self-improve

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). In scripts/self_improve.py, the dream workflow reads user-message free text from Codex local session transcripts referenced by threads.rollout_path in ~/.codex/state_5.sqlite and JSONL files under ~/.codex/sessions / ~/.codex/archived_sessions, then ingests those messages at runtime via collect_user_messages() → iter_rollout_events() → JSON parsing of payload["message"].