Workspace health
Studio › Overview is your workspace’s dashboard: what it holds, how its runs go, what needs attention, and how good each agent, skill, rule, guardrail, eval, workflow and worker is. Studio opens on it.
Spec: workspace health.
The facts come from what the workspace already records, so they’re always current and cost nothing:
- Workspace score: a ring for the workspace and one for each kind, with how many items each has and how many don’t load. Click a kind’s ring to jump to its card.
- Runs, last 30 days: runs per day (done, failed, stopped) with their cost as a line, the success rate and the total cost. Open the card for what costs the most and what fails the most.
- Eval coverage: which agents have evals, how their latest results went, and which changed since their evals last ran.
- Needs attention: files that don’t load, failing workflows and workers, low grades, agents with no evals or out-of-date evals, the auto reviewer’s denials, rules and guardrails that fired, and items nothing uses, most urgent first.
- One card per kind: open it for a table of its items, with each one’s grade, flags, uses or runs per day, its trend and its cost or eval results. Click a row for its verdict.
Every card opens in place with the arrow in its corner, and the Overview remembers which cards you opened.
A judge model rates each item against a rubric for its kind: for an agent, whether its instructions are clear, whether its tools fit, whether its evals cover it and whether its runs succeed; for a workflow, whether it handles failure, what it costs and how often it succeeds; and so on. It’s given the item’s file and what the records say about it (its eval results, runs, failures and costs), so a grade rests on evidence, not opinion. Each rubric line gets a score from 1 to 5 with a reason, and up to three suggested fixes.
An item’s grade is its overall score: A from 4.5, B from 3.5, C from 2.5, else D. A kind’s and the workspace’s grades average their items, weighted by how much each is used; a file that doesn’t load counts as a D. Grades also show as badges beside each item in Studio’s lists.
Improve on a fix puts a request in the chat for the agent to make that change: send it, and you review the edit as you would any other (its diff, approvals, undo).
Re-check judges the items that changed since they were last judged (or have never been); Re-check all judges everything again; each kind’s card has its own Re-check, and an item’s verdict has Judge again. Each first says how many items it will judge and what that should cost, and shows its progress as it goes; Stop stops it after the item being judged. Verdicts are kept for 90 days, for the trends.
Once you’ve checked by hand, the workspace is checked again every day at 02:00 by default, judging only what changed. Change it with the gear on the Overview: when (in words or as cron, such as weekdays at 6 AM, or off), the most a scheduled check may cost (estimated first: a check that would cost more doesn’t run, and the Overview says why), and the judge’s model. These are kept in the workspace’s own settings, never in the project’s files:
[health]
model = "gemini/gemini-4-flash" # empty: the auto mode's reviewer, else the default model
schedule = "every day at 02:00" # or cron, or "off"
max_cost_usd = 1 # 0: no limit
Scheduled checks run while the service runs and the workspace is open in it; a night missed runs at the next chance.