Refactor provider usage and especially oauth

[billybonks]

I reviewed the branch (main...HEAD) and updated this summary to include all public API changes.

Public API changes

1) Breaking: Claude Code custom path removed

• Removed model.start_with?("claude_code/") routing in LlmGateway::Client.
• Removed provider registration for anthropic_oauth_messages.
• Removed Claude Code adapter path (LlmGateway::Adapters::ClaudeCode::MessagesAdapter and related mappers).
• New usage pattern:
  • Use LlmGateway::Clients::Claude directly with Claude Code OAuth access tokens (sk-ant-oat...).
  • Claude now auto-detects Claude Code OAuth tokens and automatically:
    • switches to Claude Code OAuth headers,
    • prepends the Claude Code identity system prompt.
• New helper on Claude client:
  • LlmGateway::Clients::Claude#get_oauth_access_token(access_token:, refresh_token:, expires_at:) { |new_access, new_refresh, new_expires| ... }
  • Consumers are still responsible for persisting refreshed credentials.

2) Breaking: OpenAI Codex custom client removed

• Deleted LlmGateway::Clients::OpenAiCodex custom client class path.
• openai_oauth_codex provider now uses LlmGateway::Clients::OpenAi.
• LlmGateway::Adapters::OpenAiCodex::Client alias now points to LlmGateway::Clients::OpenAi.
• New usage pattern for Codex via OpenAI client:
  • chat_codex(...)
  • stream_codex(...)
  • get_oauth_access_token(...)

3) New/expanded request options in public chat/stream APIs

Across LlmGateway::Client.chat, responses, and adapter-backed usage, these user-facing options are now normalized consistently:

• cache_key
• cache_retention (short, long, none)
• reasoning (none, low, medium, high, xhigh)
• max_completion_tokens
• response_format (provider-specific mapping)

Provider behavior:

• OpenAI Chat Completions:
  • cache_key -> prompt_cache_key
  • cache_retention -> prompt_cache_retention (short -> in_memory, long -> 24h, none removes cache key)
  • reasoning -> reasoning_effort
• OpenAI Responses:
  • max_completion_tokens -> max_output_tokens
  • same cache mapping as above
  • reasoning -> { effort: ..., summary: "detailed" }
• OpenAI Codex:
  • inherits OpenAI Responses reasoning mapping
  • removes token-limit params for Codex endpoint compatibility
  • keeps prompt_cache_key but strips retention params from request body
• Anthropic:
  • max_completion_tokens -> max_tokens
  • reasoning -> thinking (budget token mapping)
  • response_format normalized to output_config
  • forwards cache_retention

4) New Anthropic cache behavior (publicly observable)

When using Claude with cache_retention, the client now automatically applies cache control:

• sets top-level cache_control, and
• sets cache_control on the last system block and last tool block.

For cache_retention: "long" against official Anthropic endpoint, ttl: "1h" is used.

5) Streaming error behavior changed (publicly observable)

Streaming mappers now raise LlmGateway::Errors::PromptTooLong when provider stream error payloads indicate context overflow (same semantic as non-streaming paths).

This applies to streaming through:

• Claude
• OpenAI Chat Completions
• OpenAI Responses / Codex

6) OAuth credential script behavior changed

Credential scripts now default to writing/merging into a shared auth file:

• ~/.config/llm_gateway/auth.json (override with LLM_GATEWAY_AUTH_FILE)
• providers are persisted under separate keys (anthropic, openai) in one JSON document.

7) Minor public-facing error detail improvement

OpenAI fallback API status errors now include response body text when no structured error payload is present.