Run errors not shown correctly in evaluation table

[junaway]

When an evaluation run fails because the OpenAI API key is empty, the evaluation table does not show the run error correctly.

To reproduce:

1. Set an empty OpenAI API key.
2. Create a prompt that uses an OpenAI model.
3. Run an evaluation on that prompt.

Expected behavior:

• The table shows the scenario failed to run because of the missing API key.
• Cost and latency show -.

Actual behavior:

• The output is empty.
• The cost and latency columns show errors about dicts.

[linear-code]

AGE-3828

[dosubot]

This looks related to a couple of recent fixes that have already been merged:

1. PR [fix] Render numeric/discrete evaluator metrics as values, not raw stats #4586 (merged June 9, 2026) specifically addressed "dict" errors appearing in evaluation table cells [1]. The issue was that raw stats objects like {"type":"numeric/discrete","count":1,"max":4,...} were being incorrectly coerced and rendered as JSON blobs in the table — which matches the cost/latency columns showing "dict" errors you're seeing.

2. PR [FE Fix] hide string-type evaluator metrics from scenario table & filters #4477 (merged May 28, 2026) fixed a similar issue with string-type evaluator outputs rendering as raw stats objects [2].

3. There's also a related issue (Evaluation Errors not clean #4377) about evaluation errors not being displayed cleanly [3], which shares the same root problem of error states not being surfaced properly in the table.

Regarding the empty output when the API key is missing — the cost and latency columns use a simple "-" fallback when values are undefined [4], but when the run fails with an error, the output cell should render an EvaluationErrorText component showing the failure [5]. The fact that it shows empty instead suggests the error result may not be properly propagated to the output column in this specific failure mode.

If you're on the latest version, the "dict" errors in cost/latency should be resolved by PR #4586. The empty output display for API key errors may still need a separate fix to ensure the error message from the failed invocation is correctly surfaced in the output cell.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.