Main now offers osm_query, and names it in the prompt, only when Overpass
is configured. Typed chat follows the same switch on both providers: its
prompt, the tools it offers and the calls it accepts.
The budget test measures the full instructions plus every tool. At this
head that is 55,210, ~290 under the 55,500 ceiling; the note said 55,370,
which that measurement does not produce.
/api/osm/features keyed its rate limit and its one-search-per-client slot
by the full peer address, so an IPv6 host could rotate addresses inside
its /64 to take both global slots and slip the 12/min limit. It now uses
clientNetworkKey, as /api/overpass does.
Picks up acd87e6c: the keyless cross-site gate on /api/osm/features
(582bfce6) and main at 93804d5d. SECURITY.md, the format-scope list and
the schema-test comments take the branch's side; the place-search budget
wording from the timeout fix is kept.
While the camera tracks an entity or looks at a point, Cesium measures
zoom distance from the target, not the Earth. Raising the ceiling to the
camera's altitude then lifted it with every outward step, so a tracked
camera could zoom out without limit. The guard now follows the altitude
only for an untracked 3D camera and otherwise keeps its 40,000 km
ceiling.
The action under way when a reply stops is answered as cancelled, since it
may have taken effect, and only later calls as skipped; a stop during a
round's first action keeps that round too. The closing note no longer
claims the actions succeeded. Tests cover the refusal body read cap and
the release of an unread refusal body.
Picks up #1527, which restores the 97 format-scope entries an earlier merge
dropped; this branch had kept the same 97. The resolved list is main's plus
this PR's 12 test and helper files, 1,222 paths, all present.
Each resolve() that raced the caller's signal added an abort listener and
never removed it, so a long-lived signal collected one listener per
completed lookup. The listener is now removed when the call settles and
is not added at all to a signal that is already aborted.
Without a limit parameter the route read null, Number(null) is 0, and the
clamp turned that into a one-feature list. An absent or empty limit now
gets the documented default of 500.
resolve_area compared the spoken name with the unit's name literally, so
"State of Texas" (admin1) and "USA" or "UK" (country) were refused with
AREA_IDENTITY_MISMATCH although the bundled matcher found exactly those
units. Both identity checks, the bundled one and the outline ladder's, now
also accept the prefix forms and same-unit country aliases the matcher
reads, at the level they imply, so the level check still applies.
"Great Britain", "Britain", "America" and "Korea" still need their own
names, and the natural-region check stays literal.
The guard's own tests pass even if createApplicationControls stops
installing it. This checks the ceiling and frame listeners are in place
after construction and gone after teardown.
Above 300 km the guard treated any shrinking pinch distance as an outward
zoom, so a two-finger tilt whose fingers drifted slightly closer had its
pitch overwritten. A two-finger gesture now counts as a zoom only when the
spacing changes more than the fingers' midpoint moves.
The guard reads its input in preRender, which Cesium only fires on frames
that draw. In request-render mode, wheel input at the ceiling doesn't move
the camera, so nothing drew and the input was applied later by an
unrelated frame. The guard now requests a frame through the render
governor whenever it has unread input.
Cesium clamps any zoom that starts above maximumZoomDistance to that
limit, in either direction, so a share link or flight that left the
camera above 40,000 km jumped to the ceiling on the first wheel tick.
While the camera is above the ceiling, the guard now uses its current
height as the limit and lowers it as the camera descends.
osm_query said "At least N" when the server capped its list, but handed
the places to the analyst engine with truncated:false, so "which of those
have emergency?" answered as if the list were whole. The handoff now
carries the cap and its caveat, and follow-ups (and their follow-ups)
stay incomplete with an unknown total.
The voice answer waited 25 s for /api/osm/features, while the server let
Overpass run for 25 s and itself waited up to 40 s per endpoint and 45 s
in all. When the browser gave up first it closed the request, the route
cancelled the upstream query, and nothing was cached, so the OSM_TIMEOUT
retry started the same query from scratch and timed out again.
The server now tells Overpass 20 s, waits 22 s per endpoint and 22 s in
all, which ends inside the client's 25 s. A test pins that ordering.
A reply stopped partway through a round now keeps the actions that ran,
answering the calls it never reached as skipped, so the history stays
paired for both providers. A failed typed turn no longer resets the
pointer, referents and follow-up memory, since its conversation continues.
The server now tells a rejected key (401, or Gemini's 400 with reason
API_KEY_INVALID, matched exactly), denied access (403) and an account out
of credit (OpenAI's 429 insufficient_quota) apart; each cools down for a
minute and gets its own message in the chat.
