v0.5: Separate persona-directed vs goal-oriented dialogue modes

[MakiDevelop]

Background

Per arXiv:2512.12775 (EACL 2026), persona fidelity degrades systematically faster when an LLM must simultaneously:

• maintain the persona (roleplay the interviewee), and
• follow task instructions (do the work)

vs. pure persona-directed dialogue (just be the interviewee).

Proposal

Architecturally separate two response modes:

1. persona_mode — pure roleplay. Bot responds in the interviewee's voice as if having a conversation. No task constraints.
2. task_mode — goal-oriented. Bot drafts an email / replies to a candidate / generates a LinkedIn post. The persona acts as a style filter applied to a task-completion response, not as the primary directive.

task_mode would use stronger persona re-injection (every 5 turns, not every 20) and prefer PPA Stage 3 with more aggressive refinement.

API sketch

async def respond(
    ...,
    mode: Literal["persona", "task"] = "persona",
    task_directive: str | None = None,  # required when mode="task"
) -> str:
    ...

Acceptance criteria

• mode parameter wired through process_turn()
• task_mode triggers tighter re-injection
• Documentation in spec on when to use which mode
• Blind test protocol updated to test both modes separately