Per-Turn Mode Switching: Ask / Plan / Edit (ref #32)

Per-turn mode switching: Ask / Plan / Edit

Related: #32

Problem

A conversation is locked to one mode (Chat or Agent) at creation time and cannot change. Common workflow: start with Chat to ask questions → want AI to write to canvas → can't → new conversation → lose context.

Root cause: mode is a conversation-level attribute. It should be per-message.

Proposal

Same conversation and shared history, with the system prompt and tool allowlist selected per turn.

"Agent" as a mode name is avoided — it already refers to the execution entity (built-in / ACP external). Two orthogonal dimensions:

agent_id: "builtin" | "claude-code" | "acp-external"
mode:     "ask" | "plan" | "edit"

Mode should be a server-enforced capability boundary, not only a behavioral instruction in the system prompt.

Reference: Claude Code's Shift+Tab (docs) — shared context, per-message mode.

Minimum implementation

1. Mode selector next to input (Ask / Plan / Edit)
2. Each mode maps to a server-side tool allowlist (not just prompt)
3. Keyboard shortcut for quick cycle
4. mode field stored per turn in .history/ JSONL

Open question: Should Plan mode produce a structured operations artifact for canvas preview before execution?

Start by tracing the existing conversation-level mode through the input flow, server-side tool authorization, and .history/ JSONL persistence; the issue names no specific files or tests. Done means Ask, Plan, and Edit can be selected per turn, the server enforces each mode's tool boundary, a keyboard shortcut cycles them, and each turn records its mode.