fix(kiro): map GPT reasoning effort fields

GPT-5.6 via Kiro needs reasoning.effort while Claude uses output_config.effort.
Resolve the effort path per-model schema (like Kiro CLI/KAS) so GPT-5.6
receives the correct structured thinking level. Claude path unchanged.

- Add resolveKiroEffortPath returning "reasoning" | "output_config" | null
- buildKiroAdditionalModelRequestFields emits schema-specific shape
- Keep prompt tags for backward compatibility
- Add OpenAI/Claude translator coverage for GPT-5.6 effort mapping

open-sse/translator/request/openai-to-kiro.js

@@ -511,12 +511,10 @@ function convertMessages(messages, tools, model) {
  *    Kiro's 2-3 minute server timeout. The suffix is stripped before being
  *    sent upstream.
  *
- * 2. Thinking / reasoning. Kiro does not accept `thinking.type` or
- *    `reasoning_effort` natively. The only way to enable reasoning is to
- *    inject `<thinking_mode>enabled</thinking_mode>` into the user content
- *    sent upstream. Detection covers Anthropic-Beta header, Claude API
+ * 2. Thinking / reasoning. Detection covers Anthropic-Beta header, Claude API
  *    `thinking`, OpenAI `reasoning_effort`, AMP/Cursor magic tags, and model
- *    name hints.
+ *    name hints. Kiro's prompt tags remain for compatibility, while supported
+ *    models also receive the same schema-specific effort fields as Kiro CLI.
  */
 export function openaiToKiroRequest(model, body, stream, credentials) {
   const messages = body.messages || [];