Skip to content

CATALOG.md

Latest commit

 

History

History
79 lines (58 loc) · 3.48 KB

CATALOG.md

File metadata and controls

79 lines (58 loc) · 3.48 KB
Raw

How the model list is built

What you see in the Copilot Chat picker is not the raw GET /v1/models payload. The extension shapes it, and the numbers below come from a real OmniRoute instance measured on 2026-08-18 (2345 raw catalog entries).

Stage Entries What happened
Raw catalog, server default 2345 every model listed twice (see below)
?prefix=alias 1396 949 duplicate ids dropped, zero models lost
Non-chat models removed 1319 26 image / video / audio / rerank rows

1. Why the raw catalog lists everything twice

OmniRoute's MODELS_CATALOG_PREFIX_MODE defaults to dual, which advertises each model under both its short alias prefix and its canonical provider prefix, for backward compatibility with client configs that hardcoded either one:

cc/claude-sonnet-4-6        ← alias prefix
claude/claude-sonnet-4-6    ← canonical prefix — same model

Both route fine; they are the same model. In a picker they read as duplicates.

The extension asks for GET /v1/models?prefix=alias, a supported per-request override (see tests/unit/models-catalog-low-noise-flag.test.ts in the OmniRoute repo), so the server sends one id per model without changing its global setting for other clients.

⚠️ Never use ?prefix=canonical. In OmniRoute's catalog.ts the canonical row is only emitted when canonicalProviderId !== alias, so providers without a distinct alias would emit nothing — that mode silently loses models. alias is the safe direction, and was verified as lossless: all 949 dropped ids were mirrors.

A second, server-independent guard drops mirror rows via their parent back-reference, so an OmniRoute too old to honor ?prefix is covered too. On the measured instance both paths converge on the same 1319 models.

2. Why some models never appear

Specialty registries. The catalog also carries image, video, audio, rerank, embedding and moderation models. OmniRoute rejects those on a chat request:

HTTP 400 — Model 'cheaperinference/nano-banana-pro' is an image-generation model
and cannot be used on /v1/chat/completions. Use POST /v1/images/generations instead.

They are filtered by the catalog's type field, so they never reach the picker.

Responses-API models are not filtered. Every Codex / GPT-5.x entry is listed as supported_endpoints: ["responses"], but OmniRoute translates those for /v1/chat/completions — cx/gpt-5.5-low and cx/gpt-5.6-sol-low both answer HTTP 200. Treating "does not list chat" as "unusable" would drop 26 working models, so only genuinely non-conversational surfaces are excluded.

Your modelFilter setting. omnicopilot.modelFilter is a regex (falling back to substring matching if the regex is invalid) applied to the model id. Empty by default.

3. Why you see providers you never configured

OmniRoute advertises models from:

  1. every provider with an active connection, plus
  2. every noAuth provider — the ones that need no credential at all. This is the "90+ free providers" part of OmniRoute, and it is intentional.

To hide the keyless ones, add them to blockedProviders in the OmniRoute dashboard settings. Nothing needs to change in the extension.

4. Refreshing

The extension caches the catalog per discovery call. OmniRoute: Refresh Models (Command Palette) or the ↻ link in the panel re-queries the server and tells VS Code the list changed. Changing any omnicopilot.* setting refreshes automatically.