AI provider integration uses raw fetch() — no native SDK support

[morozow]

Context

The original CompanionAgent example communicates with OpenAI via raw fetch() calls to https://api.openai.com/v1/chat/completions. There are no provider implementations for Anthropic or Google Gemini.

Problem

1. Manual HTTP handling. Error parsing, response normalization, header management, and retry logic are all hand-written. Native SDKs handle this automatically.
2. No Anthropic or Gemini support. Users who want Claude or Gemini must write their own integration from scratch, including format conversion (Anthropic's top-level system parameter, Gemini's Content parts format).
3. Inconsistent error handling. Raw fetch errors are generic — there is no distinction between auth failures, rate limits, network errors, and timeouts. MCP clients receive opaque error messages.
4. No token usage reporting. The raw fetch approach does not extract or normalize token usage from provider responses.

Expected behavior

• Built-in providers for OpenAI, Anthropic, and Google Gemini using their official native SDKs.
• Normalized response format (text, stopReason, usage) regardless of which provider generated it.
• Typed error classification (AUTH, UPSTREAM, TRANSPORT, TIMEOUT) for all providers.

[morozow]

Resolved. Built three provider implementations using official native SDKs:

OpenAIProvider (src/provider/providers/OpenAIProvider.ts) — uses openai npm package. Maps finish_reason to normalized stop reasons (stop → end_turn, length → max_tokens). Extracts prompt_tokens/completion_tokens → inputTokens/outputTokens. Supports AbortSignal for cancellation.

AnthropicProvider (src/provider/providers/AnthropicProvider.ts) — uses @anthropic-ai/sdk. Handles Anthropic-specific system prompt extraction (top-level system parameter). Extracts text from content blocks. Maps input_tokens/output_tokens.

GoogleGeminiProvider (src/provider/providers/GoogleGeminiProvider.ts) — uses @google/generative-ai. Converts ChatMessage[] to Gemini's Content parts format. Maps STOP/MAX_TOKENS/SAFETY to normalized stop reasons. Extracts promptTokenCount/candidatesTokenCount.

Error classification — each provider maps SDK-specific errors to BridgeError categories: AuthenticationError → AUTH, RateLimitError → UPSTREAM (retryable), connection errors → TRANSPORT (retryable), timeouts → TIMEOUT (retryable). Provider ID is included in all error details.

Peer dependencies — SDKs are optional peer dependencies. Users install only what they need. Missing SDK throws BridgeError CONFIG with installation instructions.