TrueMend is a code scanner for software written with AI assistance. It parses a codebase locally, scores every function, and reports what a careful reviewer would have caught: unchecked model output, error handling that hides failures, placeholder logic left in place.
Roughly a third of what it finds it fixes on its own, with a verified diff. The rest comes back priced, as a bounded refactor or as a decision for a person, so the backlog reads as a budget rather than a list.
Nothing is uploaded to run it. The same engine works from the terminal, in CI, and inside Claude Code, Cursor and Copilot, so one standard covers every team and every agent.
Four things that go wrong when nobody reads the code.
Each one is a pattern in the library, caught at the line, on the first run.
The catch block that hid an outage.
The retries kept failing. The catch block caught every one, rethrew nothing useful, and the dashboard stayed green for most of a day.
The engine names the wrapper at its line on the first run. The deterministic fix removes it with a diff and a re-parse, so the next failure is a failure you can see.
Cleanup ate the sprint.
The agent shipped the feature in three days. The rest of the sprint went to what it left behind, and nobody could say in advance how much that was.
The report prices the backlog before the sprint is planned: a third applied for free, the rest as bounded refactors and decisions, each with a cost attached.
The reply that became a row.
The tool returned JSON. The handler parsed it and wrote it straight to the database. One malformed reply later, the vendors table held a row nobody had asked for.
The engine flags every place a completion or tool result is used as truth, high severity, at the line, inside the review where it can still be caught.
Two teams, two standards.
One team reviewed by hand. The other let its agent merge. The agents ran on different prompts, and the same mistake passed in one repository and failed in the next.
One engine runs in the terminal, inside every supported agent and in CI, so a finding carries the same name, severity and tier wherever it appears. Nobody negotiates a second standard.
Six things it does, and the figure behind each one.
Pick a tab, or step through them in order.
15 head-to-head benchmarks. 14 wins, 1 loss, all published.
The audit has to hold up. Judged blind, against each language's own linter. The number is how often a finding does not hold up. Shorter bar is better.