Capture per-turn skill execution events in session history and transcript output while preserving human-readable agent replies and shielding tool logs from model context/routing.

~

Summary

• Adds structured per-turn skill execution events to session history (role: "skill") while preserving normal agent replies for users.
• Keeps tool event logs out of LLM context and dynamic expertise routing to avoid behavioral drift.
• Persists skill events in saved transcripts (markdown and csv) for auditability.
• Includes focused tests for skill event logging, context isolation, transcript persistence, and routing safety.

Changes

• rooms/agent.py
  • Added generate_response_with_events(...) for internal structured output.
  • Kept generate_response(...) backward-compatible (still returns str).
  • Added consume_last_skill_events() for session-level ingestion.
• rooms/session.py
  • Stores skill execution records in session.history.
  • Excludes role == "skill" records from model context.
  • Ensures dynamic selection scoring ignores skill-log entries.
• rooms/storage.py
  • Transcript writer now serializes skill events in both markdown and csv outputs.
• tests/test_session.py
  • Added tests for:
    • structured skill logging + human-readable turn output
    • exclusion of skill logs from model context
    • transcript persistence of skill events

Why

Issue #56 requires structured skill execution visibility without degrading the user-facing conversational UX.
This PR introduces an internal event trail for tooling/audit while preserving natural language replies from agents.

Validation

.\venv\Scripts\python.exe -m pytest tests/test_session.py tests/test_skills_runtime.py

• Result: 20 passed

Linked Issue

Closes #56