feat(core): full end-to-end integration of customReasoningFields across serialization and provider layers

[Yash-200608]

Description

This PR delivers a comprehensive architectural solution for issue #10292. Local LLM gateways, private enterprise proxies, and custom endpoints (like vLLM, Ollama, or WatsonX) often stream reasoning tokens inside arbitrary, non-standard JSON keys. Without explicit tracking, these tokens are dropped by standard parsers or leak unformatted into the main chat block.

Instead of a shallow string-matching workaround, this change introduces customReasoningFields as a core architectural property across Continue's configuration, validation, and stream-parsing layers. It maps non-standard payload delta keys natively into Continue's internal reasoning streams, routing them perfectly to the collapsible UI "Thinking" container.

Related Issues

Closes #10292

Changes Made

• Architecture & Typings (core/index.d.ts, JSON/YAML/control-plane schemas): Added and fully integrated customReasoningFields?: string[] across model configurations, ensuring fields pass dynamically through model-add flows and browser serialization layers.
• Encapsulation Refactor (BaseLLM, OpenAI, WatsonX): Exposed the public options accessor on the parent BaseLLM class. Replaced strict subclasses reading private internal state variables (_llmOptions), fully resolving strict TypeScript compilation and compiler isolation constraints.
• Parser Engine (core/llm/openaiTypeConverters.ts): Integrated multi-field validation inside streaming and non-streaming event-loop blocks to match, separate, and tag custom thinking deltas cleanly as role: "thinking".
• Testing Suite: Added robust unit testing maps validating edge-case scenarios for custom field extraction in both streaming and non-streaming contexts, alongside YAML configuration preservation tests.

How to Test

1. Set up a local gateway or mock endpoint streaming reasoning steps inside a custom JSON payload parameter (e.g., "my_custom_thinking_key").
2. Wire up your local project configuration block:

   {
     "model": "reasoning-model",
     "provider": "openai",
     "customReasoningFields": ["my_custom_thinking_key"]
   }
   

---

Summary by cubic

Adds end-to-end support for customReasoningFields to capture reasoning tokens from non-standard provider payloads and render them in the Thinking panel. Works in both streaming and non-streaming paths, with options preserved across config, schemas, serialization, and providers.

• New Features

  • customReasoningFields?: string[] in model config and ILLM options; preserved through JSON/YAML (packages/config-yaml, packages/config-types), control-plane schema, and serialization; accessible via BaseLLM.options.
  • Parsers map these fields to role: "thinking" for streaming (fromChatCompletionChunk) and non-streaming (fromChatResponse), with tests.

• Bug Fixes

  • Drop thinking-role chunks during FIM/autocomplete streaming to prevent polluted completions.

^{Written for commit 81de517. Summary will update on new commits.}

Review in cubic

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

1 issue found across 14 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[Yash-200608]

I have read the CLA Document and I hereby sign the CLA