Assistant: unify provider enablement and model settings

[sharon-wang]

Summary

• addresses Positron Assistant: Control provider list via Settings UI #10866

Main changes

• Individual provider enable/disable settings: Replace positron.assistant.enabledProviders array with individual positron.assistant.provider.<name>.enable boolean settings for each provider

• Per-provider model preferences: Replace positron.assistant.models.preference.byProvider object with individual positron.assistant.models.preference.<name> settings

• Per-provider custom model overrides: Replace positron.assistant.models.custom object with individual positron.assistant.models.overrides.<name> settings

• Chat panel state awareness: Welcome message now reflects enabled providers state

  • When no providers enabled
    
  • When at least one provider is enabled, but none are authenticated
    

• All providers can now be disabled: Previously, Copilot and Anthropic were always shown in the model picker regardless of settings. Now all providers (including Copilot and Anthropic) respect the enable/disable settings and are properly filtered from the model picker when disabled.

• Command positron-assistant.configureModels renamed to positron-assistant.configureProviders: better matches what it does!

• Provider config modal now takes provider id: when a provider id is passed to the configureProviders command, the modal will scroll to the selected provider. This is available via the provider enablement settings by clicking on the "language model provider dialog" link, and also via the gear icon in the Manage models... > Language Models editor

  open_provider_modal_to_provider.mp4

• Remove unused providers: Removed deprecated/unused language model providers that were never exposed in the UI:

  • Chat providers: Vertex AI, Mistral, Ollama, OpenRouter -- these providers may be compatible with the Custom Provider provider
  • Completion providers: DeepSeek, Ollama and related completion code that is unused

• Remove getSupportedProviders() API method: replaced with dynamic provider registration through ILanguageModelsService. This didn't appear to be used anywhere anyways.

Implementation decisions to call out

• models.custom renamed to models.overrides: Originally I wanted to use models.custom.<providerName> for the new per-provider settings, but ran into an issue with VS Code's configuration API: it returns a value for models.custom when any child setting like models.custom.anthropic is set, even though these are semantically different settings. This caused the migration to repeatedly trigger because it couldn't distinguish between the old object format and the new per-provider settings. Renaming to models.overrides creates clean setting namespace separation, though a different name than before.

  • Happy to switch from overrides to something else -- let me know if you have suggestions!

• Settings organized by domain, not by provider: Settings are grouped by their function (provider.* for enablement, models.* for model config) rather than having all settings for a provider under one prefix. This means a provider's settings are spread across setting namespaces, but provides better semantic grouping and discoverability in the Settings UI. Users can still find all settings for a specific provider by searching for the provider name.

• Migration preserves old values: When migrating from the old settings format, old values always take precedence. For example, if a user had "positron.assistant.enabledProviders": ["anthropic-api", "openai-api"], after migration they'll have positron.assistant.provider.anthropic.enable: true and positron.assistant.provider.openAI.enable: true. The old setting is removed after successful migration, and a one-time notification is shown.

  • Note: I'm a bit worried that the migration notification will show repeatedly for Workbench users who have admin-configured enabledProviders settings (since the admin setting can't be removed by migration). There is a "Don't Show Again" button to dismiss the notification though.

Settings change summary

Deprecated (with automatic migration):

• positron.assistant.enabledProviders → positron.assistant.provider.<name>.enable
  • 
  • 
• positron.assistant.models.preference.byProvider → positron.assistant.models.preference.<name>
  • 
  • 
  • 
• positron.assistant.models.custom → positron.assistant.models.overrides.<name>
  • 
  • 
  • 

Provider setting names: Each provider class defines its settingName property (e.g., anthropic, amazonBedrock, openAI). See extensions/positron-assistant/src/providers/*/ for the full list.

Future possible enhancements

• Auto-enable provider when clicking its settings gear icon (currently requires enabling first, then configuring)
• Display provider maturity status (experimental, preview, GA) in configuration modal

Release Notes

New Features

• Assistant: enable or disable individual language model providers via settings (Positron Assistant: Control provider list via Settings UI #10866)

Bug Fixes

• Assistant: ensure chat panel state reflects enabled providers

QA Notes

@:positron-notebooks
@:assistant

• updated e2e test fixtures and tests with updated command and settings names

Added/updated Assistant extension integration tests:

• providerMigration.test.ts - new tests for migrating legacy settings to new per-provider format
• config.test.ts - tests for settings validation
• modelDefinitions.test.ts - updated for new models.overrides setting name

Run assistant extension tests locally with npm run test-extension -- -l positron-assistant.

---

Additional notes on changes

• extensions/positron-assistant/src/providers/*: changes to files in here were to add the settingName to each provider, and remove any providers not in use
• extensions/positron-assistant/src/test/*: updated existing tests to match the new settings; added new tests for settings migration, provider enablement, whether settingName is adequately set for providers (should fail if a new provider is added but doesn't have a corresponding setting in package.json)
• extensions/positron-assistant/src/config.ts: getEnabledProviders is now handled via API, and we also pass the selected provider ID to the config dialog if a user clicks on the config dialog link in the settings UI
• src/vs/workbench/contrib/chat/*: various fixes to ensure that the chat UI shows only enabled provider information. I wanted to avoid having to add too many changes here in the upstream code, but we need to check for enabled providers and invalidate the cached provider/model in order for the chat UI to properly reflect the provider enablement state.

[github-actions]

E2E Tests 🚀
This PR will run tests tagged with: @:critical @:positron-notebooks @:assistant

^{readme  valid tags}

[sharon-wang]

Note

The issue this PR addresses is now targeted for 2026.03, so this should be merged after we branch for 2026.02!

[sharon-wang]

Confirmed with Jon that the e2e test failure with Anthropic is a known issue to be addressed separately, so I think this is ready for another review pass!

1) [***-electron] › test/***/tests/positron-assistant/positron-assistant.test.ts:344:6 › Positron Assistant Model Picker Default Indicator - Multiple Providers › Verify default model indicators and ordering for multiple providers @:win @:assistant @:web 

    Error: Expected Anthropic models to be present

    expect(received).toBe(expected) // Object.is equality

    Expected: true
    Received: false

    Call Log:
    - Timeout 30000ms exceeded while waiting on the predicate

      307 | 				expect(hasAnthropic, 'Expected Anthropic models to be present').toBe(true);
      308 | 			}
    > 309 | 		}).toPass({ timeout: 30000 });
          | 		   ^
      310 | 	});
      311 |
      312 | 	test.afterAll('Sign out of providers and clean up', async function ({ app, settings }) {
        at /__w/positron/positron/test/***/tests/positron-assistant/positron-assistant.test.ts:309:6

(this test passes for me locally)