Fostera v0.5.0 — Video generation, on-device models & weather

New Features

• Souls can look up the weather and draw it as a card, including active National Weather Service alerts for the place you asked about — so a watch standing over you is named instead of buried under "Overcast".
• Souls can pin a place on a map, and locations can be shared end to end in chat.
• Souls can search your mail and calendar and read a thread: arbitrary Gmail searches, message bodies behind your permission, calendar events in the past as well as the future, and free-time lookups.
• Added mature content for verified adults, on by default. Dark, intense fiction comes enabled on the Soul types built for it, romance-coded Souls (companion and anime companion) allow explicit content as well, and any Soul can be turned either way in its own settings. None of it reaches an account that is not a verified adult.
• Added video generation end to end, and a generated clip is now a real file you can search, remember, and play.
• Added on-device models that run locally in your browser, including a compatibility mode for devices without WebGPU and a smaller Lite tier.
• Added a next-turn cost estimate to the Resonance meter and the model picker, so you can see what your next reply will cost before you send it.
• Added an Effort picker for Premium, plus effort-aware routing that matches the model to the work being asked of it.
• Added Claude Opus 5, Qwen 3.8, and Gemini 3.1 Pro, 3.5 Flash, and 3.1 Flash-Lite.
• Added native Threads reply chains and topic posting, and LinkedIn posts can now choose their visibility (Anyone or Connections).
• Rebuilt the memories library around a knowledge spine and observatory insights, with pinning, journey synthesis, temporal recall, and still-true checks.
• Souls can edit the documents attached to their projects.
• Replaced the separate Queue tab with one Schedule surface: compose directly, edit publish times, and split posts anchor at the day and time you choose.
• Redesigned Home as one searchable workspace, with starring on every list, working hub shortcuts, and Soul orbs on project rows.
• Added per-Soul text speed, engine-first creation starters, and an adaptive temperature centered on each Soul's type.
• Added 40M Resonance and $1 entry packs.

Improvements

• Voice mode speaks the reply while it streams, so it starts talking right away instead of waiting for the whole answer to finish.
• Agentic turns reuse their cached prompt prefix instead of re-billing every tool result.
• Context compression now drops what is replaceable rather than simply what is oldest, so longer chats keep more of what matters.
• Automatic routing scores what a turn will actually cost and risk, and one stalled model can no longer eat an entire Auto turn.
• Rebuilt the appearance and Look & Feel settings, and gave each route a loading skeleton that matches its own page.
• Generated Word documents keep their formatting, with real ordered lists and proper page furniture.
• Memory applies your corrections every turn, relearns a fact you restate, and stops discarding facts you asked it to keep.
• Clearer billing throughout: annual checkout says it is billed yearly, payment failures are handled and surfaced, and the wallet and allowance meters read honestly everywhere.
• Refreshed the public site for both the productivity and companion audiences, and rewrote the demo chat to show what a Soul actually does.

Bug Fixes

• Fixed attaching an image making a reply fail. The photo was being measured against the model's context window as though it were text, which could rule out every model and leave you with "couldn't reach a model that could answer".
• A reply that stops because you are out of Resonance now says so and points you at topping up, instead of reporting that every model refused.
• A conversation that has grown too long to answer now tells you to start a new chat, rather than looking like a provider outage.
• Fixed proactive reach-outs failing whenever a Soul's most recent conversation was a project.
• Fixed voice mode cutting in and out, and stopped repeated cues from stacking on top of each other.
• Fixed the light theme hiding its own text in several places, and referral stats being invisible there.
• Fixed a closed modal leaving its whole subtree on the page.
• Fixed duplicate Souls being created when onboarding raced with itself.
• Fixed LinkedIn verification failing without saying why, and Threads topic tags failing to post.
• Fixed the PDF renderer in production, reasoning text leaking into replies, and empty replies being reported as real ones.
• Fixed scheduled posts and document turns being reported as an empty response.
• Corrected creation spend so it is attributed to the reply that caused it, and stopped generated files being named "-.png".