DevOps AI Toolkit

# PRD: Multi-Model AI Provider Support **Status**: ✅ Complete (Anthropic Production-Ready, Infrastructure Future-Ready) **Priority**: High (Implementation Complete) / Low (Alternative Provider Documentation - Blocked) **GitHub Issue**: [#73](https://github.com/vfarcic/dot-ai/issues/73) **Created**: 2024-08-20 **Completed**: 2025-10-05 **Documentation PRD**: [#140 - Multi-Provider Setup Documentation](./140-multi-provider-setup-documentation.md) (Blocked until better models) ## Executive Summary The DevOps AI Toolkit currently supports only Anthropic Claude models for AI-powered Kubernetes deployment recommendations. This PRD outlines the implementation of a provider-agnostic AI system using Vercel AI SDK, enabling support for multiple AI providers (OpenAI, Anthropic, Google Gemini) while maintaining backward compatibility. ### Problem Statement - **Vendor lock-in**: Limited to Anthropic Claude models only - **Cost constraints**: No flexibility to choose providers based on pricing - **Regional limitations**: Some providers may not be available in all regions - **Performance optimization**: Unable to select models based on specific use cases - **Fallback limitations**: No redundancy if Anthropic services are unavailable ### Solution Overview Replace the current `ClaudeIntegration` class with a new `AIProvider` abstraction powered by Vercel AI SDK, supporting multiple providers through a unified interface while maintaining all existing functionality. ## Technical Architecture ### Current State Analysis - **File**: `src/core/claude.ts` - Direct Anthropic SDK integration - **Usage**: 18 files depend on `ClaudeIntegration` class - **Methods**: `sendMessage()`, `generateYAML()`, `analyzeIntentForClarification()`, etc. - **Environment**: Requires `ANTHROPIC_API_KEY` ### Target Architecture ```typescript // New abstraction layer class AIProvider { // Vercel AI SDK integration // Support for multiple providers // Unified interface matching current ClaudeIntegration } // Provider configurations providers: { anthropic: claude-3-sonnet-4, openai: gpt-4o, google: gemini-1.5-pro } ``` ### Provider Support Matrix | Provider | Models | Environment Variable | Status | |----------|--------|---------------------|---------| | Anthropic | Claude Sonnet 4.5, Claude Opus 4 | `ANTHROPIC_API_KEY` | ✅ Existing (Validated 2025-10-03) | | OpenAI | GPT-4, GPT-4o, GPT-3.5 | `OPENAI_API_KEY` | 🔄 New | | Google | Gemini 1.5 Pro, Gemini Pro | `GOOGLE_API_KEY` | 🔄 New | **Note**: Anthropic SDK usage verified as isolated to `src/core/claude.ts` on 2025-10-03. All 10 dependent files only use `ClaudeIntegration` class, confirming existing abstraction layer. ## Implementation Plan ### Dependencies - `@ai-sdk/openai`: OpenAI provider integration - `@ai-sdk/anthropic`: Anthropic provider integration - `@ai-sdk/google`: Google Gemini provider integration - `ai`: Core Vercel AI SDK ### Configuration System ```typescript interface AIConfig { provider: 'anthropic' | 'openai' | 'google'; model?: string; fallbackProvider?: string; maxRetries?: number; } ``` ### Backward Compatibility Strategy 1. **Phase 1**: Implement new `AIProvider` alongside existing `ClaudeIntegration` 2. **Phase 2**: Create wrapper to maintain existing interface 3. **Phase 3**: Gradual migration of tools to use new provider 4. **Phase 4**: Deprecate old implementation (future release) ## Major Milestones ### ✅ Milestone 1: Core Architecture Implementation - [x] Install Vercel AI SDK dependencies (`ai`, `@ai-sdk/anthropic`, `@ai-sdk/openai`, `@ai-sdk/google`) - [x] Analyze current `ClaudeIntegration` usage patterns across 10 dependent files - [x] Design `AIProvider` interface based on actual usage needs (not speculation) - [x] Rename `ClaudeIntegration` → `AnthropicProvider` implementing `AIProvider` interface - [x] Create `VercelProvider` class implementing same `AIProvider` interface - [x] Implement provider factory pattern with configuration system - [x] Add provider-specific model defaults (`getDefaultModel()` method) **Success Criteria**: - All existing `ClaudeIntegration` methods work through new interface - Both `AnthropicProvider` and `VercelProvider` implement identical interface - Provider can be switched via environment variable (`AI_PROVIDER=vercel|anthropic`) - Provider-specific model defaults work correctly - Comprehensive unit tests pass for both providers ### ✅ Milestone 2: Multi-Provider Integration - [x] Implement Anthropic provider (migrating existing functionality) - [x] Add OpenAI provider with GPT-4 models - [x] Add Google Gemini provider integration - [x] Implement provider-specific model mapping - [x] Add error handling and fallback mechanisms **Success Criteria**: - All three providers work through unified interface - Provider switching works seamlessly - Fallback to secondary provider on failures ### ✅ Milestone 3: Tool Migration and Testing - [x] Migrate `recommend.ts` to new AI provider - [x] Migrate `generate-manifests.ts` to new interface - [x] Migrate `answer-question.ts` and other tools - [x] Update all prompt loading to work with new providers - [x] Comprehensive test suite with mocked providers **Success Criteria**: - All MCP tools work with any configured provider - Tests validate functionality across all providers - No breaking changes to existing API ### ✅ Milestone 4: Configuration and Environment - [x] Environment variable validation and setup - [x] Docker Compose configuration with multi-provider markers (future-ready) - [~] Development setup scripts for provider configuration (deferred - not needed for current Anthropic-only usage) - [x] Runtime provider switching capabilities - [~] Health checks for provider availability (deferred - can add when needed) **Success Criteria**: - ✅ Easy setup for Anthropic (production-ready) - ✅ Clear markers showing where to add future providers - ✅ Configuration files prepared for multi-provider expansion ### ⏸️ Milestone 5: Documentation and Examples (DEFERRED to PRD #140) **Status**: Blocked - Performance issues with alternative providers **Blocking Reasons**: - Google Gemini: 3x slower than Anthropic (52 min vs 17 min test suite) - OpenAI GPT-5: 2-3x slower, 91% pass rate (40/44 tests passing) **Deferred To**: [PRD #140 - Multi-Provider Setup Documentation](./140-multi-provider-setup-documentation.md) **Will Document When**: - ✅ Test suite duration within 2x of Anthropic baseline - ✅ 95%+ integration test pass rate - ✅ Next-gen models (Gemini 3, GPT-6) release **Original Scope** (moved to PRD #140): - [ ] Complete AI providers configuration guide - [ ] Updated setup documentation for all deployment methods - [ ] Code examples for each provider - [ ] Migration guide from Claude-only setup - [ ] Best practices for provider selection ### ⏸️ Milestone 6: Performance and Monitoring (PARTIALLY DEFERRED) - [x] Performance benchmarks across providers - [~] Cost analysis and optimization recommendations (deferred to PRD #139 - AI Provider Comparison & Benchmarking) - [x] Provider-specific error handling and retry logic - [~] Monitoring and observability for provider usage (deferred to PRD #137 - OpenTelemetry Tracing) - [~] Rate limiting and quota management (deferred - future enhancement) **Success Criteria Met**: - ✅ Performance benchmarks completed (Anthropic 100%, Gemini 100%, OpenAI 91%) - ✅ Provider-specific error handling implemented - ⏸️ Monitoring/cost analysis deferred to specialized PRDs ### ⏸️ Milestone 7: Feature Complete and Deployed (DEFERRED - Post-Release Activities) - [~] Feature flag for gradual rollout (not needed - already production-ready with Anthropic) - [~] Production deployment validation (ongoing operational task) - [~] User acceptance testing (post-release activity) - [~] Performance monitoring in production (deferred to PRD #137) - [~] Documentation complete and tested (deferred to PRD #140) **Deferral Rationale**: - Core implementation complete and production-ready - Anthropic provider fully tested and documented - Alternative provider documentation blocked by performance issues - Post-release validation activities beyond PRD scope ## Success Metrics ### Technical Metrics - **API Compatibility**: 100% backward compatibility with existing `ClaudeIntegration` interface - **Provider Coverage**: Support for 3+ major AI providers - **Response Time**: <5% performance degradation compared to direct Claude integration - **Error Rate**: <1% increase in AI-related errors - **Test Coverage**: >90% coverage for new AI provider code ### User Experience Metrics - **Setup Time**: Users can configure any provider in <5 minutes - **Provider Switching**: Change providers with single environment variable - **Documentation Quality**: Users successfully setup without support requests - **Cost Flexibility**: Users can optimize costs by provider selection ## Risk Assessment ### High Priority Risks 1. **API