Unified Provider Auto-Detection Across Host and Android

Summary

When no default model is configured, Trailblaze should scan the environment for available provider credentials and automatically select a working provider. This logic exists on Android today but is missing from the host/desktop side. This plan proposes unifying the auto-detection into shared code that both platforms use.

Problem

Today, Android and desktop handle “no model configured” very differently:

Behavior	Android (AndroidLlmClientResolver)	Desktop (TrailblazeDesktopAppConfig)
Scan env for available tokens	Yes — PROVIDER_PRIORITY list	No
Fallback when nothing configured	First provider with a token	Hardcoded OpenAI GPT-4.1
Fresh install with only ANTHROPIC_API_KEY	Works — picks Anthropic	Fails — tries OpenAI, no key

The desktop side already scans tokens for UI filtering (getCurrentlyAvailableLlmModelLists() via JvmLLMProvidersUtil), but doesn’t use that same logic to select the default model. The defaultLlmModel constructor parameter is hardcoded per config class.

Proposed Design

1. Extract shared auto-detection to common module

Create a LlmProviderAutoDetector (or similar) in trailblaze-models or trailblaze-common (wherever TrailblazeLlmProvider lives) that encapsulates the “scan for first available provider” logic.

object LlmProviderAutoDetector {
  /**
   * Default priority order for auto-detecting a provider from available credentials.
   * First provider with an available auth token wins.
   */
  val DEFAULT_PROVIDER_PRIORITY = listOf(
    TrailblazeLlmProvider.OPENAI,
    TrailblazeLlmProvider.OPEN_ROUTER,
    TrailblazeLlmProvider.ANTHROPIC,
    TrailblazeLlmProvider.GOOGLE,
    TrailblazeLlmProvider.OLLAMA,
  )

  fun detectFirstAvailable(
    providerPriority: List<TrailblazeLlmProvider> = DEFAULT_PROVIDER_PRIORITY,
    isProviderAvailable: (TrailblazeLlmProvider) -> Boolean,
    getFirstModel: (TrailblazeLlmProvider) -> TrailblazeLlmModel?,
  ): TrailblazeLlmModel? { ... }
}

The isProviderAvailable and getFirstModel lambdas keep the detector platform-agnostic — Android checks instrumentation args, desktop checks env vars, but the priority logic is shared.

2. Android: refactor AndroidLlmClientResolver to delegate

Replace the inline PROVIDER_PRIORITY loop in resolveModel() stage 3 with a call to the shared detector. The resolution chain stays the same (config → instrumentation arg → auto-detect), just the auto-detect step delegates.

Key file: opensource/trailblaze-android/src/main/java/xyz/block/trailblaze/android/AndroidLlmClientResolver.kt

3. Desktop: use auto-detection in getCurrentLlmModel() fallback

TrailblazeDesktopAppConfig.getCurrentLlmModel() currently falls back to defaultLlmModel (hardcoded OpenAI). Change this to:

1. Try saved settings (current behavior)
2. Try llm.defaults.model from YAML config
3. Auto-detect from environment using the shared detector
4. Error with a helpful message

This replaces the defaultLlmModel constructor parameter — or at minimum makes it the last resort after auto-detection.

Key files: - opensource/trailblaze-host/src/main/java/xyz/block/trailblaze/desktop/TrailblazeDesktopAppConfig.kt — getCurrentLlmModel() at line 156 - opensource/trailblaze-desktop/src/main/java/xyz/block/trailblaze/desktop/OpenSourceTrailblazeDesktopAppConfig.kt — passes hardcoded defaultLlmModel = OpenAITrailblazeLlmModelList.OPENAI_GPT_4_1 - Organization-specific subclasses of TrailblazeDesktopAppConfig follow the same pattern

4. Organization-specific overrides

Organization-specific config subclasses can include additional providers (e.g., a corporate LLM endpoint) in their provider priority by passing a custom providerPriority list to the shared detector rather than using the default.

Open Questions

• Should the YAML config gain a provider_priority field? Current decision: no. Keep priority in implementation for now, document it, and link to source. Revisit if users actually need to customize it.
• What happens in the desktop UI when auto-detect picks a provider? The settings dropdown should probably reflect the auto-detected selection so the user sees what was chosen. Need to check if TrailblazeSettingsRepo handles “no saved provider” cleanly.
• CLI error message — TrailblazeCli.kt:1061 already prints “No LLM configured” when getCurrentLlmModel() throws. With auto-detection, this error should only fire when genuinely no tokens are available, which is the right behavior.

Scope

This is future work — not part of the current externalize-llm-config branch. The immediate change (removing hardcoded model from ExamplesAndroidTrailblazeRule) is safe because the YAML config handles that case. This plan addresses the broader inconsistency.