Representation
Compare source and output with one tokenizer. Record read mode, omissions and recovery to measure input reduction.
Benchmarks & methodology
Measure the input.Verify the outcome.
Compare input, runtime effort and task quality. Use a reproducible baseline to evaluate LeanCTX on your own work.
Three different questions
Measure input, effort and outcome.
Judge the context representation, the cost of preparing it and the completed task separately.
Operation
Count preparation time, tool overhead, model input and output, retries and cache effects. Separate observed cost from estimates and unknowns.
Task outcome
Use the same task and acceptance criteria. Tests and review show whether the delivered context was enough.
Your repository
Benchmark your repository.
Compare context representations on your files. Save the result with the runtime version and repository revision.
lean-ctx --version
lean-ctx benchmark run . --json
lean-ctx benchmark report .These commands inspect your project. They do not run a controlled end-to-end agent experiment or establish answer quality.
From diagnostics to a decision
Make it reproducible.
- Freeze the baseline
- Pin repository revision, task, model, tool setup and evaluation criteria. Document which calls LeanCTX covers.
- Include the whole attempt
- Count recovery reads, retries, review time and failures. Separate context savings from provider billing and cached-input discounts.
- Publish the scope
- State the workload, sample size, version and measurement method. Report variation and unsuccessful runs alongside the headline.
Read the evidence
Test quality against the source.
Preserved code structure is useful evidence, but it does not establish a correct answer or a safe change. Verify the result against the original material.
Start with LeanCTX
Put context to work.
Set up LeanCTX for your tool and evaluate the difference on a real task.
Local-first. Model-agnostic. Your context.