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.
The voice debug log writes the server's timestamp after the client's record,
so a record can no longer replace loggedAt. The dev and preview servers also
send X-Content-Type-Options: nosniff with their shared browser headers.
From the findings in #710.
Co-authored-by: Sunil <[email protected]>
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
- 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follow-ups and never lets a superseded query replace follow-up memory.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.
Ported onto the split provider layout: the route now lives in
server/providers/openai/debug-log.js and its state belongs to the handler
the factory returns rather than to module scope.
/api/realtime/debug-log is the one route in its family left unhardened.
Both siblings enforce a limiter through openai/rate-limit.js, and this was
the only handler still writing with fs.appendFileSync on the request path.
- Always-on per-IP limiter (120/min, 400 global). Deliberately not the
opt-in openAiRateLimiter(): that one defaults to unlimited so throttling
cannot silently change what a user spends, and this sink spends disk and
event-loop time rather than provider quota.
- Serialized async appends with a 32 MB rotation ceiling and one retained
generation. REALTIME_DEBUG_LOG_MAX_BYTES caps a single request body; it
never bounded the file those bodies accumulate into, so any local page
could grow .gev-logs/realtime-conversations.jsonl for as long as the dev
server ran. Serializing also closes a read-then-write race on the size
check, where concurrent callers all stat under the ceiling and none
rotates.
- Generic 400/500 answers separating a malformed record or oversized body
from a failed write, instead of echoing error.message — an errno and an
absolute path for an fs failure.
- readRequestBody marks an oversized body with code 'BODY_TOO_LARGE', the
marker its sibling readRequestBodyCapped already sets, so the two cases
are distinguishable without matching on text.
Coverage extends src/tooling/localServices.test.mjs, which already drives
the real handler over a temporary sourceRoot. Both new tests fail on the
previous implementation and pass with this one.
Refs #23. Does not add authentication: these proxies have no auth
mechanism today, so introducing one is a maintainer design decision.
Sameh Khamis