Studio Radar

Radar answers repo-health questions without opening CI logs: is studio performance drifting on main, what did a run look like, which commit broke it, what shipped when. It is a Sanity Studio (dev/radar, deployed at https://radar.sanity.dev) reading a dataset that CI writes. The design record is dev/radar/SPEC.md — read the section for the view you touch, and update it when behavior changes. For regression hunting use sanity-radar-investigate; for producing the data use sanity-bench.

Tools (URL path = tool name)

Trends URL state, all shareable: ?range=30|90|all, ?branches=main,<branch>, ?layers=-band,-calibration (hide layers), ?tab=<metric group>, ?max=<series key> (one chart maximized). Releases: ?path=<test-studio path> (the path field under the header) opens every release's Test Studio link at that path. Clicking a point opens the run popover: value, percentiles, host, release context, links, and under "Suspect a regression?" the GitHub compare of the gap, Copy A/B vs previous run (the gh workflow run bench.yml … ab_from/ab_to command) and Copy investigation prompt (a paste-ready brief for an agent).

The data

Project mhfozd0z, dataset bench, private: reads need a token with access to the project (your CLI token via sanity debug --secrets, or a viewer token from sanity.io/manage; the deployed studio authenticates through your login). Writes need RADAR_SANITY_WRITE_TOKEN, which only CI has; never write from a laptop. Ids are built in one place, @repo/utils/radar-ids: lowercase, dashes, no dots (a dot makes an id a path like drafts.x, which the API scopes differently).

Rules that every consumer follows (the dashboard's queries are the reference, tools/*/data.ts):

• Project, never fetch sessions. Documents carry per-session sample arrays; a bare *[_type == "benchRun"] is megabytes.
• Time series = mode == "absolute", ordered by coalesce(git.committedAt, startedAt) (backfill runs have a historical committedAt and a recent startedAt). PR-branch runs share the type — filter git.branch == "main" for the main line.
• Same-sha runs are one point. CI re-measures commits; the charts merge them by median.
• Absolute numbers are host-relative. Always read runner.calibrationMs (higher = slower host) and runner.cpuModel next to any value; only A/B comparison.verdicts are host-independent.
• git.sha is the join key to gitCommit; benchRun.git.commit is a weak reference that dangles for PR-branch runs.

Ready-made queries: references/groq-recipes.md.

Working on the dashboard

pnpm radar                                  # sanity dev on http://localhost:3399 (needs a Sanity login; reads live data)
pnpm vitest run --project=radar             # pure modules: trend math, drift, bisect, git parsers
pnpm build:radar                            # production build (turbo)
pnpm --filter radar sync-git -- --dry-run   # preview the git-history sync (needs GITHUB_TOKEN)

• Realtime everywhere: useDocumentStore().listenQuery + useObservable, never one-shot client.fetch. New cron runs must appear without a reload.
• Charts are visx primitives (@visx/scale|shape|group|axis|responsive) inside @sanity/ui layout; no chart library.
• Debug data sources (dev server only, toolbar picker; tools/trends/debugData.ts) render steady/drift/step/host-correlated/sparse/empty datasets and synthetic release tags — use them to exercise every encoding layer and the drift feed without live data. UI is verified through these, not unit tests.
• One source of truth for "enough to care": drift thresholds import from perf/bench/stats/gate.ts. Don't invent a second statistic; SPEC.md records the baselines that were tried and rejected (step, weekday-matched) and why.
• Honesty constraint: any surface showing a ms value must make host calibration visible (CALIBRATION_EXPLAINER is the one shared wording).
• Studio conventions from AGENTS.md apply (no forwardRef, use-effect-event, no inline Translate components).

Operations

• Daily main run: bench.yml cron 05:00 UTC; git sync 05:30. A missed day is repaired with a backfill dispatch (see sanity-bench). A gap in commits: gh workflow run sync-git-metrics.yml -f backfill=true.
• Release markers only appear for tags whose commit has a gitCommit document (main-only by construction); a maintenance release cut from a branch is correctly absent.
• Cron failures alert the CI Slack channel; a hole in the series with no alert means the store step was skipped, not that the run failed.