@victormilk/pi-vibeproxy

Pi Coding Agent extension that exposes a CLIProxyAPIPlus instance ("VibeProxy") as two model providers inside Pi:

Provider name	Pi api	CLIProxyAPIPlus path
vibeproxy	anthropic-messages	POST /v1/messages
vibeproxy-openai	openai-completions	POST /v1/chat/completions

Streaming and tool/function calling are forwarded by @mariozechner/pi-ai's built-in adapters — this extension only configures the providers; it does not implement custom streaming.

Install

From npm:

pi install npm:@victormilk/pi-vibeproxy

From a local checkout of this monorepo:

pi install ./packages/pi-vibeproxy

For one-off testing without installing:

pi -e ./packages/pi-vibeproxy/src/index.ts

Configuration

Configuration lives at ~/.pi/agent/vibeproxy.json. All fields are optional.

{
  "baseUrl": "http://localhost:8317",
  "apiKey": "",
  "models": {
    "claude-sonnet-4-5": {
      "name": "Claude Sonnet 4.5 (VibeProxy)",
      "reasoning": true,
      "input": ["text", "image"],
      "contextWindow": 200000,
      "maxTokens": 16384,
      "cost": { "input": 3, "output": 15, "cacheRead": 0.3, "cacheWrite": 3.75 }
    },
    "claude-opus-4-6": {
      "thinkingLevelMap": { "xhigh": "max" }
    },
    "claude-opus-4-7": {
      "thinkingLevelMap": { "xhigh": "xhigh" }
    }
  }
}

thinkingLevelMap

Pi's thinking-level menu only shows xhigh when the model declares it via thinkingLevelMap. The extension applies built-in defaults for known Anthropic adaptive-thinking ids:

Model id pattern	Default thinkingLevelMap
*opus-4-7* / *opus-4.7*	{ "xhigh": "xhigh" }
*opus-4-6* / *opus-4.6*	{ "xhigh": "max" }
*sonnet-4-6* / *sonnet-4.6*	{ "xhigh": "max" }
(everything else)	(none — caps at high)

Declare thinkingLevelMap under a model entry to override these defaults or add a map for another model.

Context-window defaults

Opus 4.6 and Opus 4.7 also default to contextWindow: 1_000_000 and maxTokens: 128_000 (mirroring Pi's first-party catalog). All other Claude ids fall back to the family default 200_000 / 16_384. Override via contextWindow / maxTokens under the model entry.

Env-var overrides

• VIBEPROXY_URL — overrides baseUrl.
• VIBEPROXY_API_KEY — overrides apiKey.

Capabilities & defaults

CLIProxyAPIPlus's /v1/models returns only { id, owned_by }, so unspecified model capabilities use conservative defaults:

Field	Default
reasoning	false
input	["text"]
contextWindow	128000
maxTokens	8192
cost	all zeros

Declare per-model entries under models to unlock reasoning, image input, longer context windows, or accurate cost tracking.

Commands

• /vibeproxy-status — ping /v1/models and report model count.
• /vibeproxy-models — list discovered models grouped by owned_by.
• /vibeproxy-refresh — re-fetch /v1/models and re-register providers.

Smoke test (manual acceptance)

1. Start a CLIProxyAPIPlus instance (e.g. http://localhost:8317).
2. pi install ./packages/pi-vibeproxy (or pi -e ./packages/pi-vibeproxy/src/index.ts).
3. Run pi --list-models and confirm vibeproxy/... and vibeproxy-openai/... entries appear.
4. Launch Pi (pi), /model vibeproxy-openai/<some-id>, send a prompt — confirm tokens stream visibly.
5. /model vibeproxy/<some-anthropic-id>, send a prompt — confirm tokens stream visibly.
6. Issue a tool-call-bearing prompt on one of the two providers — confirm the tool call round-trips.

License

MIT