pi-bifrost

Native Pi model provider for a Bifrost gateway.

Routing is by vendor prefix from alias (anthropic.claude-…, us-gov.xai.grok-…) or a non-platform id (openai/gpt-6-sol). No version regexes. New GPT/Grok/Claude ids inherit the vendor wire.

/v1 is the OpenAI-compatible surface. Anthropic does not go there. The provider baseUrl is the gateway root; each model sets its own path.

Install

pi install npm:@josephjohncox/pi-bifrost

Or clone and link:

git clone https://github.com/josephjohncox/pi-bifrost.git
cd pi-bifrost && bun install
ln -sfn "$(pwd)" ~/.pi/agent/extensions/pi-bifrost

Connection lives in ~/.pi/agent/models.json — same file Pi already uses. Do not add a models array; that replaces the live catalog. Overrides use modelOverrides:

{
  "providers": {
    "bifrost": {
      "baseUrl": "https://bifrost.example",
      "api": "openai-completions",
      "apiKey": "$BIFROST_API_KEY",
      "headers": { "x-bf-mcp-include-tools": "" },
      "exclude": ["internal/*"],
      "modelOverrides": {
        "vendor/some-id": { "input": ["text"] }
      }
    }
  }
}

Last good catalog is cached at ~/.pi/agent/bifrost-catalog.json (ids/metadata only).

Session refresh

On factory load and every session_start (startup / reload / new / resume / fork) the provider refetches and reregisters. /bifrost-refresh does the same on demand.

Defaults (src/defaults.ts)

• x-bf-mcp-include-tools: "" on every request
• nemotron-nano-12b-v2 is text-only
• models.json modelOverrides / exclude always win

CLI

pi-bifrost list
pi-bifrost sync
pi-bifrost zed

sync only touches ~/.pi/agent/models.json connection fields. If this provider is loaded, do not put a models array on providers.bifrost.

Zed

pi-bifrost zed writes Zed's own LLM providers (Agent / inline model picker). It is not ACP and not the Pi TUI.

It updates ~/.config/zed/settings.json:

Each available_models row gets max_tokens (context), max_output_tokens (generation cap), and wire capabilities from the live catalog.

Thinking in Zed's UI is a settings flag, not a catalog field:

• OpenAI-compatible: Zed hides thinking unless reasoning_effort is not none. This writer sets "high" for reasoning models. That is a single default effort, not Pi's xhigh / max ladder.
• Anthropic-compatible: set "mode": { "type": "adaptive" } or Zed treats the model as non-thinking. Haiku 3 is left without a mode.

The command also refreshes any existing openai/anthropic-compatible entry that already points at the same gateway URL (or has no URL yet), so leftover aliases stay in sync.

Reload Zed after pi-bifrost zed. Pick Bifrost - OpenAI or Bifrost - Anthropic in the model dropdown.