A Pi package that shows the real downstream model behind OmniRoute combos in Pi's footer — updated as soon as the first real token arrives.

OmniRoute combos like auto/claude-sonnet, auto/coding:reliable, or fusion_free are not models — each request is routed to a real upstream model. Pi only sees the combo id in message.model, so "which model am I actually talking to?" is surprisingly hard to answer at a glance.

Two things make it worse:

• OmniRoute's streaming responses only carry x-omniroute-route-class in HTTP headers. The real-model header (x-omniroute-model) is emitted as an SSE trailer comment at the end of the stream, which Pi's SSE parser drops.
• pi-ai's OpenAI-completions parser writes output.responseModel ||= chunk.model. OmniRoute prepends a model: "omniroute" keepalive chunk, so responseModel latches to "omniroute" and never updates when the real model arrives in later chunks.

Net effect: on many combos, Pi's responseModel is stuck on the literal string "omniroute" for the entire stream.

This extension patches globalThis.fetch once and tee()s any OmniRoute event-stream response, reading the real model straight out of the SSE body — before pi-ai's parser gets to it. As soon as a chunk arrives whose model isn't "omniroute", the footer flips to combo → real-model.

at request start:              auto/claude-sonnet …
after first real token:        auto/claude-sonnet → claude-opus-4-7
gateway fallback (only if OmniRoute never reveals a real model):
                               auto/claude-sonnet → (gateway fallback)

The fetch tee is scoped: it only activates for responses whose content-type is text/event-stream and which carry an x-omniroute-route-class header. Non-OmniRoute traffic is untouched.

For non-streaming responses the extension falls back to the x-omniroute-model header (via Pi's after_provider_response hook), then to message.responseModel.

From GitHub:

pi install git:github.com/xerosanyam/pi-omniroute-model-trace

Reload Pi (/reload in the TUI) or restart after installation.

The footer chip omni-model updates automatically after every OmniRoute reply. To re-show the last value as a notification:

/omni-model

The package also ships a skill (omniroute-model-trace) that teaches the agent how the extension resolves the real model and how to cross-check via session files, gateway call logs, or a direct curl to the gateway.

• Pi Coding Agent
• Pi routed through an OmniRoute provider (provider id omni), e.g. via the omniroute-pi-ext-integration extension

MIT