Rate Limits**: Different providers have different limits - **Mitigation**: Implement per-provider rate limiting 2. **Response Format Variations**: Providers may return different formats - **Mitigation**: Vercel AI SDK standardizes responses 3. **Breaking Changes**: Provider API changes could break functionality - **Mitigation**: Use stable SDK with version pinning ### Medium Priority Risks 1. **Cost Implications**: Different providers have different pricing - **Mitigation**: Document cost implications clearly 2. **Provider Availability**: Some providers may have regional restrictions - **Mitigation**: Support multiple providers as fallbacks 3. **Vercel AI SDK Issues**: Community reported concerns about AI SDK RSC development pause, error messages - **Status**: ✅ **Validated - Not Applicable** (2025-10-03) - **Analysis**: AI SDK RSC concerns only affect React Server Components usage. Our backend Node.js usage is unaffected. SDK remains actively developed (AI SDK 5 released 2025, 2M+ weekly downloads). - **Mitigation**: Continue monitoring SDK releases, maintain fallback capability to direct provider SDKs if needed ## Documentation Impact ### Files Requiring Updates - `README.md`: Multi-provider support in features section - `docs/mcp-setup.md`: Provider configuration instructions - `docs/setup/development-setup.md`: Environment variables for all providers - `docs/setup/docker-setup.md`: Multi-provider Docker configuration - `docs/setup/npx-setup.md`: Quick start with provider selection - `CLAUDE.md`: Development guidelines for provider abstraction ### New Documentation Required - `docs/ai-providers-guide.md`: Comprehensive provider configuration guide - Provider-specific examples and troubleshooting guides ## Future Considerations ### Phase 2 Enhancements (Future PRDs) - **Model Fine-tuning**: Support for custom fine-tuned models - **Local Models**: Integration with local LLM providers (Ollama, etc.) - **Advanced Routing**: Intelligent provider selection based on query type - **Cost Optimization**: Automatic provider switching based on cost thresholds - **A/B Testing**: Compare responses across providers for quality optimization ### Compatibility Planning - **API Evolution**: Plan for provider API changes and SDK updates - **Performance Monitoring**: Long-term performance trends across providers - **User Feedback**: Collect data on provider preferences and usage patterns ## Decision Log ### ✅ Decision: Vercel AI SDK Validated as Optimal Choice - **Date**: 2025-10-03 - **Decision**: Proceed with Vercel AI SDK as specified in original PRD (no alternative framework needed) - **Rationale**: - Comprehensive research comparing Vercel AI SDK vs LangChain vs LiteLLM vs LLM.js vs Ax - TypeScript-first with excellent type safety (matches our codebase requirements) - Built-in streaming support (critical - we use `stream: true` in `claude.ts:184`) - 18+ provider support, 2M+ weekly downloads, production-proven (Perplexity, Chatbase) - Active development (AI SDK 5 released 2025, AI Gateway available for production) - Simple API matching current usage patterns (minimal migration effort) - Framework-agnostic (not tied to Vercel platform for Node.js backend usage) - Reported concerns (AI SDK RSC development paused) don't affect our backend use case - **Alternatives Rejected**: - **LangChain**: Too complex, requires more boilerplate, overkill for simple text generation needs - **LiteLLM**: Python-first design, requires proxy server deployment (wrong architecture for TypeScript project) - **LLM.js/Ax**: Too new, insufficient production validation - **Impact**: No changes to original PRD approach - validation confirms technical direction - **Owner**: Development Team ### ✅ Decision: Interface-First Implementation Strategy - **Date**: 2025-10-03 - **Decision**: Adopt interface-first design approach based on actual `ClaudeIntegration` usage patterns - **Rationale**: - Anthropic SDK already well-isolated to `src/core/claude.ts` (verified across codebase) - 10 dependent files only use `ClaudeIntegration` class, not raw Anthropic SDK - Existing code structure already acts as effective abstraction layer - Can extract interface from real usage patterns rather than speculative design - Enables parallel development of both providers with validation before full migration - **Implementation Approach**: 1. Analyze current `ClaudeIntegration` usage patterns across 10 dependent files 2. Design `AIProvider` interface based on actual needs (not speculation) 3. Rename `ClaudeIntegration` → `AnthropicProvider` implementing interface 4. Create `VercelProvider` implementing same interface 5. Build provider factory with configuration system 6. Update imports to use interface type 7. Write integration tests validating both providers - **Impact**: - Updated Milestone 1 tasks to reflect interface-first approach - No code movement needed - just interface extraction and new implementations - Clearer migration path with testable intermediate states - **Owner**: Development Team ### ✅ Decision: Provider-Specific Model Defaults Required - **Date**: 2025-10-03 - **Decision**: Add provider-specific default model configuration to `AIProvider` interface - **Context**: Current implementation has hard-coded Anthropic model in `claude.ts:181`: `'claude-sonnet-4-5-20250929'` - **Rationale**: Different providers have different model naming conventions and optimal defaults - **Implementation**: Add `getDefaultModel()` method to `AIProvider` interface - **Default Model Mapping**: - Anthropic → `claude-sonnet-4-5` - OpenAI → `gpt-4o` - Google → `gemini-1.5-pro` - **Impact**: Adds requirement to Milestone 1 implementation tasks - **Owner**: Development Team ## Work Log ### 2025-10-03: Milestone 1 - Interface Design Phase Complete **Duration**: ~2-3 hours **Phase**: Core Architecture - AIProvider Interface Design **Completed PRD Items**: - [x] Analyze current `ClaudeIntegration` usage patterns across 10 dependent files - Evidence: Systematic analysis found only 3 methods actively used: `sendMessage()` (15 calls), `analyzeIntentForClarification()` (1 call), `isInitialized()` (1 call) - Identified 9 unused methods excluded from interface design (no bloat) - [x] Design `AIProvider` interface based on actual usage needs - Evidence: Created `src/core/ai-provider.interface.ts` with minimal 5-method interface - Interface based on real usage patterns, not speculation - Added only 2 new methods required for multi-provider support: `getDefaultModel()`, `getProviderType()` - [x] Implement provider factory pattern with configuration system - Evidence: Created `src/core/ai-provider-factory.ts` - Factory supports explicit configuration via `create()` and environment-based detection via `createFromEnv()` - Includes helper methods: `isProviderAvailable()`, `getAvailableProviders()`, `isProviderImplemented()` - [x] Add provider-specific model defaults - Evidence: Defined `PROVIDER_DEFAULT_MODELS` constant with defaults for all Phase 1 providers - Anthropic → `claude-sonnet-4-5-20250929`, OpenAI → `gpt-4o`, Google → `gemini-1.5-pro` **Architectural Decisions**: - **Extensible provider architecture**: Used `string` for provider type instead of enum to support 19+ Vercel AI SDK providers (not just 3) - **Provider-specific env vars**: Followed industry standard (`ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, `GOOGLE_API_KEY`) - **Minimal interface design**: Only included methods with evidence of actual usage **Files Created**: - `src/core/ai-provider.interface.ts` - Core AIProvider interface and type definitions - `src/core/ai-provider-factory.ts` - Provider factory with environment-based configuration **Next Session Priorities**: - Install Vercel AI SDK dependencies (`ai`, `@ai-sdk/anthropic`, `@ai-sdk/openai`, `@ai-sdk/google`) - Implement `AnthropicProvider` class (rename from `ClaudeIntegration`) - Implement `VercelProvider` class for OpenAI and Google support - Write integration tests validating both provider implementations **Milestone 1 Progress**: 57% complete (4 of 7 items) ### 2025-10-04: Milestone 1 - Provider-Agnostic Implementation Complete **Duration**: ~4-5 hours **Phase**: Core Architecture - AnthropicProvider Implementation & Provider-Agnostic Refactoring **Completed PRD Items**: - [x] Rename `ClaudeIntegration` → `AnthropicProvider` implementing `AIProvider` interface - Evidence: Created `src/core/providers/anthropic-provider.ts` implementing all 5 AIProvider methods - Migrated all functionality while keeping existing Anthropic SDK (no Vercel AI SDK for Anthropic) - All 5 interface methods implemented: `sendMessage()`, `analyzeIntentForClarification()`, `isInitialized()`, `getDefaultModel()`, `getProviderType()` - Original `ClaudeIntegration` removed, all imports updated to use `AIProvider` interface **Additional Work Completed (Provider-Agnostic Refactoring)**: - **Made entire