Skip to main content
A deviation is a recurring type of misbehavior - for example “retrieval returned no sources”, “overstated a claim the sources don’t support”, or “skipped a required verification step”. Glassray finds concrete examples across many traces (instances) and clusters them into a deviation type, so you review a pattern once rather than triaging the same problem trace by trace. Each deviation is found against your spec - the generated description of what your agents are supposed to do - and is linked to the flow it affects.

​
Lifecycle

A deviation moves through a Linear-style lifecycle. Every transition is recorded with a short reason, so nothing changes status silently.
If you’ve connected Slack, Glassray posts each deviation to your channel - by default only once a deviation is Confirmed, so you’re not pinged for candidates that are still being investigated. You can lower that threshold to be notified as soon as a deviation is Suspected, so you can confirm or dismiss it right from the card.

​
Finding more examples

While a deviation is Suspected, Glassray is still building the case - scanning your recent trace history for more instances so you can see how often it really happens. Once it recurs in enough examples, Glassray promotes it to Confirmed automatically; until then it stays Suspected. You can also kick off this search yourself from the deviation.

​
What happens when you confirm

Confirming a deviation - promoting it to Confirmed, yourself or automatically - generates a fix:

A proposed fix

Glassray generates a code-level fix suggestion - a prompt or logic change - that you review and apply like any pull request. If Slack is connected, the deviation’s thread gets a “generating a fix” update.
The fix is a proposed diff you review - Glassray does not auto-merge it. You stay in control: review the suggestion, attach it to the failing trace as evidence, and ship it through your normal review process.

​
Where deviations come from

Most deviations come from Glassray’s own scan against your spec. But you can also surface them from outside the loop:

From feedback

A negative signal pinned to a trace is actioned - Glassray sorts that trace into a flow and creates or appends to a deviation.

From Slack

Describe a misbehavior to the Glassray bot and it can log a deviation from the conversation and scan for more instances.
You can review, confirm, and inspect deviations in the dashboard, or programmatically over the MCP server with list_deviations, get_deviation, and update_deviation_status.