Fix: inference-parser misses /chat/completions requests from OpenAI-compatible proxies (opencode, litellm)

[msteinder]

Problem

The inference-parser plugin dispatches on the request path in OnRequest with:

case "/v1/chat/completions", "/v1/completions":

OpenAI-compatible proxies such as litellm (used by opencode) strip the /v1 prefix and post to /chat/completions directly. This falls through to the default arm, so pctx.Extensions.Inference is never populated — resulting in missing model, token counts, and method fields in abctl session views.

Observed in abctl output when running opencode through authbridge-proxy in demo mode:

9     14:13:56.91   out   │┌ req   —         —                   ete-litellm.ai-mode…
9     14:13:59.27   out   │└ resp  —         —         200  2.36s ete-litellm.ai-mode…

Compare with Claude Code (which posts to /v1/messages — matched by the Anthropic path), where all fields are populated correctly.

Root cause

authbridge/authlib/plugins/inferenceparser/plugin.go — the OnRequest switch only covers /v1/chat/completions and /v1/completions. The /v1-less variants /chat/completions and /completions are not listed.

The response path (OnResponseFrame) is unaffected — its default branch already routes to the OpenAI parsers, so response parsing works once the request side is fixed.

Fix

Extend the OnRequest switch case to include the /v1-less variants:

// Before
case "/v1/chat/completions", "/v1/completions":

// After
case "/v1/chat/completions", "/v1/completions", "/chat/completions", "/completions":

No other changes needed.