Evaluation fixtures, tool-call assertions, and assistant edits you approve

New

Fixture editor on evaluation suites — author simulated tool responses with a searchable tool picker, an argument matcher whose schema keys insert with a click, and a response prefilled from the tool's real example payload. See fixtures.
Required tool calls on test cases — one editor asserts a tool was called, optionally with matching arguments, call counts, a required call order, and a total tool-call budget. See suites and cases.
Test setup on test cases — freeze the simulated time (entered as wall-clock time in a chosen timezone), the channel, and which suite fixtures the case runs against, so time-dependent cases stay deterministic across reruns. See suites and cases.
Delete suites and cases — permanent, takes the run history with it, and refused while a run is in progress. See deleting suites and cases.
Knowledge faithfulness check on test cases — tick it on and an AI check verifies the answer only asserts what the retrieved knowledge supports. Separate from knowledge grounding, which checks the right sources were retrieved: an agent can retrieve perfectly and still invent details. Scored n/a when nothing was retrieved, and adds 1 interaction per case execution, shown in the estimate. See scoring.
Author fixtures with the assistant — the in-dashboard assistant can now look up a tool's argument schema and example payload, then create and edit a suite's fixtures in chat. Runs still start from the dashboard only. See fixtures.

Run details now badge each tool call as simulated or blocked, separate test-configuration errors from infrastructure errors, and group scores into execution, factual and quality tiers. See scoring.
Assistant change sets keep their approved, applied or discarded status when you return to a conversation, and can now update an agent's starter questions. See editing with the assistant.
Widget sandbox gains Save to Agent — persist the previewed title and heading to the agent's draft config instead of re-typing them in the editor. See web widget.
Delete confirmations — deleting an evaluation suite, case or fixture now shows a confirmation dialog spelling out what goes with it, instead of a browser prompt. See deleting suites and cases.
Fixture storage — evaluation fixture matchers and responses are now stored encrypted at rest.
In-dashboard assistant — reliably finds answers from the product documentation again. Nothing to do on your side.
Run details — the raw evaluator detail now shows tool and knowledge source names next to their ids.
Dashboard chat shows typing dots for the whole streamed reply, so a pause for a tool run or a thinking model no longer looks like a frozen cursor.

Knowledge grounding scored 0% even when retrieval worked — the evaluation trace never captured which knowledge sources the agent used. Every retrieval path now reports them, so the grounding score reflects what actually happened. Rerun any suite whose grounding scores looked wrong.
Questions phrased as a rewrite of a topic — "expand on the refund policy" as an opening message — were treated as formatting instructions and answered without consulting knowledge. They now retrieve like any other knowledge question.