[mmabrouk]

Context

In the old app, the scenario drawer of an evaluation run shows, for the app output and for each evaluator, a button that opens that step's trace, plus the time, tokens and cost of that step. Its header also has full screen, collapse all, copy as JSON and a Form or JSON switch. The /m drawer had none of this. It had only one "Open trace" button, for the app output. So in /m you could not open an evaluator's trace or see what one step cost.

Stacks on #8196 (scenarios table), which also has #8183 (run header) and #8164. The trace drawer message view from #8154 is already in that branch.

Changes

Each section of the drawer (Inputs, Outputs, and each evaluator) now has a title that folds the section. Outputs and each evaluator also get a trace button and badges for time, tokens and cost:

Outputs                          [trace] 370ms  # 39   $0.000005
Exact Match                      [trace] 201μs
Answer correctness (LLM judge)   [trace] 887ms  # 126  $0.000009

The trace button opens that step's trace in the /m trace drawer, with the readable message view. The badge text comes from formatMetricValue, so it reads the same as the scenario table.

The header now has four tools: full screen (wide screens only, because a phone already shows the drawer full screen), collapse or expand all sections, copy the scenario as JSON, and a Form or JSON switch. The JSON looks like this:

{"inputs": {"country": "Japan"}, "outputs": {"outputs": "Tokyo"},
 "evaluators": {"Exact Match": {"success": true}}}

If an evaluator failed, its entry is {"error": "<message>"}. When you move to another scenario, the drawer keeps your view choice. Previous, Next, the scenario picker and copy id work as before.

There are no new requests. The step metrics come from the scenario's metric rows (data[stepKey]), and the page already reads those. The trace ids come from the step results, and the page already reads those traces in one request per page. A new shared helper, stepMetricStats(metrics, stepKey, metricKey) in @agenta/entities/evaluationRun, reads one step's metric. scenarioMetricStats now uses it.

The Annotate tab uses the same body component. It does not pass the new view prop, so its layout and save path stay as they are. All new code is inside the drawer's lazy chunk.

What the tests cover

Risks, most costly first:

1. A step shows another step's numbers or trace, for example the app's cost under an evaluator. Tests: "Each step shows its own time, tokens and cost", "A step's trace button opens that step's trace in the trace drawer", and "A step without a trace has no trace button".
2. The drawer sends a request for each step or each scenario. Test: "The traces of a page are read in one request, and moving on reads none".
3. Copy or JSON gives wrong data, for example a failed evaluator looks like a pass. Tests: "JSON shows the whole scenario as data, and Copy as JSON copies that text" (the test checks the exact text) and "An evaluator that failed shows its error in the JSON view".
4. The tools stop working: sections do not fold, or the view resets when you move to another scenario. Tests: "Collapse all folds every section, and a section's title opens it again" and "The chosen view stays when moving to the next scenario".
5. Unit test for stepMetricStats: it reads only the given step, and a missing step or null data returns null.

The tests use the real atoms and hooks. They mock only HTTP, the trace fetch, the clipboard, time and the router. Mutation check: I made four changes to the code, one at a time, and each change made one of the tests above fail:

• read the metrics from the wrong step
• open the wrong trace id
• ignore the folded state
• drop the evaluator error

One case has no test: two runs of the same evaluator in one scenario. The code shows the first run that has a trace.

Tests / notes

• Types pass for @agenta/mobile, @agenta/entities and @agenta/oss. eslint and prettier pass on the changed files. The tests of the changed files pass, plus the related run, scenario and Annotate tab tests (73 tests).
• speed:check passes. Run page 1055.6 KB (limit 1068.9 KB). /evaluations 1071.5 KB (limit 1093.9 KB). App shell 886.8 KB (limit 908.7 KB).
• I checked the change live against a seeded run. The numbers match the old app for the same test case (370ms, 39, $0.000005 for the app; 887ms, 126, $0.000009 for the LLM judge). I did not check the old app's per-field type tags and per-field Text/JSON switch. They are not ported; the Form/JSON switch for the whole drawer replaces them.

What to QA

• Open an automatic run, go to Scenarios and open a row. Outputs and each evaluator show a trace button and their own time, tokens and cost.
• Click an evaluator's trace button. The trace drawer opens on that evaluator's trace, not on the app's trace.
• Click "Collapse all sections", then click one section title. Only that section opens. Then switch to JSON, click Next, and check that the drawer stays in JSON.
• Click "Copy as JSON" and paste. The text matches the JSON view.
• On a phone width, the badges wrap under the section title and the page does not scroll sideways.
• Regression: the Annotate tab of a human run still shows the plain inputs and output next to the annotation form, and saving an annotation still works.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review: 1

• release/.*

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

Run configuration

• Source mode: Auto
• Repository: Agenta-AI/agenta
• Branch: feat/m-parity-evaluation-focus-drawer
• File: .coderabbit.yaml
• Commit: f97b3fd
• Layers: Repository YAML (base), Organization UI (inherited)
• Review profile: CHILL
• Plan: Advanced
• Run ID: b88517a7-27bb-42ee-a59f-df00ea2313f0

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn

[mmabrouk]

🤖 The AI agent says: @coderabbitai review

[coderabbitai]

Rate Limit Exceeded

@mmabrouk have exceeded the limit for the number of chat messages per hour. Please wait 8 minutes and 11 seconds before sending another message.

[github-actions]

Website preview

Preview URL: https://pr-8207-agenta-website-preview.mahmoud-637.workers.dev

Built from f97b3fdbc3eeefef88cf6c6ccf061e4adff8885e. This comment updates in place on every push.

[github-actions]

Railway Preview Environment

Item	Value
Status	stopped
Commit	f97b3fdbc3eeefef88cf6c6ccf061e4adff8885e
Manual expiry (UTC)	Not granted
CI deadline (UTC)	Not granted
Workflow	https://github.com/Agenta-AI/agenta/actions/runs/42281103992

Dependent tests finished. Post /preview for manual review. Environment absence verified; retained storage requires separate provider accounting.

Post /preview as a new comment to request a one-hour preview. Repeating it while active does not extend the hour.
Preview data is disposable. Expiry is enforced by the next successful cleanup sweep, not an exact timer.
Environment deletion stops its compute. Provider-retained storage and permanent services may still be billed.