@stablekernel/pi-bifrost

Pi extension that registers a Bifrost LLM gateway as a model provider, with per-project virtual-key resolution delegated to gateway-cli.

What it does

• Registers two providers:
  • bifrost-openai — the OpenAI-compatible surface, every chat model the gateway serves.
  • bifrost-anthropic — the Anthropic Messages surface, Claude models only. The Anthropic wire carries prompt-cache control blocks that only Claude accepts, so non-Claude models are registered on the OpenAI surface instead of being listed twice with one copy broken.
• Fetches each surface's model list, limits and pricing through gateway-cli models --wire=… at startup — no catalog is checked in.
• Seeds reasoning capability per model: catalog-flagged models register reasoning: true. On the OpenAI surface they also get a thinkingLevelMap translating pi's thinking levels to the model's effort values (unapplicable levels are hidden, never passed through by name) plus supportsReasoningEffort, so pi's per-request effort picker sends the provider's own values. On the Anthropic surface the capability is flagged and pi's Claude thinking handling applies. Toggle-only reasoning models get no effort map. The default effort itself is pi's own thinking-level default.
• Never holds the key: pi resolves it per request via the !gateway-cli key print command form, so the per-project key cascade (environment override → project → user default) applies on every turn.

/bifrost-refresh

Reconciles the model list on demand, in pi and omp alike (one package, one implementation, one command):

• With gateway-cli on PATH it runs gateway-cli sync — the same reconcile as the background sync — and shows that command's own output. It rewrites the saved selection and the harness configs on disk.
• Without it it asks the gateway directly: $BIFROST_DISCOVERY_URL (or $BIFROST_GATEWAY_URL) with $BIFROST_VK, falling back to the key in ~/.config/bifrost/keys/default.env. It then reports how many models the gateway offers and which are new or gone against the ids this session registered. This mode reports only — it writes nothing, and the key is never printed.

Either way the answer arrives as a notification, and any failure is one readable line rather than an error thrown into the session.

Restart the session to use a newly added model. Pi and omp build a provider's model list once, while the session starts, and neither can re-register one to a running session — so the report names the new models and says to restart, and nothing here fakes a reload.

Datasheet supplement (off by default)

Wildcard-provider models — e.g. a fresh Fireworks launch that the CLI's models.dev source has not cataloged yet — never reach gateway-cli models, so they cannot be listed here either. When configured, this extension also reads the datasheet feed Bifrost itself syncs from and appends the wildcard-provider ids the CLI list lacks; CLI-listed ids always win. Applies to the OpenAI-compatible surface only; the Anthropic surface stays Claude-only.

Activation (first wins):

1. env GATEWAY_DATASHEET_URL
2. ~/.config/bifrost/datasheet.json — {"url": "..."}
3. <project>/.gateway/datasheet.json — same shape

The URL must be https://; a source with anything else is treated as absent.

Behavior: the feed is fetched once per hour and cached under ~/.cache/gateway-bifrost/datasheet-cache.json (or $XDG_CACHE_HOME), with a 3s hard timeout and a stale-cache fallback on failure. A failed first fetch also triggers a background cache fill, so a slow or cold feed costs one session rather than every start. Unconfigured or unreachable, the supplement is silently skipped — startup never blocks on it.

Provenance: supplement-only models are tagged limits_source: "datasheet" and cost_source: "datasheet" in the gateway-cli models payload they ride in. The tags stop there — pi's provider config has no provenance field, so they are not forwarded into the registered shapes.

Requirements

• gateway-cli on PATH (or the GATEWAY_CLI_BIN env var pointing at it), connected once with gateway-cli connect.

Without it nothing is registered — there is no credential to hand pi — but /bifrost-refresh still reports the gateway's live list from the environment (see above).

Install

pi install npm:@stablekernel/pi-bifrost

Works in omp (the Pi fork) too — its plugin loader accepts the same package and pi.extensions manifest:

omp plugin install npm:@stablekernel/pi-bifrost

License

UNLICENSED — published for installation convenience; all rights reserved. Intended for use within the organization that operates the Bifrost gateway.