Remove 'one moment' phrase from prompt — tool is now instant

Stage 1 speed: drop internal GPT-5.5 fuzzy match (local matcher), thread tool, tighter timeouts

Live product image updates as assistant speaks (latest-mention wins)

24h cache; warmup stock XML; fallback to catalog when stock fails (with model-number disambiguation)

Drop transcript-based main detection; trust GPT/tool for product selection

Allow brief wait phrase on FIRST stock query only; subsequent calls instant

Tighten voice prompt: 1-2 sentences max, no filler, single-sentence stock answers

Replace thumbnails with overlay arrow nav; click image to pause auto-rotation

Revert "Fix Marlin switching bug: persist tool_link across sub-responses, treat digits as distinctive, drop + glue"

Revert "Fallback: when stock XML fails, still show product image from Trek catalog index"

Fallback: when stock XML fails, still show product image from Trek catalog index

Fix Marlin switching bug: persist tool_link across sub-responses, treat digits as distinctive, drop + glue

Strict transcript-based main product detection (all 4+ chars tokens must match)

Equal-size monitors; landscape thumbnails; rails fit in side gap

Move gallery thumbnails outside monitor: vertical rails on left+right

Drop color suffix from product name; thumbnails already show color visually

Drop transcript-based main product detection (false positives); show main only via tool

Strip size from displayed name; use main link when showing color variant

Always show main product; switch to color variant only if user mentions a color

Remove 'one moment, checking' prompt — tool is now instant

Parse XML once into hash index; build in background thread (no event loop blocking)

Revert "Default to main product; switch to variant ONLY when user explicitly asks color/size"

Default to main product; switch to variant ONLY when user explicitly asks color/size

Use new XML schema: gallery (8 images), color/size variant matching

Switch Trek XML feed to /output/2688003925 (richer fields, multi-image)

Replace bg image with pure-CSS dual monitors (Apple Studio Display style)

Use real dual-monitor image as backdrop, inset avatar/product into screens

Two side-by-side monitors: assistant left, product showcase right

Light theme inside CSS monitor frame: bezel + stand, white screen, dark text

Show main product by default, switch to variant on color/size mention, prompt aware of visuals

Stop assistant from reading product URLs aloud (prompt + strip from tool result)

Detect product mentions from assistant transcript when tool isn't recalled

Fix product image: match by exact productLink from tool result instead of name tokens

Add split-screen product display: avatar left, product image right

Revert "Log AudioContext sample rate for mobile diagnosis"

Revert "Show audio sample rate on screen for mobile debugging"

Show audio sample rate on screen for mobile debugging

Log AudioContext sample rate for mobile diagnosis

Fix Turkish orthography (use proper diacritics) and lang=en for English text

Use original Trek SVG logo in brand header

Trek-branded modern UI: red accent, Rajdhani type, futuristic frame

Update assistant image (with headset mic)

Show full image without cropping (3:2 ratio, contain)

Replace 3D Three.js avatar with static image (LFS)

Use API default voice (same as generic)

Migrate session.update to GA realtime API format

Remove beta header for GA realtime API

Upgrade to gpt-realtime-2 model

Add /bh/products /bh/warehouses /bh/inventory cached JSON proxies for tavsiye system

/warehouse-xml endpoint: Trek PHP ve diger programlar icin merkezi XML kaynagi (CORS open, 5dk downstream cache)