Skip to content

Proxy external mode hardcodes native FC — no prompt-injection fallback #53

@antoinezambelli

Description

@antoinezambelli

Proxy external mode hardcodes native FC — no prompt-injection fallback

Problem

ProxyServer in external mode creates a LlamafileClient with mode="native" (proxy.py:169). There's no way to configure prompt-injected FC mode when connecting to an external backend.

This means the proxy only works with backends/models that support native tool calling via OpenAI format. Models without native FC support (or backends like vLLM serving models with incomplete tool-calling templates) can't use forge's prompt-injection fallback through the proxy.

In managed mode this isn't an issue — the eval runner and setup_backend() both accept mode and pass it through.

Proposed fix

Add --mode flag to forge-proxy CLI and mode parameter to ProxyServer.__init__(). Pass through to LlamafileClient(mode=...) in external mode.

Context

Came up from a user asking about vLLM + forge. vLLM serves OpenAI-compatible endpoints but not all models have native FC templates. Prompt-injection mode would make the proxy work with any OpenAI-compatible backend regardless of native FC support.

Metadata

Metadata

Labels

No labels
No labels

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions