Repository navigation
Releases
Releases · kabaBZ/magpie-otari-provider
v0.3.0 - two-layer routing with Otari policies and Otari-Tags
Two-layer routing
Magpie picks the intent tier; Otari picks the real provider, enforces budget, and handles failover.
Added
- Otari routing policies are exposed as Magpie models by policy name (
otari/fast,otari/strong,otari/long). Otari-Tagson every request:app=magpie,session=<id>,model=<id>, so usage can be filtered per conversation.npm run setup:routingcreatesfast/strong/longOtari policies with provider failover.docs/two-layer-routing.mdwith the full setup, verification and demo script.
Fixed
- Organization provider-key models now use the callable selector (
<key-prefix>:<model>) instead of a native id Otari does not accept. - Policy models send the bare policy name, matching Otari's model spelling.
Verified
magpie provider test otari->chat 592 ms - fast- Local gateway: a quick question routes to
otari/fast; a test-writing prompt routes tootari/strong. - 41 unit tests pass.
Upstream feedback
Full Changelog: v0.2.1...v0.3.0
v0.2.1 — organization provider labels actually work
Fixes the organization-provider label introduced in v0.2.0.
Fixed
- The plain
GET /api/v1/modelslisting reports every model as
pricing_source: "configured". Otari only marks a model
pricing_source: "organization"when the listing is asked for one
organization provider key by name (?provider=<key>). - The plugin now reads
GET /organizations/me/provider-keys, then asks
GET /models?provider=<key>for each usable key, and overlays those rows
onto the catalog list. No key name is hard-coded. - Models served by your own key now show the key name, e.g.
[mydeepseek] deepseek/deepseek-flash.
Notes
- Falls back to the organization name, then the provider name, when a
key listing cannot be read. - Model discovery still works when the Dashboard session has expired;
the prefix is simply not added.
Install
magpie plugin add npm:magpie-otari-providerv0.2.0 — label models served by your own provider keys
Models served by provider keys your organization added under Organization → Providers are now labelled, so they can be told apart from deployment-hosted ones.
Added
pricing_source: "organization"models carry anorganizationOwnedflag.- Display name is prefixed in this order:
- the provider key name (
[mydeepseek]), when the Dashboard session can read it and it is unambiguous for that provider; - the organization name (
[kaba's organization]); [organization Inference providers: <provider>];[organization Inference providers]when the provider is unknown.
- the provider key name (
The prefix is display-only: model IDs and requests are unchanged.
Notes
- The label needs a Dashboard session for the key and organization names. Without one it falls back to the provider name, and model discovery still works.
Install
magpie plugin add magpie-otari-providerv0.1.1 — publishable package fix
Fixes a release-blocking bug in v0.1.0.
Fixed
- Removed a self-referential
file:.pkgtest/...entry fromdependencies.
v0.1.0 failed to install for every user because npm tried to unpack
that local tarball. v0.1.1 has no runtime dependencies at all.
Install
magpie plugin add magpie-otari-providerv0.1.0 — Otari provider for Magpie
First release of the Otari provider plugin for Magpie.
Features
- Dynamic Chat model discovery from
GET /api/v1/models(145 models on the hosted gateway). - Per-million-token pricing, with
/pricing/currentas a fallback. - Dashboard session login by cookie or email/password, plus API-key-only mode.
- Balance, key budget, reset time, 30-day usage and routing visibility.
- OpenAI-compatible Chat Completions, streaming and non-streaming.
Install
magpie plugin add magpie-otari-providerNotes
freeis only set when Otari states a per-token input and output price of zero.- Tiered prices show the base tier; a long-context request can cost more.