Find the code worth improving
Follow a finding back to the code. Improve the work around it.
Make the codebase easier for agents to work in. Start with one finding, inspect its cause, and check the behavior you preserve.
Start with the latest scan
Open Code health. Check its revision and whether the relevant analyzer completed. Open a finding to inspect the affected file or function in Code explorer.
Choose a finding related to real work: a function that is repeatedly hard to change, a dependency boundary that causes confusion, or a check that catches actual defects. Keep the behavior checks attached to the change.
Investigate one finding
- Open its evidence and relevant code. Understand the responsibility behind the number.
- Propose one useful change: a clearer boundary, a smaller function, or a better verification recipe.
- Compare the before/after code and pass the same behavior checks. Read the related agent work too.
Scan profiles and coverage
Setup queues a full working-copy baseline and a bounded structural lookback: the latest 10 commits plus up to 20 earlier samples across the repository's lifetime.
Structural scans measure size, function complexity, and duplication locally. Full scans add applicable compiler, lint, unused-code, React, dependency, and smell analysis.
Read each analyzer's state. A missing tool or dependency can block a check. Open its reason, restore the prerequisite, and retry. The rest of a scan does not turn that missing result into zero.
Analyzer reference
| Analyzer | Evidence |
|---|---|
| cloc | Code, comment, and blank-line counts |
| Fallow | Complexity, function size, duplication, and structural evidence |
| Knip | Unused file, export, and dependency candidates |
| Oxlint | Lint rules and locations |
| TypeScript | Compiler diagnostics |
| React Doctor | React diagnostics |
| Dependency Cruiser | Edges, cycles, and boundary findings |
| Qlty | Independent metrics and code smells |
Full scans may need installed project dependencies or additional tools. Qlty is a standalone executable; the app repository documents its supported installation.
History, working copy, and ongoing collection
Historical scans reconstruct structural measurements from Git revisions now. They cannot recover agent feedback, runtime, or tokens that were never recorded. Full historical checks need an appropriate environment at that exact revision.
The working-copy scan includes current uncommitted code. Keep it distinct from committed history. Today's dependencies are not automatically valid for older commits.
Collection runs while the tracker is running. In Settings, inspect new-commit collection and its scan profile. In Code health, run a full scan when you need current diagnostic results.
Read the evidence → · Investigate a missing result →
Inspect the checks your project declares
Open Code health → Checks to see tests, lint, type checks and builds declared in package scripts, CI and literal verification blocks in project instructions. Inspect the source, saved result, exit, duration, commit and run time. A check without saved evidence shows Not run.
Timing series show median and p90 within matching recipe, scope, environment, machine and local/CI context. Flaky observed requires serial clean repeats on the same commit with different outcomes. Dirty or legacy runs without clean snapshots remain Insufficient runs. No disagreement observed describes only that sample. It is not a reliability score.