More consistent agent behavior and reliable session endings

• Per-agent prompt configuration Each agent now owns its own role, objectives, engagement rules, and methodologies. Cleaner separation makes agent behavior more consistent and easier to evolve.
• More reliable session endings Recovered cases where some models occasionally failed to end sessions properly, so conversations now close cleanly across more model variants.
• Clearer session lifecycle Models now reliably distinguish between wrapping up a conversation and ending the session, so sessions end when they should.

Embed integrations now correctly fire navigation callbacks instead of silently dropping them.