Migrate OpenAI provider to Responses API and add Codex models

[datlechin]

Problem

The OpenAI provider in TablePro uses /v1/chat/completions. OpenAI's official guidance is that Responses API is the path for new projects. Reasoning models (gpt-5-codex, gpt-5.3-codex, gpt-5.5) lose capability on Chat Completions: tool calling with reasoning is not supported on GPT-5.4+, no reasoning summaries, no parallel tool calls, weaker caching. Today users who pick a Codex or GPT-5 model in TablePro get a degraded version of it.

Proposed solution

Refactor the OpenAI side of the AI layer:

1. Split OpenAICompatibleProvider into two providers:
   • OpenAIResponsesProvider — speaks /v1/responses, used by the OpenAI provider type.
   • ChatCompletionsProvider — keeps /v1/chat/completions for OpenRouter, Ollama, Custom.
2. Add reasoningDelta / reasoningSummary cases to ChatStreamEvent and route them in AnthropicProvider and OpenAIResponsesProvider. Render in a collapsible panel above each assistant reply.
3. Pass strict: true on function tools through ChatToolSpec.
4. Stay stateless — send full [ChatTurn] each call, do not use previous_response_id. Local chat history stays as-is.
5. Curate the OpenAI model list: gpt-5-codex, gpt-5.3-codex, gpt-5.5, gpt-5.4-mini surface at the top.
6. Optional reasoning-effort dropdown (low/medium/high/xhigh) shown only for reasoning models.
7. Image input in the chat composer for OpenAI and Anthropic.

Auth: existing AIKeyStorage (Keychain). No new credentials. ChatGPT Plus/Pro sign-in is not available to third-party apps; surface that in the API key field copy.

Out of scope (explicitly)

• No codex CLI binary bundled.
• No codex app-server subprocess.
• No new "Codex" provider type — Codex is a model under OpenAI.
• No previous_response_id server-side state.
• No Agents SDK / multi-agent handoffs.
• Azure OpenAI stays out for now (different endpoint shape, header style).

Alternatives considered

• Embed codex app-server over JSONL JSON-RPC (mirrors Copilot LSP). Rejected: ships a large Rust binary, agentic file editing is not a database-client feature.
• Add a separate Codex provider type. Rejected: marketing leaking into architecture; the canonical surface is the model name on the OpenAI provider.
• Keep Chat Completions everywhere. Rejected: blocks reasoning + tools on GPT-5.4+, worse caching, official guidance is to migrate.

Related database type

N/A / General