Repository navigation
Description
The umbrella (#11038) acceptance criteria 3 and 6 have no implementation: nothing yet takes a live Queen actor event and updates one widget, and nothing measures event-to-visible-paint latency. T27F's data layer exists and is sealed (specs/t27f/ui_plan.t27 catalog + validate_plan(); specs/t27f/spec_graph.t27; the routing law in specs/t27f/frontend.t27); what is missing is the smallest end-to-end bridge, measured. This issue is that bridge.
This is the minimal vertical slice proposed in #11038. It reuses existing specs and adds no new system: one existing actor event source, one catalog widget id, one renderer target, one measurement hook. T27F Live (epic #10266, issues #10260-#10265) is the parent track; this issue does not implement work_graph, scene, live_stream or face in full -- it proves the wiring once and puts a number on it.
Prior art read on 2026-10-22 (why this shape): Vercel's json-render (vercel-labs/json-render) proves catalog-bounded generative UI with many render targets; OpenUI (thesysdev/openui) proves a streaming-first UI language with first-party runtimes; Phoenix LiveView proves server-state diffing to a browser. None of them is spec-sealed or provenance-labelled, which is T27F's ground. The measurement hook borrows the shape of Goose's ObservabilityHook (correlated requestId, typed lifecycle events, duration and usage) because #11038 identified it as the missing instrumentation pattern.
Boundary
- specs/t27f/live_slice.t27
- specs/t27f/frontend.t27 :: route
User Scenarios
Given a running Queen actor runtime whose actors emit events already counted by specs/queen/telemetry.t27, when one such event reaches the slice bridge, then exactly one widget from the ui_plan catalog is re-drawn and the new frame carries the event id it was derived from.
Given a 10-minute soak of actor events, when the run ends, then the summary line reports p50/p95/p99 event-to-visible-paint latency in milliseconds, the redraw count, and the dropped/duplicated event counts, all recomputable from the recorded spans alone.
Given an event whose widget cannot be drawn this turn (renderer target down), when the bridge picks a stand-in, then the frame is labelled with the stand-in reason and nothing undrawn is claimed made, per claim_ok in the routing law.
Requirements
- FR-001: The slice MUST draw exactly one widget id that already exists in the catalog of
specs/t27f/ui_plan.t27; this issue MUST NOT add a widget id to that catalog. - FR-002: Every rendered update MUST carry a span record with a correlated request id, the actor event id, and start / first-paint / end timestamps, so that latency is recomputable from the log alone.
- FR-003: The bridge MUST live as a
.t27spec withtestandbenchblocks, and its runtime output MUST be produced byt27ccode generation (L2); hand-written foreign glue MUST stay within the 40-lines-per-file cap. - FR-004: The slice MUST NOT edit
specs/queen/telemetry.t27; new metric names it needs MUST be declared in the new spec, in the naming style telemetry.t27 already uses. - FR-005: The renderer target MUST be one that already exists today; the slice MUST NOT introduce a second renderer, a new UI framework or a new event bus.
- FR-006: A failed draw MUST be visible: the span record MUST distinguish drawn, stand-in and refused outcomes.
Success Criteria
tri test specs/t27f/live_slice.t27exits 0 on the lab, including the spec's owntestblocks.- A 10-minute soak reports p50, p95 and p99 event-to-visible-paint latency in ms, with
0dropped event ids and0duplicated event ids. - The measured numbers and the method are written to
live_slice_bench.md, and that file lists the histogram bucket edges used. - The routed surface appears in the
routetable ofspecs/t27f/frontend.t27, and the claim gate still passes:tri test specs/t27f/frontend.t27exits 0. - No new dependency appears:
git diff --stat origin/masterfor the slice branch touches only the two Boundary paths.
Activity
gHashTag commentedon Nov 17, 2026
Spec exists and passes typecheck + tests. M10 milestone complete.