codebase provider-agnostic**: - Removed all "claude" references from shared code (53+ locations across 18+ files) - Removed all "anthropic" references from shared code (8 files) - Updated `dotAI.claude` → `dotAI.ai` throughout codebase - Made all error messages, comments, and logs provider-agnostic - **Core files refactored** (`src/core/`): - `index.ts` - Changed interface from `claude: ClaudeIntegration` to `ai: AIProvider` - `error-handling.ts` - Generic AI service error messages - `capability-scan-workflow.ts` - Uses factory instead of direct API key checks - `unified-creation-session.ts` - Uses `aiProvider.isInitialized()` check - `schema.ts` - Accepts AIProvider via dependency injection - **Tool files refactored** (`src/tools/`): - `recommend.ts` - Uses `dotAI.ai` instead of `dotAI.claude` - `remediate.ts` - Generic AI provider error suggestions - `version.ts` - Complete refactor: `anthropic` → `aiProvider` with `providerType` field - `build-platform.ts` - Updated to use new interface - `answer-question.ts`, `generate-manifests.ts` - Updated imports - **Interface files refactored**: - `src/interfaces/mcp.ts` - Made comments generic - `src/mcp/server.ts` - Made initialization comments generic **Integration Tests Updated**: - `tests/integration/tools/version.test.ts` - Updated to validate new `aiProvider` structure - Fixed flaky test in `tests/integration/tools/remediate.test.ts` (pod crash timing issue) - **Test Results**: 43/44 passing (98% success rate) - **Build Status**: ✅ All TypeScript compilation successful, no errors **Architectural Decisions**: - **Interface validation strategy**: Kept Anthropic SDK for AnthropicProvider (didn't switch to Vercel AI SDK) - Rationale: Validates existing code works through new interface before adding complexity - Benefit: Proves interface design is correct and complete - **Provider-agnostic naming**: All shared code uses "AI provider" terminology instead of specific providers - **Backward compatibility**: Maintained all existing functionality while refactoring structure **Files Created/Modified**: - Created: `src/core/providers/anthropic-provider.ts` - Complete Anthropic implementation - Modified: 18+ core and tool files to use new interface - Modified: Integration tests to match new structure - Removed: `src/core/claude.ts` (migrated to anthropic-provider.ts) **Next Session Priorities**: - Install Vercel AI SDK dependencies (`ai`, `@ai-sdk/openai`, `@ai-sdk/google`) - Implement `VercelProvider` class for OpenAI and Google support - Write provider switching integration tests - Begin Milestone 2: Multi-Provider Integration **Milestone 1 Progress**: 71% complete (5 of 7 items) ⬆️ **UP from 57%** ### 2025-10-05: OpenAI GPT-5 Integration Testing & Bug Fixes **Duration**: ~3-4 hours **Phase**: Multi-Provider Integration - OpenAI Provider Validation **Completed Work**: - [x] **Multi-provider timeout accommodations** - Increased integration test timeouts to accommodate OpenAI's slower processing - remediate.test.ts: Manual Mode 5min → 20min, Automatic Mode 5min → 30min - recommend.test.ts: Full workflow 10min → 20min - manage-org-data-capabilities.test.ts: Full scan 16min → 45min - Rationale: OpenAI GPT-5 significantly slower than Anthropic/Gemini for AI-intensive operations - [x] **kubectl command escape sequence fix** (remediate.ts:1019-1020) - Root cause: OpenAI generates kubectl JSON parameters with escape sequences: `{\"apiVersion\"...}` - Fix: Added defensive code to strip escape sequences before command execution - Implementation: `fullCommand.replace(/\\"/g, '"')` - Impact: Works universally for all providers without breaking existing functionality - [x] **Test stability improvements** - Fixed race condition in pod crash detection with retry loop - Replaced fixed 30s wait with polling mechanism **Integration Test Results**: - **Anthropic Claude Sonnet**: 44/44 passing (100%), 17 min total duration ✅ - **Google Gemini 2.5 Pro**: 44/44 passing (100%), 52 min total duration (3x slower) ⚠️ - **OpenAI GPT-5**: 40/44 passing (91%), 2-3x slower ⚠️ **Technical Discoveries**: - **Performance**: Anthropic fastest, Gemini 3x slower, OpenAI 2-3x slower - **Command generation**: Different AI providers generate kubectl commands with varying formats (escape sequences) - **Compatibility**: Code fix approach preferred over prompt engineering for cross-provider consistency **Files Modified**: - src/tools/remediate.ts - Escape sequence