What it is for
The queue of what the AI proposes. It runs overnight, looks at the last 24 hours, and nothing it does changes data: everything becomes a pending proposal, and approval belongs to a person.
What it proposes
- Merging errors that are the same. It looks at groups that seem to be the same bug under different fingerprints and proposes the grouping rule that merges them. Approving creates the rule.
- Whose fault it is. It classifies new errors as ours, the customer, a third party, or the environment. Approving fills in the category.
The AI gets things wrong. A wrong grouping rule merges two distinct bugs and makes one of them disappear from the list, and nobody finds out. A pending proposal costs one click and does not cost a lost bug.
What each button does
Every card says it before the click. Approving a merge creates a grouping rule: from then on, occurrences matching its conditions count as a single error. Approving a classification only sets the error's category, a label for filtering; status, owner and priority do not change. Rejecting either one changes nothing and archives the proposal for good.
Classifications can be approved all at once, from the button in the queue header. Merges go one by one: a wrong rule joins two different bugs, and that deserves a read.
How to turn it on
Turn on an analyst under Parts: Claude, OpenAI or Ollama (local, if the error cannot leave the building). With no analyst on, this screen stays empty and the job says so.
The Run now button does not wait for the night.
Being told when a new proposal arrives
The menu shows the pending count. Beyond that, the overnight run publishes an
insight saying how many proposals showed up, which goes out through the
channel you already use. One notice per run, not one per proposal: a night can
produce forty.
The same proposal does not come back
Each proposal is signed by its content, not by the text the model wrote. Rejecting a proposal stops it from coming back tomorrow in different words.