional if OPENAI_API_KEY is already set
API_BASE=https://...        # optional, for custom endpoints

The examples

File DSPy module DB file What it shows
cot_simple.py ChainOfThought ctxrot-cot.db Hello world — single-call tracking
react_multihop.py ReAct ctxrot-react.db Multi-hop QA with Wikipedia search — context growth across iterations
custom_module.py custom Module ctxrot-custom.db Pipeline of 3 LM calls with growing prompts
rlm_reasoning.py RLM ctxrot-rlm.db REPL reasoning loop — most dramatic context growth
optimizer_bootstrap.py BootstrapFewShot + ChainOfThought ctxrot-optimizer.db Compile-time LM-call volume — one session per trainset example

Run one

# Pick any example
python examples/react_multihop.py

# Then view the dashboard
ctxrot --db ctxrot-react.db

Once you have a session to look at, try:

  • Local rot metrics — ctxrot analyze --db ctxrot-react.db
  • Export to JSONL — ctxrot export --db ctxrot-react.db --all -o react.jsonl
  • Deep analysis — ctxrot deep-analyze --db ctxrot-rlm.db (the RLM example tends to surface the most interesting rot patterns)

Next