Sample NDepend Web Reports
Reports of NDepend.AI Fix Runs
▶ Watch NDepend.AI (5 min)Eight runs on eight open-source code bases fixed by NDepend.AI driving Claude Code (Opus as the premium model, Sonnet as the economy model). Each report opens on its AI Assistant tab: the complete log of the run, from the planning pass to the fix rounds, the parallel agent sessions, the tests, the change review and the final report, next to the NDepend report of the fixed code, diffed against the code at the start of the run.
| Code base | Issues fixed | Duration | Cost | Coverage | Tests (+written) | Bugs found |
|---|---|---|---|---|---|---|
| eShopOnWeb | 468 / 493 95% | 1h 21min | $44 | 45% → 97.5% | +506 619 green | 3 |
| NopCommerce (large scale fix) | 3,711 / 3,781 98% | 5h 23min | $261 | 23% → 43.6% | +3,476 4,580 green | 8 |
| Jellyfin (large scale fix) | 1,514 / 1,644 92% | 11h 23min | $543 | 31% → 56.2% | +4,303 8,017 green | 17 |
| Flow Launcher (large scale fix) | 2,837 / 3,109 91% | 6h 17min | $294 | 11% → 60.8% | +2,744 3,236 green | 5 |
| Spectre.Console (large scale fix) | 967 / 1,071 90% | 1h 31min | $64 | 71% → 98.5% | +1,507 2,747 green | 7 |
| Microsoft.FeatureManagement | 283 / 292 97% | 1h 06min | $25 | 78% → 99.6% | +200 347 green | 1 |
| Carter | 222 / 223 99% | 39min | $13 | 68% → 99.6% | +107 271 green | 0 |
| NodaTime | 354 / 387 91% | 1h 26min | $48 | 98% → 99.4% | +313 23,029 green | 0 |
The figures are those of each report's final report:
- Issues fixed: the issues of the targeted set that NDepend's re-analysis finds fixed.
- Duration: the wall-clock duration of the run.
- Cost: the cost billed by the agent's provider.
- Coverage: the code coverage before and after the run.
- Tests (+written): the tests the run wrote and the tests green at the end.
- Bugs found: the potential bugs the agent reported in the application code while writing tests and fixing issues. The run leaves them unfixed and lists each one in its final report, to check by hand.
NDepend.AI Fix Runs with Eight Coding Agents
The same run on a small experiment code base TrainTrain, driving eight different coding agents: the same 22 issues to fix (architecture, naming, visibility, code coverage and a few issues to suppress on purpose), the same NDepend verification of every fix. Each report opens on its AI Assistant tab, with the complete log and the final report of the run.
| Coding agent | Issues fixed | Suppressed | Left on purpose | Still to fix | Duration | Cost | Coverage | Tests (+written) | Bugs found |
|---|---|---|---|---|---|---|---|---|---|
| 14 / 22 64% | 3 | 3 | 2 | 6min 35s | $1.65 | 69% → 77.4% | +7 29 green | 2 | |
| 13 / 22 59% | 4 | 3 | 2 | 5min 44s | ~$0.22 * | 69% → 77.2% | +7 29 green | 1 | |
| 14 / 22 64% | 4 | 3 | 1 | 16min 38s | ~$1.63 * | 69% → 78.7% | +15 37 green | 0 | |
| 14 / 22 64% | 3 | 3 | 2 | 9min 10s | $2.11 | 69% → 77.2% | +6 28 green | 2 | |
| 13 / 22 59% | 4 | 3 | 2 | 23min 28s | 18.8 credits | 69% → 77.2% | +12 34 green | 0 | |
| 15 / 22 68% | 4 | 3 | 0 | 19min 52s | $2.98 | 69% → 79.2% | +15 37 green | 0 | |
| 13 / 22 59% | 4 | 3 | 2 | 12min 01s | $5.51 | 69% → 77.2% | +12 34 green | 0 | |
| 13 / 22 59% | 4 | 3 | 2 | 39min 18s | $0 * | 69% → 89.4% | +17 39 green | 2 |
The figures are those of each report's final report:
- Coding agent: under its name, the premium model / the economy model of the run.
- Suppressed: the issues hidden by a SuppressMessage attribute the agent added, not fixed and left out of the fixed figures.
- Left on purpose: the issues the agent declined with a reason, a human decision. In all eight runs, the same 3 methods
left untested:
GetTrain(),BookSeats()andGetBookingReference()create their ownHttpClientand call a remote HTTP service, with no way to pass in a fake: they need an integration test, not a unit test. - Still to fix: the issues left, branches the tests found unreachable, dead code to delete by hand (for Kiro, 2 coverage issues not reached, to retry in another run).
- Cost: * Codex and Copilot CLI report no cost of their own: the run's tokens priced at the published rates (partial for Copilot CLI). Kiro bills in credits. OpenCode ran a free model.
Reports Compared with a Baseline (usual NDepend reporting, no AI)
For these sample reports are imported:
Such serverless report can be generated by invoking... NDepend.Console.exe $absolute-path-to-an-NDepend-project-file$...from your CI/CD running on Windows, Linux and Mac
You can generate the report using a trial license, a Build-Machine license |
|