Skip to content

changes.md

Latest commit

 

History

History
24 lines (18 loc) · 2.7 KB

changes.md

File metadata and controls

24 lines (18 loc) · 2.7 KB

Codex API Support: Feature Parity & UI Overhaul

Summary

This pull request introduces full feature parity and explicit UI support for the OpenAI Codex backend (chatgpt.com/backend-api/codex/responses). The codebase is now entirely backend-agnostic and smoothly transitions between Anthropic Claude and OpenAI Codex schemas based on current authentication, without losing features like reasoning animations, token billing, or multi-modal visual inputs.

Key Changes

1. Codex API Gateway Adapter (codex-fetch-adapter.ts)

  • Native Vision Translation: Anthropic base64 image schemas now map precisely to the Codex expected input_image payloads.
  • Strict Payload Mapping: Refactored the internal mapping logic to translate msg.content items precisely into input_text, sidestepping OpenAI's strict v1/responses validation rules (Invalid value: 'text').
  • Tool Logic Fixes: Properly routed tool_result items into top-level function_call_output objects to guarantee that local CLI tool executions (File Reads, Bash loops) cleanly feed back into Codex logic without throwing "No tool output found" errors.
  • Cache Stripping: Cleanly stripped Anthropic-only cache_control annotations from tool bindings and prompts prior to transmission so the Codex API doesn't reject malformed JSON.

2. Deep UI & Routing Integration

  • Model Cleanups (model.ts): Updated getPublicModelDisplayName and getClaudeAiUserDefaultModelDescription to recognize Codex GPT strings. Models like gpt-5.1-codex-max now beautifully map to Codex 5.1 Max in the CLI visual outputs instead of passing the raw proxy IDs.
  • Default Reroutes: Made getDefaultMainLoopModelSetting aware of isCodexSubscriber(), automatically defaulting to gpt-5.2-codex instead of sonnet46.
  • Billing Visuals (logoV2Utils.ts): Refactored formatModelAndBilling logic to render Codex API Billing proudly inside the terminal header when authenticated.

3. Reasoning & Metrics Support

  • Thinking Animations: codex-fetch-adapter now intentionally intercepts the proprietary response.reasoning.delta SSE frames emitted by codex-max models. It wraps them into Anthropic <thinking> events, ensuring the standard CLI "Thinking..." spinner continues to function flawlessly for OpenAI reasoning.
  • Token Accuracy: Bound logic to track response.completed completion events, fetching usage.input_tokens and output_tokens. These are injected natively into the final message_stop token handler, meaning Codex queries correctly trigger the terminal's Token/Price tracker summary logic.

4. Git Housekeeping

  • Configured .gitignore to securely and durably exclude the openclaw/ gateway directory from staging commits.
⚡