@d3ara1n/pi-provider-stepfun

StepFun (阶跃星辰) provider for pi — registers pay-as-you-go and Step Plan channels through StepFun's Chat Completions API.

Providers

Both channels accept the same API key — set both env vars to it if you only have one. The difference is billing and model set:

• stepfun bills per token (CNY) and includes step-1o-turbo-vision.
• stepfun-plan consumes Step Plan subscription allowance and includes step-router-v1.

StepFun also supports Anthropic-compatible Messages, but this extension uses its documented Chat Completions API for both channels. The base URLs above are not interchangeable: calling the ordinary API does not spend Step Plan allowance.

Models

step-3.5-flash-2603 is the Agent/Coding-optimized snapshot of step-3.5-flash — faster and more token-efficient, tuned for coding and agent frameworks.

Step 5 Preview also accepts video through the API, but pi's model input declaration only supports text and images. StepFun documents up to 1M output tokens for this model; this extension caps individual responses at 64K so pi can reserve room for input within the 1M context. Its Chat Completions compatibility is based on the published model documentation, not yet verified with a live API request.

step-router-v1

A routing model: it automatically dispatches each request to either deepseek-v4-pro (complex reasoning, 1M context) or step-3.7-flash (routine calls, 256K context). pi advertises the router's larger context, but the selected engine's actual limit applies; very long requests may fail if routed to Flash. Charges depend on the engine selected for each request.

Note that step-router-v1 emits an [Advisor consultation] … [End of advisor consultation] planning block inside its text output — this is the router's internal planning, not something to act on.

Pay-as-you-go Pricing

The ordinary API bills in CNY per million tokens, using any gift balance before the topped-up account balance. Current official rates:

Cost estimation in pi's status bar is disabled (set to 0), since it assumes USD. These prices are references, not amounts charged by the extension.

Step Plan Subscription

Step Plan's current Token Plan measures model use in Credits (1M Credits = ¥1 of model usage). Credits are issued monthly, can be spent at any time during the month, and do not roll over. The plan channel does not draw on the ordinary API account balance.

Quarterly and annual plans are paid up front, but Credits are still issued monthly. Subscribers can buy a separate 30-day booster when their monthly allowance runs out: ¥49 for 400M Credits or ¥99 for 1,600M Credits. Model usage is converted to Credits from its actual API charges; step-router-v1 uses the price of the engine selected for each request. These are the published domestic tiers; check the subscription page for current prices in your region.

The older Coding Plan used request counts and five-hour/weekly limits. It cannot be purchased again, but existing subscribers with successful auto-renewal can keep it; upgrading to Token Plan is irreversible. Neither plan requires a different endpoint or provider ID. See the official upgrade notice.

Installation

pi install npm:@d3ara1n/pi-provider-stepfun

Or add to ~/.pi/agent/settings.json:

{
  "extensions": [
    "/absolute/path/to/pi-extensions/packages/pi-provider-stepfun"
  ]
}

Configuration

Get an API key from the StepFun console, then set the env var for the channel(s) you use:

export STEP_API_KEY="your-api-key"        # stepfun (pay-as-you-go)
export STEP_PLAN_API_KEY="your-api-key"   # stepfun-plan (Step Plan)

Or store it through /login, or manually in ~/.pi/agent/auth.json:

{
  "stepfun": { "apiKey": "your-api-key" },
  "stepfun-plan": { "apiKey": "your-api-key" }
}

Notes

• The Step 3.x Flash models always reason. Turning thinking off in pi only omits the reasoning parameter — the model still thinks at its default level. This is a model characteristic, not something the provider can disable. step-1o-turbo-vision does not reason.
• step-3.5-flash-2603 accepts only low and high reasoning effort; pi's medium is mapped to low. For models with three documented effort levels, pi's extra levels are mapped into the low/medium/high range.

Dependencies

None — standalone provider with no pi-extension dependencies. It uses pi's built-in openai-completions streaming.