pi-sensenova

Pi extension for SenseNova, an OpenAI-compatible provider. It registers the sensenova provider with a seed model catalog, then refreshes the catalog from /v1/models in the background and persists it for offline starts.

Install

pi install npm:pi-sensenova

Or install from git:

pi install git:github.com/pgciq/pi-sensenova

To try it for a single run without persisting:

pi -e .

Configuration

Set the API key before starting pi:

export SENSENOVA_API_KEY="your-api-key"

• Base URL: https://token.sensenova.cn/v1
• Provider id: sensenova
• Auth: SENSENOVA_API_KEY env var (the key is kept as an env reference, so pi marks the provider unconfigured rather than sending a placeholder when the variable is absent)

Or let Pi manage the credential interactively:

/login sensenova

The key is stored in ~/.pi/agent/auth.json; SENSENOVA_API_KEY remains a fallback.

• OpenAI-compatible streaming — text models stream through pi-ai's openai-completions API.
• Reasoning / thinking models — deepseek-v4-flash, deepseek-v4-pro, and glm-5.2 expose a thinking-effort level map (minimal, low, medium, high, xhigh, max).
• Image generation — models matching sensenova-u1, sensenova-u1.5, sensenova-u1-fast, sensenova-u1.5-fast (and sensenova-u1.5-lite) route to the verified /v1/images/generations endpoint. Generated images are saved under .pi/generated-images/; supported terminals also receive a TUI Image entry (rendered inline when the terminal supports a graphics protocol), while the saved path is reported as a clickable file:// link in the TUI. Print/RPC mode reports the saved path as plain text.
• Image editing — sensenova-u1.5-lite routes to /v1/images/edits when the prompt includes reference images.
• Dynamic model discovery — a seed list is available immediately; the full list is fetched from /v1/models and cached. Discovered models survive restarts and offline starts.

Seed models

| Model id | Notes | | --------------------- | ------------------------------ | | sensenova-6.7-flash-lite | Reasoning-capable text model | | deepseek-v4-flash | Reasoning-capable text model | | glm-5.2 | Reasoning-capable text model |

Models discovered via /v1/models are added automatically.

Model discovery (non-blocking)

pi-sensenova registers a seed model catalog synchronously at load (so pi starts instantly) and refreshes it from /v1/models in the background via pi's refreshModels callback — it never blocks startup on the network.

• The seed list is always available immediately, even offline or without SENSENOVA_API_KEY.
• A successful background refresh replaces the seed list and is persisted to pi's provider cache, so discovered models survive restarts and offline starts.
• Every network call is bounded by a timeout and degrades to the seed list on any failure.

Usage

pi --model sensenova/deepseek-v4-flash "你好"

Generate an image (text prompt) with an image model:

pi --model sensenova/sensenova-u1.5-lite "一只戴帽子的猫"

Commands

The extension registers the following commands:

| Command | Description | |---|---| | /sensenova-models [image\|vision\|audio\|video\|reasoning\|tools] | List SenseNova models with capabilities, context/output limits; an optional filter narrows the table. | | /sensenova-usage | Show token/cost usage accumulated in the current Pi process. |

Examples:

/sensenova-models
/sensenova-models vision
/sensenova-models image
/sensenova-usage

/sensenova-usage is based on message_end usage reported by completed assistant messages and is therefore local to the current Pi process. SenseNova's OpenAI-compatible API does not expose a uniform account-level billing/usage endpoint through this provider, so this command is not an account invoice.

Development

The extension depends on @earendil-works/pi-ai (a peer dependency provided by pi at runtime). createAssistantMessageEventStream is imported from the bare package, and openAICompletionsApi is resolved defensively: newer pi-ai builds expose it only via the @earendil-works/pi-ai/api/openai-completions.lazy subpath, while older builds export it from the bare package. The fallback import keeps the extension loading on both layouts. Validate with:

node --test