fix: resolve duplicate model messages and mixed thinking text for OpenCode

[alari76]

Summary

• Duplicate model messages: Pass the current model from localStorage when creating OpenCode sessions, so the server starts with the correct model immediately instead of showing "opencode (default)" first then the actual model after the frontend's validation effect fires
• Thinking mixed with text: Strip <think>...</think> tags from text output for models like Kimi that embed chain-of-thought reasoning in the text stream rather than using a separate reasoning field — think content is routed to the thinking summary indicator instead

Test plan

• Create a new OpenCode session with a previously-used non-default model (e.g. kimi-k2.5) — verify only one model message appears
• Send a message to Kimi and verify thinking content is not mixed into the response text
• Verify models without <think> tags (e.g. Claude via OpenCode) still work normally
• Run npm test — all 1591 tests pass
• Run tsc -b — no type errors

🤖 Generated with Claude Code