The provider owns authentication, dynamic /models discovery, the OpenAI-compatible transport, and model-specific request compatibility. It does not proxy requests through another service.

Why this package exists

The original pi-provider-alibaba@1.0.1 correctly implemented the route and transport layer, but marked every dynamically discovered model as reasoning: false, text-only, 128K context and 8K output. That made native Pi and downstream hosts hide reasoning controls even for models that support them.

This maintained package keeps the original provider architecture and adds the missing capability projection:

Alibaba-compatible system role instead of OpenAI developer;
top-level enable_thinking and reasoning_effort;
model-specific reasoning levels;
current context, output, and image capabilities for known model families;
filtering of image/audio/video-only entries from the chat catalog;
contract and request-body tests.

Install in Pi

Do not install this package together with pi-provider-alibaba: both register the same provider IDs.

pi install npm:pi-provider-alibaba-complete

Set the credential belonging to the product you purchased, restart Pi, then select the matching provider and model:

# Token Plan
export ALIBABA_TOKEN_PLAN_API_KEY='<your-token-plan-key>'

# Coding Plan
export ALIBABA_CODING_API_KEY='<your-coding-plan-key>'

# Pay-as-you-go DashScope API
export ALIBABA_API_KEY='<your-api-key>'

China and international routes use different endpoints. A key that can list models on the wrong route can still fail the real completion request, so always select the route matching the key's product and region.

Install in DeepSeek Harness

dsh plugin --profile web add pi2dsh
dsh plugin --profile web add pi-provider-alibaba-complete

export ALIBABA_TOKEN_PLAN_API_KEY='<your-token-plan-key>'
NODE_USE_ENV_PROXY=1 dsh web

The NODE_USE_ENV_PROXY=1 prefix is needed only when your network requires the Node process to use HTTP_PROXY / HTTPS_PROXY.

The models appear as native DSH routes, including groups such as Alibaba Token Plan (CN). The DSH model picker reads the provider's Pi metadata and exposes only meaningful reasoning levels.

Reasoning controls

Model family	Selectable levels	Alibaba wire format
DeepSeek V4 Pro / Flash	Off, High, Max	`enable_thinking` + `reasoning_effort`
dated DeepSeek V4 variants	Off, Low, High, Max	`enable_thinking` + `reasoning_effort`
Qwen 3.8 Max	Off, Low, Medium, XHigh	`enable_thinking` + `reasoning_effort`
Qwen 3.6 / 3.7 hybrid models	Off, High	`enable_thinking`
always-thinking models	their real non-disableable level	no fake Off control

Unknown models stay conservative until their protocol is known. The provider does not guess a reasoning format which a newly released model might reject.

Verification

npm install
npm run verify

The release was additionally verified against a real China Token Plan:

native Pi with DeepSeek V4 reasoning enabled;
raw DeepSeek V4 high and off requests;
raw Qwen 3.8 Max medium request;
DeepSeek Harness model picker showing Off / High / Max;
two complete DSH agent loops across a process restart, including tool call, tool result, second model request, and final answer;
all main DSH requests carried enable_thinking: true, reasoning_effort: "high", and a system message;
the credential appeared in zero persisted test files.