Engineering
Recognize a problem you fixed before
A customer reports a failure that resembles an earlier incident.
Every run searches your own past diagnosed work first, and every approval and rejection is kept, so the second time starts further along. The ones that keep coming back become known issues written from real diagnoses, not an article somebody has to maintain.
How Novix gets involved
Every run, at the memory stage.
The third stage of every run, before a model has been asked anything expensive. There is nothing to enable: the corpus is your own past diagnosed work, and it grows on its own.
How the agent handles it
No model, and the corpus is deliberately narrow
MemorySearches your own past diagnosed work for a similar incident.A distinctive-weighted match over your own past diagnosed tickets, using what the reporter actually sent. Live context is excluded on purpose: connectors attach the same recent commits to every ticket, and that boilerplate used to decide the score.
Two bars, not one
Showing prior art is a soft suggestion. Calling two tickets near-duplicates is a claim that two customers hit the same bug, so it has to clear a higher threshold and evidence floors underneath it.
Did the fix actually work last time
The evidence summary records whether any match was a fix the customer confirmed, and which affected areas this diagnosis shares with a past resolved one. A missing signal is an honest empty row, never a hidden one.
It checks whether a fix regressed
A new defect is compared against tickets fixed in the last 60 days. A strong match is flagged as a suspicion, never an action: a person confirms or dismisses it, and a confirmed one carries a caution into future diagnoses of that area.
Known issues write themselves
A defect diagnosed confidently enough to be worth patching also drafts a plain-words known issue, once per source ticket. A person edits it and a manager publishes it before anybody outside can read it.
What your team gets
- Prior art on the ticket, with how similar and why
- A regression flag when something looks like it came back
- A known-issue draft written out of a real diagnosis
It would rather miss a duplicate than invent one
Every threshold is set for precision, because a duplicate link is a claim made to a person and other features build on it. And there is no article editor: known issues are a byproduct of fixing the bug, not a writing job.
Try it on a real customer problem.
Connect the place reports or product signals already arrive. The same agent takes it from evidence to a tested pull request.