[Story]: AI exit interview (guided topic blocks)

[adoLime]

User Story

As a Leaver, I want a guided AI interview about my topic blocks (open tasks, decisions, stakeholders, systems, unwritten rules), so that I can leave implicit knowledge behind without any writing effort.

Context & Motivation

• Type: Functional
• Parent epic: 10 EPIC - Quality Radar & Technical Debt Management #60 (10 EPIC – Quality Radar & Technical Debt Management / Knowledge Retention & Handover)
• Teams: team:ai, team:backend, team:frontend
• Dev guide: docs/handover-dev-guide-phase1-en.md § 4 (Story C) · Concept: docs/handover-feature-konzept.md § 7
• Reusable: Buddy agent pattern (buddy_persona.py, routes/buddy.py), ScriptedLLMClient, redaction.py, chatbot SSE UI (features/chatbot/, services/sse.ts)

Acceptance Criteria

• Persona with a topic-block agenda in priority order (open tasks → decisions → stakeholders → systems/access → unwritten rules); one question per turn
• Per block: Markdown summary, confirmed/corrected by the leaver
• Session can be interrupted/resumed; memory fold for long sessions (compact pattern)
• Corpus evidence stored as citations; "I don't know" is allowed and logged
• Transcript persisted in the backend (AI stateless); PII redaction before any further use
• Multi-turn deterministically testable (ScriptedLLMClient)

Sub-Tasks (by team)

AI

• persona.py (mirror of buddy_persona.py: topic-block agenda, one question per turn, German)
• interview_agent.py (stateless multi-turn: message list + prior_summary, block progress in the request)
• Routes POST /handover/interview/turn, POST /handover/interview/compact
• ScriptedLLMClient scenarios + guardrails (stateless, projectId fail-closed, query fence, citation validation)

Backend

• HandoverInterviewSession/Message entities (BuddySession pattern) + persistence
• SSE endpoints for interview turns
• PII-redacted transcript storage; raw transcript visible to leaver + PM only

Frontend

• Interview chat UI with block progress (reuse MessageRow, MessageCitations, ChatComposer, SSE parser — same token/citation/done/error contract)

Dependencies

A → B → C → D (B/C can overlap after A, D needs C). Interview transcript is the input for the dossier (Story D).

[daniilperkin]

Heads-up from #319: the team decided to retire /chat, and sprintstart-frontend#206 deletes features/chatbot/ (including MessageRow, MessageCitations, ChatComposer) once Buddy has taken over its features.

For the interview UI, it would be safer to build on features/buddy (BuddyThread / BuddyMessage / BuddyComposer, buddyService SSE: token / citation / tool_use / done / error) than on the chatbot components. Anything the interview needs from chat (reasoning panel, citations drawer) is being moved into features/buddy as part of #206.