Skip to content

v1.0.1 — the assistant on your own OpenAI key, reminders and recurring events

Latest

Choose a tag to compare

@tengosro tengosro released this 09 Nov 00:09

What's new:
• "Remind me tomorrow at 8 to call the bank" becomes a calendar event with a reminder, even when
it belongs to no project.
• A recurring event ("every Monday at 9 a meeting") is written into the calendar day by day;
when you don't say until when, the assistant usually asks.
• On your own server with an OpenAI key the assistant works with reasoning models too (e.g.
gpt-5.6-luna) without extra settings: a refused reasoning setting is retried with "none", then
without it, and what worked is remembered. Before, every turn ended with error 400.
• Drawing through OpenAI defaults to gpt-image-1-mini (many new keys no longer have gpt-image-1);
set KB_IMAGE_MODEL for another one. The error message no longer shows doubled braces.
• New guide in the docs: Your own OpenAI key — the assistant and drawing with one key, which
model to pick, what we measured and what to do when something fails.

Verified: 10 runs before × after with gpt-5.6-luna — reminder 3/10 → 10/10, recurring meeting
2/10 → 10/10, other tasks unchanged; no regression with another assistant model; targeted suites
ai-chat-openai-strict, ai-obrazky, proces-asistent, ai-chat, chat-rezimy, help index, i18n, en-first;
full regression (290 suites) green.

Upgrade note: no migration. If you added KB_CHAT_OPENAI_EXTRA with reasoning_effort none for
OpenAI, you can keep it (it takes precedence) or delete it. As always git pull and rebuild the image.

Česky

Novinky:
• „Připomeň mi zítra v 8 zavolat do banky“ asistent zapíše jako událost v kalendáři
s připomínkou, i když nepatří k žádnému projektu.
• Opakovanou událost („každé pondělí v 9 porada“) asistent rozepíše do kalendáře na jednotlivé
dny; když neřeknete, do kdy, obvykle se zeptá.
• Na vlastním serveru s klíčem OpenAI asistent funguje i s uvažujícími modely (třeba
gpt-5.6-luna) bez nastavování navíc: odmítnuté nastavení přemýšlení zopakuje s „none“, pak bez
něj, a co prošlo, si pamatuje. Dřív každý tah skončil chybou 400.
• Kreslení přes OpenAI má výchozí model gpt-image-1-mini (gpt-image-1 řada nových klíčů už nemá);
jiný nastavíte v KB_IMAGE_MODEL. Chybová hláška už neukazuje zdvojené závorky.
• Nový návod v dokumentaci: Vlastní klíč OpenAI — asistent i kreslení s jedním klíčem, jaký model
zvolit, co jsme naměřili a co dělat, když něco nejde.

Ověřeno: 10 běhů před × po s gpt-5.6-luna — připomínka 3/10 → 10/10, opakovaná porada
2/10 → 10/10, ostatní úkoly beze změny; s jiným modelem asistenta beze zhoršení; cílené sady
ai-chat-openai-strict, ai-obrazky, proces-asistent, ai-chat, chat-rezimy, nápověda, i18n, en-first;
plná regrese (290 sad) zelená.

Poznámka k upgradu: bez migrace. Kdo kvůli OpenAI přidal KB_CHAT_OPENAI_EXTRA s
reasoning_effort none, může ho nechat (má přednost), nebo smazat. Jako vždy git pull a přestavba image.