Merge commit 34906c9a resolved the end-of-list conflict in scripts/format-scope.json with the branch's side, which dropped the 97 entries main had appended since Oct 26. Those files all still exist; 64 of them are tests that runtime discovery does not cover, so format:check had stopped checking them.
Each entry goes back at its position on main before the merge. All 1680 files pass npm run format:check.
#1308 moved /api/overpass to the keyless cross-site gate: foreign origins
and cross-site fetches are still refused, but forwarding headers are not,
so the map works behind a reverse proxy. Voice place search spends the same
operator Overpass, so it now uses the same gate; with the stricter one,
place search failed behind a proxy while the map kept working. The prompt
lint note records the merged size, 55,370 with an operator Overpass.
Picks up #1308 (Overpass admission), #114 (Sea Marks and Ski Pistes),
#1206 (radio country coverage) and #1435. Resolutions:
- SECURITY.md: main's rewritten cross-site paragraph, with voice place
search listed beside /api/overpass among the keyless map routes.
- scripts/format-scope.json: both sides. A merge on main dropped 97 entries
(being restored there); keeping them, this PR's 12 and main's 2 new ones
keeps every file in format:check. All 1,222 paths exist.
- Translation catalogs: main's two new keys and this PR's 47, no overlap.
- Layer count 36 (33 + OSM Places + Sea Marks + Ski Pistes); both schema
pins re-derived, the first-run length is exactly 30,919 + 292 + 107.
Typed chat sessions no longer end on the voice inactivity timeout, and a
failed or stopped reply keeps the conversation: Stop cancels only the
reply in progress, and map actions that already ran stay in the model's
context. Each upstream request and map action gets its own deadline, old
tool results shrink instead of the history overflowing, and step limits,
a missing or rejected key, the OAuth lock and throttling each get their
own message.
- Server: OpenAI chat is off under GEV_PREFER_CODEX_OAUTH; chat throttles
default to the provider's throttle; replies carry API_KEY_MISSING,
API_KEY_REJECTED, CODEX_OAUTH_REQUIRED and RATE_LIMITED codes; upstream
Retry-After passes through, and a rejected key (Gemini's invalid-key 400
included) cools down for a minute
- Pinokio forwards the chat throttle settings
- UI: the reply reveal publishes only the message list; the border orbit
runs only while busy; layers attach with a button or Enter; the log is
aria-busy while a reply types out; auto-scroll respects a reader who
scrolled up; Escape closes the attach menu; the chat gear says Settings;
a note says chat bills the OpenAI key when voice uses ChatGPT sign-in
- Layout does less work while panel height animates; unused blurs removed
- Gemini voice start ends chat; the voice preview clears on every new
session
- Tests for the route gates, client limits, attachments and the fixes
The two Update-branch merges resolved the locale conflicts by keeping
this branch's side, which dropped nine strings main had added (Utilities
panel, drawing colours, cockpit exit). Rebuild each catalog as main's copy
plus Sea Marks and Ski Pistes, and re-pin the voice tool schema length and
digests on top of main's Utilities target.
Tightening the AREAS rule dropped "to outline it (draw=true)", and on Gemini
"Outline the state of Texas" stopped passing draw: none of 5 runs set it, so
the model drew the state again with annotate_map. With the clause back all 4
runs called resolve_area with draw. The prompt is 55,263 characters with an
operator Overpass, ~240 under the 55,500 ceiling.
From review:
- Card lines the actions write in English (area basis, part counts, merged
duplicates, OSM caveats, the imagery window) are shown in the UI
language through the analyst card's note path; the model still gets the
English. One OSM place reads "1 hospital".
- Same-named clarification choices show their region code (WA, DC).
- OSM Places shows as unavailable once the server says no Overpass is
configured, and its points are depth-tested past 2.5 km so they never
show through the globe.
- find_imagery puts the operator's sources and mode back when a box is
refused, and a voice search cancelled mid-flight leaves "Search
cancelled" instead of an empty strip.
- The README lists OSM Places with the other Utilities.
From review:
- The browser client for /api/osm/features gets its own tests (request
shape, abort, network failure, coded refusals), and a 200 whose body is
not JSON is now a failure instead of zero places.
- An end-to-end test runs resolve_area, osm_query and an analyst follow-up
through the real runner and checks the follow-up keeps the area; it fails
with the wiring reverted.
- The Realtime token test pins Overpass off and compares the full tool
schemas again instead of names taken from the environment.
From review: a row inside a large area pays a ray cast over every edge, so
20,000 rows against a 6,000-vertex outline made ~200 ms long tasks. The scan
now also yields once ~8 ms have passed, checked every 256 rows. An ambiguous
region refusal also tells the model the way back: resolve_area with the
picked candidateId, then analyst_query over that area.
manjunath22466