Preserve tool-call JSON for deterministic local inference

[latent-variable]

Summary

• add arguments_json to FunctionCall so we persist the exact tool-call payload returned by the model
• have the OpenAI client reuse the stored JSON when constructing follow-up requests (falling back to a canonical dump only if the string is missing)
• capture the raw JSON when parsing responses, keeping both the parsed dict (for tool execution) and the untouched string (for replay)

Problem

When Mini-Agent is configured to send requests to a local LM Studio endpoint (or any local serving stack with KV cache), each subsequent request must be byte-identical for the cached portion of the context. Today the request builder re-serializes every tool call in the transcript using json.dumps(..., sort_keys=True). That changes key ordering, whitespace, or float formatting compared to what the model actually emitted, meaning the tool call that was prepended to Request #2 is different from the one the model saw during Request #1. LM Studio therefore treats the assistant history as a cache miss, reprocessing all prior tokens (~12k tokens per turn in our setup), wasting latency and compute.

Testing

• python - <<'PY' … (arguments_json present) ✅
• python - <<'PY' … (fallback path) ✅
• Manual: local LM Studio run now reports 99.88% prompt reuse on Request feat: Add Windows support with PowerShell integration #2 (previously ~33%).

[AkairoDev]

Thanks for this excellent PR! This is a well-designed optimization for local LLM inference with KV cache.

The solution is elegant:

• Preserves original JSON string for deterministic replay
• Falls back gracefully when arguments_json is missing
• Backward compatible

The performance improvement (33% → 99.88% cache hit rate) is impressive! 🚀

However, there are merge conflicts with the current main branch. Could you please rebase and resolve the conflicts?

Looking forward to merging this! 🙏