Models and Jev
Because the assistant is built on pi, every provider and model supported by pi is available. This page manages four cards: the main model and its thinking level, the Jev API key, LLM provider logins, and web search providers.
主模型 (Main model)
The model list is loaded from pi's agent directory (combining built-in models, the pi.dev catalog, and any custom models.json). The catalog is fetched at bot startup and refreshed automatically once a day (unless PI_OFFLINE=1); clicking 刷新目录 in the page header refreshes it on demand.
Clicking the current model chip opens a searchable modal over the full catalog. The main model must support vision input, as image understanding and sticker selection rely on it; models without vision remain visible in the list with a 不支持视觉 badge and cannot be selected. Next to the chip, the 思考 dropdown sets the model's reasoning effort (关闭, minimal, low, medium, high, xhigh, max) and takes effect immediately.
Jev(出口审查)
Jev is TypeSafe AI's decision model. It never writes replies; instead, it answers structured yes/no probability questions (0–1) and multi-level scoring questions (0–2)—covering group join-in, reply timing, quote decisions, turn-queue ordering, per-bubble outbound safety checks, unknown sandbox domains, and candidate note scoring at rollover. Because every outbound message passes through Jev review, if Jev has no key or cannot be reached, outbound review fails closed and no messages are sent.
This card shows Jev's status (正常 / 未设置 key / 调用失败) and a 设置 key / 更换 key button. The API key is write-only: it is stored on the bot side and never returned to the browser or written to logs.
The mode, threshold, and custom criterion for each configurable Jev decision live on the page for that feature: 回复时机 and 引用 on Conversations and policies, 需要程度 (queue ordering) on Advanced settings, and 出口审查 on Safety.
服务商 (LLM login)
Lists every model provider registered in pi along with its supported login flows (OAuth or API key). Logged-in providers appear first, with the first 15 shown inline and the rest searchable via the 更多 modal. Starting a login relays pi's interactive prompts (authorization URLs, device codes, or API key inputs) over a WebSocket and stores the resulting credentials exclusively on the bot side. Providers that only read ambient environment variables display pi's setup hint below the row.
搜索服务商 (Search providers)
Configures the providers used by the web-search and page-fetch tools (currently DeepSeek and Parallel). Keys are write-only here as well. Click the active provider chip at the top of the card to choose which logged-in provider to use; if only one search provider is logged in, the assistant uses it automatically even before an explicit selection is saved (the card notes this alongside the chip).
DeepSeek shares its API key with the LLM provider login above: signing in or logging out on either card updates both at the same time.