Refact — Local-First AI Coding Assistant

Practical Notes

• Quant: the repo describes an IDE-embedded flow where the plugin runs a local refact-lsp engine per workspace.
• Quant: validate setup by running one agent task end-to-end and measuring time saved over 3 repeated tasks (same prompt, same repo).

What to standardize before rollout

Refact becomes much more valuable when teams standardize:

1. Provider policy: which providers are allowed for which repos (open-source vs private).
2. Default models: one for chat, one for agent work, one for embeddings if needed.
3. Task boundaries: define which actions require explicit approval (deps updates, migrations, deploy scripts).

Suggested first workflows

• “Explain module X” + “write a unit test for function Y”.
• “Refactor a file” with a measurable constraint (max 10 lines changed; no behavior change).
• “Fix a failing test” with reproduction steps and a time budget.

Use the same 2–3 workflows across the team so you can compare outcomes consistently.

FAQ

Q: Is it only for chat? A: No. The README positions it as an agent that can plan, execute, and iterate in engineering workflows.

Q: Do I have to use one provider? A: No. It supports multiple provider families; choose per your policy.

Q: How do I avoid risky changes? A: Define approval-gated actions and start with read/modify-only tasks before automating merges/deploys.