Spaces:
Paused
Paused
Commit History
-
Mozg odpowiada tylko na podpisane wywolania; bez placu zabaw; awaria to blad, nie pozegnanie
- respond(context_json, sig): HMAC-SHA256 z ARIA_BRAIN_SECRET, fail-closed, limity rozmiaru - _try_it usuniete (publiczne wywolania modelu na kredyt organizacji) - jezyk po granicach slow, honorowany wybor z Workera; role tylko user/assistant - blad inferencji rzuca gr.Error zamiast udawac wyjscie ARII - gradio i huggingface_hub przypiete; README bez hf_oauth8f9bde4
-
aria: fix instruction-leak (frame+lang moved into system msg, convo ends on real user turn); react ~every 3rd msg; 'byt'->'inteligencja'; 400-char limit
Malina +1: Claude Opus 4.8- orchestrator: per-turn frame + language directive now live in the SYSTEM message instead of trailing user turns -> ARIA stopped answering the instruction itself ('no wiadomo że po polsku'); reaction-only turns get a tiny synthetic user turn - persona: reactions encouraged ~every 3rd emotional message (as an add-on, not instead of the reply); SYSTEM_PROMPT 'byt' -> 'inteligencja' - message length cap 100 -> 400 (server events.js, input maxlength, client) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>2e61db0 -
Update brain/orchestrator.py
Verified signature18d0fde -
brain: switch to aya-expanse-32b (your 32B/gemma-4 weren't available; gemma-4 broke Aria entirely)
Malina +1: Claude Opus 4.8- model: CohereLabs/aya-expanse-32b (works, strong Polish; kept bill_to=NextNeural) - restore reactions in build_user_frame (SYSTEM_PROMPT untouched), de-emphasized so the model still replies with words; both directions (Aria reacts + comments user reactions) - reply context + silent gender (never verbalizes the assumption) - robust output parsing: accept content/text/message keys; split multi-sentence replies into separate lowercase no-period bubbles (chat style, model-agnostic) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>5ea523e -
zmiana na google/gemma-4-31b-it
Verified signature11e098d -
fix zmiany modelu
Verified signature09c43be -
Model change to Qwen/Qwen2.5-14B-Instruct
Verified signatureb6af9cd -
Tescik 2
Verified signature113d2e6 -
Tescik
Verified signaturefec9e45 -
brain: force reply language to match user's last message (was defaulting to Polish on English)
Malina +1: Claude Opus 4.8Add _lang(ctx) (last user message) and append an explicit per-turn directive ('Reply in ENGLISH' / 'Odpowiedz po polsku'); reuse it for signoffs. Verified: EN in -> EN reply, PL in -> PL reply. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>961d1cd -
aria: fix role mixup in rehydrated history; reliable language-matched idle nudges; anti-echo
Malina +1: Claude Opus 4.8- DO: rehydrate context by MERGING user events + ARIA bubbles in time order with correct roles (was loading only user events -> model couldn't tell who said what) - idle nudges generated locally from PL/EN pools, language = user's LAST message, no-repeat; always sends (model sometimes returned nothing) - drop bubbles that echo the user or repeat a prior line - persona: explicit role note + 'never copy the user'; signoff lang = last user msg Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>1308d69 -
aria: randomized two-bubble farewell on inference failure (12 leave + 11 bye per lang)
Malina +1: Claude Opus 4.8Out-of-credit/error now sends a casual two-bubble sign-off (leave line -> bye line), randomized and language-matched, e.g. 'muszę spadać' + 'siemanko' / 'gotta go' + 'catch you later'. DO mirrors it for Space-unreachable. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>98ca451 -
brain: bill inference to NextNeural org via bill_to (was charging personal token owner)
Malina +1: Claude Opus 4.8InferenceClient(bill_to='NextNeural') routes Inference Providers usage to the org instead of the personal account that owns HF_TOKEN. Also unblocks replies since the org has billing/credit. Configurable via ARIA_BILL_TO; requirements bumped for bill_to. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>988cdc2 -
aria: casual sign-off on inference failure instead of robotic error
Malina +1: Claude Opus 4.8On brain failure (esp. 402 no-credit) ARIA now bails like a person — 'muszę lecieć' / 'gotta go for now' (language-matched) — instead of 'ej coś mi się zcięło'. Same for Space-unreachable in the DO (casual bubble, not a red error). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>466ef03 -
aria: 27B model, messenger style (lowercase, no period), 25s idle nudges, cluster tuning
Malina +1: Claude Opus 4.8- brain: switch to google/gemma-3-27b-it (vision, multilingual, ~3x cheaper than 72B) - persona+schema: human messenger style — lowercase start, no trailing period; idle produces a short varied ping (halo?/jesteś?/you there?), never repeats last msg - DO: idle 25s; inject system_idle into history so brain nudges (not duplicates); dedup idle bubble vs last reply; debounce 900ms so quick msgs cluster Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>8a87a60
-
fix(brain): use Qwen2.5-VL-72B-Instruct (7B-VL not offered by any provider)
Malina +1: Claude Opus 4.8The default model returned 'model_not_supported', producing the 'ej, coś mi się zcięło' fallback for every message. Switch to a vision-capable model that's actually enabled; log the real exception instead of swallowing it. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>8e49975
-
ARIA: event-driven multimodal agent (brain + Cloudflare backend + Messenger UI)
Malina +1: Claude Opus 4.8- brain/: persona, event/bubble contracts, vision orchestrator; app.py exposes api_name=respond - cloudflare/aria-worker: AriaConversation Durable Object, R2 media, usage (20 msgs/5h) - cloudflare/pages-integration: Messenger UI + _worker.js merge snippet - cloudflare/db: additive D1 migration - ARIA_ARCHITECTURE.md Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>4131875