Valet 0.3.0

Added

• assistant.access takes a key per tool: plex.plex_stop_stream: control under plex: read lets that one tool act, and downloads.torrent_control: none or valet.report_issue: none takes one away from the chat, commands, rules, plans, pins and live cards alike. config check names a key that matches no tool, with the tools that instance has. Without dotted keys nothing changes.
• assistant.reasoning: auto (as before), off, low or high. off turns a Qwen-style model's thinking off in the chat (chat_template_kwargs: { enable_thinking: false }, which servers that don't know it ignore); low and high send reasoning_effort; Anthropic maps them to its thinking budget. Background calls keep the model's default.
• assistant.options passes temperature, topP and maxOutputTokens to the model. Left out, the server's defaults.
• assistant.tools: all (the default, as before), lazy or auto. lazy sends Valet's own tools and one line per integration, and loads an integration's tools when a question needs them; auto is lazy once all the tools come to about 6k tokens. A much smaller prompt on a big setup, for one more step on questions about an integration. When it loads some, the chat shows a "Picked tools" step.
• The docs say how to keep the assistant's context small: which settings helped in testing against a local model, and which made answers worse.

Changed

• A Preply lesson counts as an online meeting: its classroom link gets the ring before it starts, the reminder and a Join button, like Meet or Zoom.
• The wall display lays out a page like the main page: a section's columns per screen ({ phone, tablet, desktop, wide }) and a card's own cols and rows, with cards in a row ending flush. The demo's House page uses it: the cameras over two rows, the thermostat beside who's home and the doors, then the energy and the weather.

Fixed

• The assistant's last step, the one kept for answering, could loop on a small model: a tag or "Let me check…" hundreds of times, until the server's limit. It now doesn't think first and stops after about 400 tokens, and stray tags a model leaves in its answer (</think>, </thinking>, <trans>) aren't shown.