Keeps the voice-core prompt rules and adds main's world-query and
vessels-vs-visuals rules; the tool catalog wraps the deixis-aware runner.
Also fixes two things the branch had before the merge:
- Live voice could not start: OpenAI rejects track_entity's top-level
anyOf. Model-facing tools now omit top-level combinators; the runner
and benchmark still require a query or a referent.
- The Contacts-centred analyst answer dropped the viewport coverage note.
- Voice layer manifest (src/voice/layerManifest.js) generates the layer enums,
aliases and analyst field hints; a test fails when a catalog layer has no entry.
- analyst_query validates fields, operators, units, values and scopes, counts up
to 250,000 loaded records per layer, reports complete/partial/unanswered, tells
LAYER_OFF / NOT_READY / FEED_UNAVAILABLE apart from zero, refuses mismatched
follow-ups and never lets a superseded query replace follow-up memory. Seven
more layers expose analyst records.
- Tool results carry a code-built say line and a display card; the voice card in
the mic control shows captions, plan steps and the result
(src/voice/resultDisplay.js is the one adapter). Slow tools narrate their
current step out of band; holding Space while the assistant talks cuts it off.
Turn spans go to the debug log; caption transcription is metered; the log
omits what was said unless GEV_VOICE_LOG_CONTENT=1.
- Point and ask: the Space-down pointer snapshot resolves "this", "that one" and
"around here" through 'pointer' targets, and "the second one" through
referent:n from the last numbered list.
- Prompt states each policy once (instructions + tools 48.5k -> 46.7k chars).
- Voice benchmark harness (scripts/qa-voice-bench.mjs, scripts/voice-bench/)
shares its phrase table and grading with scripts/qa-voice-routing.mjs.
NHC stamps an advisory with its nominal time and writes it about half an
hour early, so the 5-minute future limit rejected the whole feed each
advisory cycle and the layer showed no storms. Allow 90 minutes ahead.
Toggling datacenters, dams, or submarine cables off only set `show = false`
on their data sources. Cesium's DataSourceDisplay still walks every
visualizer of a hidden source each frame, and the ~11.5k hidden entities
(hundreds of MB) stayed resident, so the whole scene rendered ~2x slower
for the rest of the session after a single on/off round trip (owner field
test 2026-09-13).
disable() now removes the data sources from the viewer and drops the built
records; the parsed dataset (NDJSON features / TeleGeography JSON) is cached
so enable() rebuilds without a refetch or re-parse. A disable that lands
while a load is in flight releases the finished build instead of parking it
hidden. destroy() clears the cache.
Measured with a headful Chrome harness (flights+military+satellites on,
camera rotating, 250-frame windows), before -> after:
- entities resident after OFF: 11,537 -> 0
- DataSourceDisplay.update per frame after OFF: 1.09 ms -> 0.02 ms
- scene.render per frame after OFF: 13.9 ms (control 6.6 ms) -> 5.2 ms
(fresh baseline 5.0 ms in the same run)
- JS heap 30 s after OFF: 1.6 GB and climbing -> 342 MB after GC
Tests: the GeoJSON lifecycle and cables suites assert disable leaves zero
sources/contexts and re-enable rebuilds with fetch called once.
Two independent triggers turned the global status chip red while every layer
was still populated (reproduced at JFK in Cockpit; timeline in the issue).
adsb.lol proxy: a non-OK upstream status was relayed verbatim even with a
12 s-old cache in hand, and the browser layer answers a 429 with a 45 s
cooldown in which every refresh reports "adsb.lol rate limited" without a
network call. The proxy now serves its cached body as STALE on a 429/5xx and
backs off from upstream for Retry-After (bounded), mirroring the OpenSky
proxy. A failure with nothing cached is still relayed.
Mapped installations: the "zoom in" guidance was stored in `error`, so the
manager reported a refresh failure and the chip a LOAD FAILED whenever the
view could not be bounded (Cockpit looks at the horizon). The status is now
set without an error string; the row reads the prompt from `statusMessage`,
and the chip ignores `error` for guidance statuses, matching the carve-out
layerFeedState already has.
Bug reports now require reproduction steps, environment, the doctor
report, and logs. Feature requests require a data source and license.
Blank issues are disabled; security reports are routed to private
vulnerability reporting via the contact links.
Bilawal Sidhu