# Model catalog comparison

Environment for every capture: Linux web client, Chromium preview browser, dark appearance, 1440×900 CSS pixels (2880×1800 PNG), OpenCode 1.19.6. Both T3 homes began empty with T3CODE_DEV_SEED=0. No live user data was copied.

Base: 634ee43deb6d5fc0feae77153d33df1e97ffb303. Head: f10e300c1a787d0b0cdfdb14d183edca352b7917.

For each checkout, install with `vp i`, then start `T3CODE_DEV_SEED=0 OPENCODE_CONFIG="$CATALOG_JSON" vp run dev`. CATALOG_JSON is the downloaded catalog.json absolute path. Only the synthetic catalog provider is enabled in that config. No model request is sent to it.

Create an empty git repository at DEMO_DIR. Add it with `node apps/server/src/bin.ts project add "$DEMO_DIR" --title 'Model catalog demo' --base-dir .t3`. Create history with `node apps/server/src/bin.ts thread start --base-dir .t3 --provider devReplay --model showcase --title 'Catalog comparison' "$DEMO_DIR" 'Show the model catalog.'`.

Pair an incognito browser with the checkout's startup pairing token. In Settings → Providers, choose Add provider → OpenCode → Next → Next → Add instance. Wait for its six models to load. OPENCODE_CONFIG is inherited from the dev server environment. Return to a new draft in Model catalog demo and hide the thread details panel.

Run runCatalogProof(tools, tabId) from driver.js in the tool orchestration environment. It clears the composer and types @Hydrogen, @Carbon, @Wonder and @Argon in sequence, then searches the model picker for Gemini 4, clears search, selects the OpenCode provider, and expands Show more. It records the complete sequence and takes each still only after two images 500 ms apart are pixel-identical.

Observed on base: Hydrogen offered; Carbon, Wonder and Argon absent from mentions. Only Hydrogen and Ember appear before Show more. Search still finds all three Gemini models.

Observed on head: Hydrogen, Carbon and Wonder offered; Argon absent. Carbon, Hydrogen, Ember and Wonder appear before Show more. Argon and Spark remain behind Show more. Search still finds all three Gemini models.

The scenario runner was not used because its documentation excludes provider-specific behavior; this proof exercises the real OpenCode catalog discovery path rather than the replay provider's model inventory. Replay was used only to create synthetic thread history.
