Skip to content
fold-kit-experiments

Service Incident Lab

An app-owned AG-UI reference demo at /agent-analysis, with a second retained host at /agent-analysis/retained. Its backend is a deterministic synthetic fixture, not a live agent, production monitoring system or remediation service.

Useful Outcome

Run analysis to inspect a service latency change against a validated baseline. The chart uses fixed fixture domains, a clearly illustrative 100ms p95 budget, and baseline outlines for changed services. Table selection supplies a keyboard path to the service inspector; local filtering does not restart the stream.

The normal fixture raises API latency from 60ms to 145ms, preserves it through a reordered snapshot, then adds a fourth service. Counts come from valid records. Newly added services have no earlier measurement and are labelled New rather than given a manufactured comparison.

The invalid fixture first accepts the same spike, then attempts an atomic patch that changes latency to 999ms and removes a missing service. Rejection evidence is separate from validated measurements: no partial change reaches chart, table or inspector. The last valid data, baseline and selection remain available. Restart analysis begins a new normal run requiring its own first valid snapshot; it is not an in-run resync or resume acknowledgement.

Ownership

  • Astro owns the document, client island and navigation facts. The host retains the same island across the two demo routes; it does not store agent state.
  • The FoldKit Model owns run identity, first-snapshot baseline, last valid Wire, local inspection and bounded rejection/evidence state. Commands/subscriptions own side effects; only accepted data is supplied to chart geometry.
  • Viz supplies pure Cartesian geometry and parent-owned rendering helpers. It has no AG-UI dependency and cannot load data or execute agent work.
  • The endpoint validates SDK input and app-owned scenario data, limits actual request bytes to 16 KiB and streams SDK-encoded SSE from a request-owned scope.

Each new run retains the previous valid Wire and its matching baseline while connecting. Its first valid snapshot replaces both. Later snapshots update Wire without resetting that run’s baseline; comparison uses stable service IDs, not row order. Failed, cancelled and completed runs retain useful measurements.

Cancel stops the active client stream without starting a replacement. Route exit cancels an active run, retained stayed navigation preserves it, and unmount disposes its scope. Closing HTTP is not proof that a real agent stopped computing.

The supported inbound subset is run start/finish/error plus state snapshot/delta. Other valid event families are ignored. The demo does not claim full AG-UI client compliance, automatic retry, a backend run registry or provider integration.

Local Checks

Terminal window
bun run --filter @opsydyn/web dev
bun run --filter @opsydyn/web test:e2e
bun run check
bun typecheck
bun run test

The two browser files cover real retained navigation, responsive geometry, keyboard selection, coherent baseline comparisons and invalid-update recovery at desktop and 390px viewports. They exercise the real local Astro SSE endpoint, not the visual prototype’s fixed state switching.

See the implementation plan and qualification record for current execution evidence. Browser CI, compiled Cloudflare endpoint streaming, screen-reader acceptance, deployed behaviour and real backend compute cancellation remain separate qualification gates. The locally qualified and reviewed demo is committed as 083dff8 and merged into local main. The subsequent user approval authorises pushing this work and the previously reviewed commits; hosted workflow results are separate from the local qualification evidence.