fix: normalize LLM output types in profile serialization

[ygh1254]

Summary

• Add type coercion helpers (_coerce_to_str, _coerce_to_str_list) to safely convert dict/list LLM outputs into plain strings
• Normalize fields at construction time via OasisAgentProfile.__post_init__ — covers bio, persona, country, profession, gender, interested_topics
• Defensive coercion in serializers (_save_twitter_csv, _save_reddit_json) as a second safety net before string operations like [:150] and .replace()
• Normalize in _generate_profile_with_llm() immediately after json.loads() to catch structured outputs before they reach the dataclass

Root Cause

LLM JSON parsing (json.loads) can produce dict or list values for fields declared as str (e.g. bio, persona, country). The serialization code then crashes on string operations like slicing ([:150]) or .replace() applied to non-string types.

Traceback from issue:

File "backend/app/services/oasis_profile_generator.py", line 1296, in _save_reddit_json
  "bio": profile.bio[:150] if profile.bio else f"{profile.name}",
KeyError: slice(None, 150, None)

Testing

The coercion logic handles:

• bio as dict → extracts text via common keys or json.dumps
• persona as dict → same
• country as list → joins into comma-separated string
• interested_topics as nested dicts → flattens to list[str]

Closes #154

[666ghj]

Thank you for tracing the serialization crash to structured LLM values. We preserved your original commit and merged the maintained version in #737.

The follow-up normalizes once when OasisAgentProfile is constructed, recursively flattens nested topics, handles None without exposing the literal text, includes MBTI, and removes competing coercion policies from the serializers. Four profile and serialization tests passed. This PR is now closed as superseded by the merged implementation.

🤖 Sent by the MiroFish repository maintenance agent.