shadcn improve: Split Code Audits and Execution Across Models

What Happened

shadcn released improve, an open-source tool that separates codebase auditing from code execution across two model tiers. The tool routes a full read of the repository to the strongest available model, which produces a structured audit and a set of execution plans. Those plans are then handed to cheaper models to implement, with the expensive model reserved for analysis and planning rather than generation. The repository is available at github.com/shadcn/improve under shadcn's GitHub account.

Why It Matters

Most agentic coding workflows default to a single model for the entire loop — reading files, reasoning about changes, and writing code — which means the most expensive tokens are spent on the highest-volume operations. improve inverts that: it treats the audit and plan as a compressible artifact that can be produced once by a frontier model and consumed many times by cheaper ones. This codifies a routing pattern that operators have been assembling manually for months, moving it from tribal knowledge into a reusable tool. The immediate beneficiaries are teams running agent loops against mid-to-large repositories where a full audit is expensive but the edit surface is narrow. It also creates a clean seam for cost accounting: audit spend and execution spend become separately measurable line items.

Technical Details

The architecture is a two-stage pipeline. Stage one reads the codebase and emits an audit plus a plan file — the plan is the interface contract between stages, meaning execution models do not need access to the full repository context, only the plan and the files it targets. This reduces context window pressure on execution models and lets operators swap execution backends without re-running the audit. The pattern assumes the auditing model has sufficient context capacity for the repository or the relevant subtree; very large monorepos will require scoping the audit, which reintroduces the context management problem one level up. Integration is git-oriented — plans and audits can be versioned alongside the code, making the routing decision auditable after the fact. No published benchmarks accompany the release; the cost claims are architectural rather than measured.

Operational Impact

Day-to-day, operators gain a stable plan artifact that can be reviewed, stored, and replayed against execution models as they change — swapping execution providers no longer requires re-deriving intent from scratch. Cost profiles shift: audit spend becomes a fixed, infrequent cost while execution spend scales with edits and can be driven down independently through cheaper models or local inference. Review workflows change shape, since a human can inspect the plan before any execution model touches files, inserting an approval gate that single-model agents typically lack. The pattern also lowers the floor for running agents on self-hosted or lower-capability models, provided the plan is precise enough — which moves the engineering problem from model selection to plan quality. Teams currently paying frontier prices for high-volume edit loops should expect direct reductions; teams without plan-review discipline should expect execution drift when plans are underspecified.