fix, execution logic validation - tests/integration/tools/remediate.test.ts - Timeout increases, race condition fixes - tests/integration/tools/recommend.test.ts - Timeout increases - tests/integration/tools/manage-org-data-capabilities.test.ts - Timeout increases ### 2025-10-05: Documentation Deferred - Performance Gap Analysis **Duration**: ~1 hour **Phase**: Documentation Planning & Decision **Decision**: Defer user-facing documentation until performance improves **Rationale**: - Implementation complete and fully functional ✅ - Alternative providers have significant performance issues: - Gemini: 3x slower (52 min vs 17 min) - OpenAI: 2-3x slower, 91% test pass rate - Documenting slow features creates poor user experience - Next-gen models (Gemini 3, GPT-6) expected to improve performance **Actions Taken**: - Created PRD #140: Multi-Provider Setup Documentation (blocked) - Defined unblock criteria (2x performance, 95%+ tests) - Updated PRD #73 status to "Implementation Complete - Documentation Deferred" - Documented performance benchmarks for future comparison **Next Steps**: - Monitor for new model releases (Gemini 3, GPT-6) - Test new models when available using existing integration suite - Document providers that meet performance criteria (PRD #140) ### 2025-10-05: OpenAI Debugging Deferred - Strategic Decision **Duration**: 5 minutes **Phase**: Final Project Closure Decision **Decision**: Defer OpenAI test failure debugging until next-gen models release **Rationale**: - **Core implementation complete**: All 3 providers functional, architecture solid - **Primary provider perfect**: Anthropic at 100% test pass rate (44/44) - **Not worth debugging now**: Next-gen models (GPT-6) will likely fix reasoning issues - **Documentation already blocked**: PRD #140 waiting on performance improvements anyway - **Strategic timing**: Better to validate new models when released rather than debug old ones **Current State**: - ✅ Anthropic Claude Sonnet: 100% pass rate, 17 min duration (production-ready) - ✅ Google Gemini 2.5 Pro: 100% pass rate, 52 min duration (functional but slow) - ⚠️ OpenAI GPT-5: 91% pass rate (40/44), 2-3x slower (needs improvement) **Deferred Work** (to be revisited with new models): - Debug remaining 4 OpenAI test failures - Improve OpenAI test pass rate from 91% → 95%+ - Performance optimization investigation **Next Review Trigger**: Release of GPT-6 or significant GPT-5 improvements ### 2025-10-05: Deployment Configuration Updates - PRD Complete **Duration**: 15 minutes **Phase**: Final Infrastructure Preparation **Completed Work**: - [x] **Docker Compose updates** (`docker-compose-dot-ai.yaml`) - Added comment marking where to add `AI_PROVIDER` variable in future - Clarified ANTHROPIC_API_KEY required for AI analysis - Clarified OPENAI_API_KEY required for embeddings (not optional) - Future-ready structure for adding Google/other provider keys - [x] **Helm Chart updates** (`charts/`) - Added comment in `values.yaml` marking where to add future provider keys - Updated OpenAI key comment to clarify it's for embeddings - Added comment in `templates/deployment.yaml` marking where to add AI_PROVIDER env var - Maintained backward compatibility with existing deployments **Configuration Approach**: - **Anthropic-only** as documented production configuration - **Clear markers** showing where to add alternative providers later - **OpenAI preserved** for embeddings (required for capabilities/policies/patterns) - **No unsupported providers** included in configs **PRD Status**: ✅ **COMPLETE** - Core implementation: 100% complete - Infrastructure: Future-ready for multi-provider expansion - Documentation: Deferred to PRD #140 (blocked on model improvements) - Monitoring: Deferred to PRD #137 (OpenTelemetry) - Cost analysis: Deferred to PRD #139 (Benchmarking) ## References - **Vercel AI SDK Documentation**: https://ai-sdk.dev/docs/introduction - **Current Implementation**: `src/core/claude.ts` - **GitHub Issue**: [#73](https://github.com/vfarcic/dot-ai/issues/73) - **Provider Documentation**: - Anthropic: https://docs.anthropic.com/ - OpenAI: https://platform.openai.com/docs - Google AI: https://ai.google.dev/docs --- **Last Updated**: 2025-10-05 (PRD Complete - Infrastructure Future-Ready) **Next Review**: When new major AI models released (GPT-6, Gemini 3) for alternative provider documentation **Stakeholders**: DevOps AI Toolkit Users, Contributors, Maintainers