pi-provider-qoder

A pi extension that talks to Qoder the way the official qodercli does — both regions, no separate CLI binary to install.

pi install npm:@zenodinh/pi-provider-qoder
# or: omp install npm:@zenodinh/pi-provider-qoder
# or straight from the repo: pi install git:github.com/zenodinh/pi-provider-qoder
# remove it later with: pi remove npm:@zenodinh/pi-provider-qoder

Quick start

Log in once. Paste a PAT from the global or China integrations page, or leave the prompt empty to sign in through the browser (global region only):

/login qoder

Then pick a model with /model — or skip the picker:

pi --list-models qoder                # the ids this account offers
pi --provider qoder --model <id>      # one turn, global
pi --provider qoder-cn --model <id>   # one turn, China

What you get

• Two regions in one extension — qoder (global) and qoder-cn (China), each with its own account, catalog and cache.
• Two ways to log in — paste a PAT, or approve a browser sign-in; a PAT in the environment logs the provider in at startup.
• The live model catalog — fetched from your account and cached per region, so the picker offers what Qoder actually serves.
• Thinking where the model has it — pi's levels map onto each model's enable_thinking and reasoning_effort support.
• Agentic tool use — native tool calls plus DSML markup in the text stream, with tool images passed back to the model.
• Robust streaming — Qoder's double [DONE] envelope, hidden <thinking> markup, queue waits, and one continuation when a stall cuts an answer short.
• Signed requests — every request carries the COSY signatures Qoder expects, so nothing sits in the middle.

Providers

Both register together; each is a separate region with its own account and Credits.

Commands

Models

• The catalog is live — /model/list per region, cached at ~/.pi/agent/qoder-models-cache.json and qoder-cn-models-cache.json, refetched when over an hour old or tied to another account.
• The id is the display name — whitespace stripped, e.g. Qwen3.8-Max; that is the id you give /model or --model.
• The context window is the largest tier the catalog advertises; offline, a four-tier fallback (Auto, Ultimate, Performance, Efficient) keeps the picker usable.
• Thinking is per model — effort-based models expose pi's levels as reasoning_effort values, toggle-only models treat any level as on, and off sends enable_thinking: false.
• Billing is in Credits, not USD — Qoder charges Credits and pi's cost figures are derived from them at 75 Credits per USD where a rate is measured; Qoder's own credits / original_credits / billable ride on the usage object.

Configuration

Everything below is optional; nothing needs to be set to use the extension.

Limits

• Credits are not dollars. pi's cost is a conversion of Qoder's Credits, not the bill; the bill itself lives in Qoder's ledger, per account.
• Retries come from pi's agent loop. settings.retry re-dispatches a failed turn; this provider adds no retry of its own on a billable chat POST.
• Cache warming follows pi's cacheWarming setting. When pi asks for a refresh, this provider approves it; every refresh spends Credits, bounded by QODER_WARM_BUDGET. An older qoder-warm-approval.json is no longer read and is safe to delete.
• The legacy gateway is the default. v2 is an opt-in, BYOK-oriented transport that only some models may use; AGENTS.md names the switch and the eligibility table.

Development

npm install
npm test              # offline suite, replays the recorded fixtures
npm run test:coverage # the same suite plus coverage/lcov.info
npm run test:live     # re-record live protocol fixtures (needs QODER_PAT / QODERCN_PAT)
pi -ne -e ./src/index.ts --list-models qoder   # this checkout, not the npm install

CI gates, the release process and the host contracts to re-check on a pi upgrade live in AGENTS.md.

Credits

This project is a fork of OnlyTomInSecond/pi-provider-qoder, which itself is a fork of simonsmh/pi-provider-qoder. Thanks to both for the protocol work this fork builds on.

License

MIT