feature about chat completions reasoning, support claude model enable thinking & interleaved thinking , and support gemini-3-x thinking

[caozhiyuan]

This pull request introduces significant improvements to the translation logic between Anthropic and OpenAI message formats, especially around the new interleaved thinking protocol for Claude models. It adds support for thinking blocks in both streaming and non-streaming translation flows, updates protocol and API versions, and refines how message content is handled to ensure compatibility and correctness.

Anthropic <-> OpenAI translation improvements:

• Added support for interleaved thinking protocol for Claude models, including injecting reminders and enforcing thinking block requirements in system and user prompts. Thinking blocks are now correctly mapped to and from OpenAI's reasoning_text and reasoning_opaque fields, with appropriate filtering and content handling. [1] [2] [3] [4] [5] [6] [7]
• Updated the translation logic to include a thinking_budget parameter, calculated based on the model's capabilities and request payload, and passed through to copilot as needed. [1] [2]

Streaming translation and state handling:

• Refactored the streaming translation logic to modularize event handling, explicitly manage thinking block state, and ensure correct emission of content and tool call blocks. This includes new helper functions for handling message start, thinking text, content, tool calls, and finish events. [1] [2] [3]
• Updated the AnthropicStreamState interface and related state initialization to track thinkingBlockOpen. [1] [2]

Protocol and API version updates:

• Bumped Copilot and API version numbers to 0.35.0 and 2025-10-01 respectively, and changed the openai-intent header from conversation-panel to conversation-agent for more accurate intent signaling. [1] [2]

Message content and tokenization:

• Adjusted message token counting to ignore the reasoning_opaque field, preventing it from being double-counted in token calculations.
• Refined message content mapping to separate text and thinking blocks, ensuring only text blocks are included in the main message content for OpenAI compatibility. [1] [2] [3]

Type and interface enhancements:

• Extended the AnthropicThinkingBlock interface to include a signature field, supporting signature-based validation and filtering for thinking blocks.

These changes collectively ensure robust support for the new Claude thinking protocol, improve translation accuracy, and future-proof the integration with updated API versions.

[mrhanhan]

I believe that regularization could be used to transform the Claude model here, which would be better.

like:

function translateModelName(model: string): string {
  // Subagent requests use a specific model number which Copilot doesn't support
  if (model.startsWith("claude")) {
    const newModel = model.replaceAll(
      /(claude-(?:haiku|sonnet|opus)-\d+)-(\d+)(?:[.-]\d+)?/g,
      "$1.$2",
    )
    consola.log("Use Model:", model, newModel)
    return newModel
  }
  consola.log("Use Model:", model)
  return model
}

effect:

[caozhiyuan]

@mrhanhan Actually, this piece of translateModelName code is already useless . claude setting change like this

    "ANTHROPIC_BASE_URL": "http://localhost:4141",
    "ANTHROPIC_AUTH_TOKEN": "dummy",
    "ANTHROPIC_MODEL": "gpt-5.2",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "gpt-5.2",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "gpt-5-mini",
    "CLAUDE_CODE_SUBAGENT_MODEL": "gpt-5-mini",
    "DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "BASH_MAX_TIMEOUT_MS": "600000"

[caozhiyuan]

You can also see it in README.md https://github.com/ericc-ch/copilot-api?tab=readme-ov-file#manual-configuration-with-settingsjson