5 updates
Welcome to the sixth issue of the Moda weekly update. Last week we introduced more Home widgets, retention heatmaps and saved cohorts. This week we have been working on testing agent changes against production traces and making it easier to follow an investigation across the dashboard.
Test a model change against recorded calls
Replays lets you compare several models on your agent's recorded calls, so you can inspect how a change behaves on the work your agent actually receives.
Start a model comparison, choose the models to test and review their replies alongside the recorded context. Results show what each test found, with individual calls available in a side panel. You can inspect the examples behind a result before deciding whether to change your agent.
Fix tests and regression checks also live in Replays. A fix test compares a proposed change with a replay without that change. A regression check looks for problems the change could introduce elsewhere. Results distinguish extra friction from other outcomes, so a reply that makes the user do more work is visible in the comparison.
You can stop a running test from the dashboard. Replays is available to workspace members where the feature is enabled.
Follow up without losing the evidence
Ask now supports follow-up questions tied to a problem, user, use case or theme from an earlier answer.
When an answer points to a particular group of failures, you can ask about that finding and keep the relevant subject in scope. Linked evidence lets you open the source and check the answer against what happened.
You can also choose a time window for an Ask thread. The answer names the period it covers, and saved answers keep their original sources and time window. This makes it easier to return to an investigation later without confusing an earlier answer with today's data.
Answers stream into the panel as they are prepared, including structured result blocks. Progress is visible while Ask works, and you can rate an answer from the panel.
One time range across the dashboard
The dashboard now starts on the last three days and remembers the time range you choose as you move between pages.
Choose seven days on Home, then open Problems, Signals, Users, Topics or Traces to keep investigating that period. The shared choices are one, three, seven, fourteen, thirty and ninety days, though some pages offer fewer ranges. Users shows thirty days for a remembered fourteen-day range. Traces shows thirty days for a remembered fourteen- or ninety-day range. Your choice is remembered for your account and workspace in the current browser.
A link with an explicit time range takes precedence, so opening a shared link preserves the period its sender selected. Page-specific choices such as custom dates or All time stay on that page. Home widgets with their own saved range keep it.
If Home has no traces in the last three days, it offers a button to show the last seven days.
Choose where @moda answers in Slack
You can now choose which Slack channels allow answers from @moda, independently of where you send alerts.
In the Slack connection settings, link a channel to your Moda workspace and enable answers there. This gives your team a place to ask about production traces while keeping the choice separate from alert delivery.
If your Slack workspace connects to more than one Moda workspace, the channel's binding determines which workspace answers the question.
Inspect your harness from the CLI
The Moda CLI now has a local component map and a way to compare maps across code changes.
moda harness skeleton --path=. --out=/tmp/before.json --json
Run it again after your change with --out=/tmp/after.json, then compare the two:
moda harness skeleton-diff /tmp/before.json /tmp/after.json --json
The comparison lists added, removed, moved and changed components and relationships. Line-number shifts alone do not count as changes. The map includes detected agents, prompts, tools, models and their connections, with coverage and known gaps reported alongside the result.
The scan runs locally without an LLM or network calls. It records component locations and ownership, including excerpts of up to 80 characters from inline prompts. Check these excerpts before sharing a map.