fix: respect OPENCODE_DISABLE_MODELS_FETCH flag in ModelsDev.get()

[b3nw]

What does this PR do?

Fixes OPENCODE_DISABLE_MODELS_FETCH to actually work.
Previously the flag only prevented refresh() from fetching - but get() would still read cached data or fall back to build-time embedded macro data. So even with the flag set and cache deleted, users still saw models.dev models merged in.
Now get() returns {} immediately when the flag is set, so only locally configured providers/models appear.
export async function get() {
if (Flag.OPENCODE_DISABLE_MODELS_FETCH) return {}
// ...
}
Related: anomalyco#4959

How did you verify your code works?

Ran OPENCODE_DISABLE_MODELS_FETCH=true bun dev models and compared output to ~/.config/opencode/opencode.json:

• Output showed exactly 64 models across 7 providers (antigravity, gemini-cli, nano-gpt, nvidia-nim, chutes, longcat, mistral)
• Every model in output matched a model explicitly defined in local config
• No models.dev merged models appeared

Greptile Summary

This PR fixes OPENCODE_DISABLE_MODELS_FETCH to fully prevent models.dev data from being used. Previously, the flag only prevented refresh() from fetching new data, but get() would still read cached data or fall back to build-time embedded data. Now get() returns an empty object immediately when the flag is set.

Key changes:

• Added early return if (Flag.OPENCODE_DISABLE_MODELS_FETCH) return {} at the start of ModelsDev.get() on line 79
• When enabled, only locally configured providers/models appear (no models.dev providers merged in)
• PR author verified output shows exactly 64 models from local config with no models.dev models

Testing: Manual verification with OPENCODE_DISABLE_MODELS_FETCH=true bun dev models confirmed correct behavior.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The change is a simple, well-targeted fix that adds a single early return guard. The logic is straightforward, matches the existing pattern in refresh(), and has been manually tested. The empty object return is safe since the consuming code already handles empty provider records (as seen in provider.ts:608-609 where it uses mapValues which works fine with empty objects).
• No files require special attention

Important Files Changed

Filename	Overview
packages/opencode/src/provider/models.ts	Added early return in get() to respect OPENCODE_DISABLE_MODELS_FETCH flag

Sequence Diagram

sequenceDiagram
    participant User
    participant Provider
    participant ModelsDev
    participant Flag
    participant Cache
    participant Macro

    Note over User,Macro: Before this PR (Flag set but still returns data)
    User->>Provider: Get providers
    Provider->>ModelsDev: get()
    ModelsDev->>Flag: Check OPENCODE_DISABLE_MODELS_FETCH
    Flag-->>ModelsDev: true
    Note over ModelsDev: Flag is set, but get() continues...
    ModelsDev->>Cache: Read cached models.json
    alt Cache exists
        Cache-->>ModelsDev: Return cached data
    else No cache
        ModelsDev->>Macro: Load embedded data
        Macro-->>ModelsDev: Return build-time data
    end
    ModelsDev-->>Provider: Return models.dev providers
    Provider-->>User: Returns local + models.dev merged

    Note over User,Macro: After this PR (Flag properly respected)
    User->>Provider: Get providers
    Provider->>ModelsDev: get()
    ModelsDev->>Flag: Check OPENCODE_DISABLE_MODELS_FETCH
    Flag-->>ModelsDev: true
    ModelsDev-->>Provider: return {} (early exit)
    Provider-->>User: Returns only local providers

Loading

[shuv1337]

Merged via #308