What makes it different
What zerd does differently from most coding agents, from reviewing the real diff to checking every file path it cites.
| Most agents | zerd |
|---|---|
| Say "done" | Run checks on the changed files, then have a different model review the real diff |
| Show every review finding | A verifier tries to refute each one; high-severity claims are reproduced in a sandbox or capped |
| Quietly invent a file or line | Every path:line in the answer is checked against your tree |
| Re-send the whole file every turn | Remember what was read; send only new lines or a diff |
Ask the prompt to behave ("never read .env") | Enforce it in code: confined tools, a path-scanning shell, an OS sandbox |
| Die when a provider's quota runs out | Fail over to another provider and tell you |
| Forget what you told them | Turn your corrections into lessons in your own AGENTS.md |
| Break an old caller while "refactoring" | Signature guard sends it back if a public parameter disappeared |
| Hand a weak model the same rails as a strong one | Repair what open models get almost right: Read becomes read_file, old_str becomes old_string, a stringified edit list is parsed, a cut-off reply is continued |
| Find a type error at the end of the run | Per-edit diagnostics: the project's own tsserver, ruff, gofmt, node --check report only the problems that edit introduced |
And one rule under all of it: every token and every tool call should earn its place.