Entroly: measure how much coding-agent context your repo can discard locally

[juyterman1000]

I built Entroly to reduce the amount of repository context sent to coding-model providers without hiding the files that matter.

The core is local-first: Entroly indexes and selects context on-device under a token budget. If you proxy a cloud model, the selected prompt still goes to that provider; Entroly reduces what is sent rather than adding another destination.

What the existing release benchmarks show

Measured input-token reduction is workload-dependent. The committed benchmark artifacts currently show the largest reductions on large-context workloads, while short prompts and tasks that already fit the budget correctly save less or pass through unchanged.

• Accuracy-retention table and reproduction commands: README benchmarks
• Committed benchmark artifacts: benchmarks/results
• Architecture details: docs/DETAILS.md

Try the local smoke check

pip install entroly
cd /path/to/your/repo
entroly verify-claims

This writes a bounded .entroly_verification.json smoke report with sampled token savings, indexing speed, and file coverage. For a real ROI estimate, run a representative proxy pilot and inspect the dashboard over your actual workload.

Feedback wanted

I am looking for adversarial tests and real pilot workloads, especially:

• repositories where relevant files are difficult to retrieve
• coding-agent tasks where context selection loses an important dependency
• small-repo workloads where compression should deliberately pass through
• integration friction with a specific agent and version

Disclosure: I built Entroly. It is Apache-2.0 licensed. Please report concrete failures or reproducible edge cases rather than treating the benchmark numbers as universal promises.