pi-byteplus-modelark-provider

A Pi extension that registers your BytePlus ModelArk Coding Plan subscription as a model provider in Pi.

One API key covers the whole plan — Seed, GLM, Kimi, DeepSeek and GPT-OSS — and the model list is discovered from BytePlus rather than hardcoded, so new plan models appear without waiting for an extension release.

If you have used pi-byteplus-modelark, this is the same idea with two changes you might care about:

• No export BYTEPLUS_API_KEY=… in your shell profile. You sign in with Pi's own /login, and the key is stored in ~/.pi/agent/auth.json.
• No hand-maintained model list. The extension reads the model list BytePlus publishes for the Coding Plan, layers in context/output limits, and falls back gracefully when BytePlus changes that page.

Requirements

• Pi 0.85.x or newer.
• Node 20 or newer (what Pi itself requires).
• A BytePlus ModelArk Coding Plan subscription. The Coding Plan endpoint rejects a pay-as-you-go-only key.

Quickstart

1. Install

From npm:

pi install npm:pi-byteplus-modelark-provider

Or straight from git, pinned to a tag:

pi install git:github.com/jstokke/[email protected]

Or over HTTPS:

pi install https://github.com/jstokke/[email protected]

Already have the old pi-byteplus-modelark installed? Remove it first — both register a provider, and the old one pins a stale model list:

pi remove npm:pi-byteplus-modelark

2. Sign in

Get an API key from the BytePlus ARK console (region ap-southeast-1), then inside Pi:

/login byteplus

Pi prompts with a masked secret input, checks the key, and saves it. You only do this once.

3. Pick a model

/model

You should see the current Coding Plan line-up under BytePlus ModelArk, for example byteplus/glm-5.2, byteplus/kimi-k2.5, byteplus/deepseek-v4-pro, or byteplus/ark-code-latest for the plan's auto-router.

What it does

| | | | :--- | :--- | | Provider id | byteplus | | Display name | BytePlus ModelArk | | Endpoint | https://ark.ap-southeast.bytepluses.com/api/coding/v3 (Coding Plan) | | Wire | OpenAI Chat Completions | | Auth | Pi /login (masked prompt → auth.json), env vars as fallback |

• Uses the Coding Plan endpoint, so requests consume your subscription instead of being billed on top of it.
• Discovers the plan's models from the list BytePlus publishes, cached 24h, with the live /models catalog as a metadata source and fallback, and ten known-good ids bundled as a last resort.
• Knows each model's limits — context window, output cap, vision and reasoning support — instead of shipping every model as "128k, text only".
• Sends reasoning correctly. ModelArk accepts reasoning_effort (none | minimal | low | medium | high | xhigh | max), so Pi's thinking picker actually works, and glm-5.2 gets the extra off/xhigh levels BytePlus documents for it.
• No runtime dependencies. Plain Node built-ins; the HTTP surface is fully mocked in tests.

Managing it

pi update --extensions   # update, whichever source you installed from
pi remove npm:pi-byteplus-modelark-provider
pi remove git:github.com/jstokke/pi-byteplus-modelark-provider

Alternatives and escape hatches

Headless / CI setups can skip /login and use an environment variable:

export BYTEPLUS_API_KEY="ark-…"   # or ARK_API_KEY, the name BytePlus uses

Raising the output cap if Pi reports Response was truncated before completion.:

export BYTEPLUS_DEFAULT_MAX_TOKENS=65536

Per-model caps go in ~/.pi/agent/byteplus-model-overrides.json:

{
  "deepseek-v4-pro": { "maxTokens": 200000 },
  "Seed 2.0 Lite":   { "contextWindow": 262144 }
}

If BytePlus restructures its docs page, point the discovery at a mirror rather than waiting for a release:

export BYTEPLUS_PLAN_DOC_URL="https://mirror.example.com/coding-plan"
export BYTEPLUS_NO_ENRICHMENT=1   # or skip the scrape entirely

Every environment variable, the discovery layers, the curated limits and their provenance, and the troubleshooting messages are documented in src/README.md.

Notes and limitations

• Built and tested against Pi 0.85.x. @earendil-works/pi-coding-agent and @earendil-works/pi-ai are optional peer dependencies: Pi provides both at runtime, and the extension installs no runtime dependencies of its own.
• I only have a Coding Plan subscription, so that is the only tier this has been exercised against. A pay-as-you-go-only key will be rejected by the Coding Plan endpoint — that is BytePlus' behaviour, not the extension's.
• BytePlus does not document a model-listing endpoint for the Coding Plan. The live /models call is therefore treated as best-effort: it is used for metadata and as a fallback, never as the authority on what the plan serves.
• Only the OpenAI-compatible wire is registered. The plan also exposes an Anthropic-protocol endpoint, but it serves the same models, so registering it would just duplicate every entry in the picker.
• The model list is parsed out of a public BytePlus documentation page, and the Coding Plan endpoint is undocumented. Either can change without notice; npm run smoke is what reports drift.

Development

git clone https://github.com/jstokke/pi-byteplus-modelark-provider.git
cd pi-byteplus-modelark-provider
npm install
npm test          # 113 unit tests, fully mocked, ~0.6s
npm run typecheck
npm run smoke     # parses the live BytePlus docs page

To iterate without reinstalling, symlink the source into Pi's extension directory:

mkdir -p ~/.pi/agent/extensions
ln -s "$(pwd)/src" ~/.pi/agent/extensions/byteplus

See CONTRIBUTING.md for the project layout and the conventions this repo cares about.

Disclaimer

Not affiliated with, endorsed by, or sponsored by BytePlus, ByteDance or Zhipu AI. "BytePlus", "ModelArk", "Seed", "GLM", "Kimi", "DeepSeek" and "GPT-OSS" are trademarks of their respective owners, used here only to describe what this extension interoperates with.

The extension targets an undocumented Coding Plan endpoint and parses a public BytePlus documentation page to discover models. Check that this fits BytePlus' terms for your own use.

License

MIT © Joachim Stokke