fix(eval overview): hide non-output metrics for evaluator steps

[mmabrouk]

Summary

• Restrict evaluator metrics in Overview to evaluator output namespaces (attributes.ag.data.outputs.* and normalized equivalents)
• Filter both live run metrics and fallback evaluator metric definitions using the same namespace check
• Prevent annotation infra metrics (duration, cost, tokens, errors) from showing as evaluator metrics in the Overview section

Testing

• Not run (frontend-only filtering change)

---

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview, Comment	Mar 11, 2026 9:09pm

[devin-ai-integration]

Devin Review found 1 potential issue.

View 4 additional findings in Devin Review.

[devin-ai-integration]

🟡 normalizeMetricPath produces paths for outputs. prefix that isEvaluatorOutputMetric will always reject

When normalizeMetricPath receives a path starting with outputs. (e.g., "outputs.score"), it produces "attributes.ag.outputs.score". The new isEvaluatorOutputMetric filter then checks this against EVALUATOR_OUTPUT_PATH_PREFIXES, but none of them match "attributes.ag.outputs." — note the missing data. segment.

Root Cause

At evaluatorMetrics.ts:161, normalizeMetricPath maps outputs.X → attributes.ag.outputs.X:

if (trimmed.startsWith("outputs.")) return `attributes.ag.${trimmed}`

But the EVALUATOR_OUTPUT_PATH_PREFIXES at evaluatorMetrics.ts:30-35 does not include "attributes.ag.outputs." — it only includes "attributes.ag.data.outputs." (with the data. segment). So isEvaluatorOutputMetric("attributes.ag.outputs.score") returns false, and the metric is silently dropped at line 179.

Compare with the data. prefix handling at line 160: normalizeMetricPath("data.outputs.score") → "attributes.ag.data.outputs.score" which correctly passes the filter.

This inconsistency means any evaluator definition whose metric path starts with outputs. (e.g., "outputs.score") will have that metric silently excluded from fallback metrics.

Impact: In practice, the standard extractMetrics flow (evaluators.ts:87-100) produces bare key names like "score" which hit the default branch of normalizeMetricPath and correctly get prefixed with attributes.ag.data.outputs.. So this bug would only manifest if an evaluator definition provides a metric path explicitly prefixed with outputs., which is a supported but apparently uncommon code path in normalizeMetricPath.

(Refers to line 161)

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[github-actions]

Railway Preview Environment

Status	Destroyed (PR closed)

Updated at 2026-03-12T14:36:12.125Z