[psanders]

While a run is in progress, the benchmarks page shows it live. Live is on for every run.

What changes

• live.json (perfkit/live/v1). Zod schema plus schemas/live-v1.schema.json. The agent writes <run root>/live.json every 15 s (temp file, then rename): run id, version, commit, status (running, finished, failed with reason), project, estimated end, and every benchmark in perfkit.yaml order with its state (queued, running, done, not-run).
  • Capacity: the current step (from the clock), every step the agent already judged (offered, placed, calls, failed, p50/p95/p99, failed %, ok, stop reason), heldCps, and once done result (limit, lower bound, stop reason).
  • Soak: elapsed, warm-up and duration, calls placed and failed so far, raw RSS per process (at most 300 points), and growthMBPerHour, which stays null until there is a full hour of data after the warm-up.
  • It reuses the analyzer: evaluateCapacityStep now returns the step's numbers, and the soak's least-squares growth is moved into memoryGrowthMBPerHour, which the verdict also calls.
• Serving it. GET /live.json from a node:http server on 127.0.0.1:<live.port> (default 8787, --live-port):
  • responses carry Cache-Control: no-store;
  • CORS allows the origin of project.siteUrl and http://localhost:*, plus the ngrok-skip-browser-warning header, and the server answers the OPTIONS preflight;
  • it answers 503 before the first snapshot and 404 for anything else.
  • With NGROK_AUTHTOKEN, an @ngrok/ngrok tunnel opens on live.domain or NGROK_DOMAIN. The server and tunnel close when the run ends.
  • On DigitalOcean the agent serves it and stays up 30 s after the run so open pages see it end. With --local, the run/collect process serves it from this machine.
• Where the page finds it. You can set project.live.url or live.url in perfkit.yaml, or let it default to https://<live.domain>/live.json. The URL goes into report.json project.live.url, an additive optional field. ?live=<url> on the page overrides it.
• Live view in the app. Built from the approved "State / Live (preliminary)" frame: live picker, preliminary banner with a ticking "updated N s ago" (warning after 60 s, "connection lost, retrying" after 3 min or a failed poll), header, progress cards and overall bar, and live numbers.
  • Capacity uses the report's own latency chart, drawn up to the current step, with a "now" marker, the steps not run yet shaded, and a "So far" card with the last published limit.
  • Soak uses the report's memory chart with the warm-up band, a "now" line and the MB/h note.
  • When the run finishes, the page says the report is being published and draws it once data/runs.json lists the run. Run pages never go live.
  • Respects prefers-reduced-motion. The final report is unchanged: live CSS and chart options are added only when live.
• perfkit serve --site <dir> [--port 8080]. A static server for local previews. The README has the local recipe.
• Workflow. run.yaml takes an optional NGROK_AUTHTOKEN secret and passes it only to the start job, which puts it into the agent container's environment.

Verified

• npm test (626 passing), npm run typecheck, npm run lint.
• A real perfkit run --local: a 3-step ramp, then a 3-minute soak with a 1-minute warm-up, watched with perfkit serve and headless Chromium. The page went live, updated, showed "Run finished; the report is being published", then drew the final report once it was rendered into the site.

Notes for review

• The @ngrok/ngrok native binding for the agent's Alpine container is installed on the load generator at start (npm install of the pinned 1.7.0 in node:22-alpine), only when a token is set. It never fails the run.
• Docker keeps a container's environment in its config on the droplet, so the token lives there until the droplet is deleted. It is never in run.json, a log or the report.

🤖 Generated with Claude Code