Epic: TUI Command Palette for agent management

[bug-ops]

Overview

The TUI is currently read-only — side panels display metrics but offer zero interactivity. Users cannot manage skills, MCP servers, memory, or view config without restarting or editing files manually.

This epic introduces a command palette (: key, vim-like) — a fuzzy-searchable overlay for executing agent management commands directly from the TUI.

Architecture

• TuiCommand enum + dedicated mpsc channel from TUI → Agent loop
• AgentEvent::CommandResult variant for responses back to TUI
• Command palette overlay widget with fuzzy filtering and category prefixes (skill:, mcp:, memory:, view:, config:)
• Commands registered declaratively for easy extension

Phases

Phase 1: Infrastructure + Read Commands

• Command channel (TUI → Agent → TUI)
• Command palette widget (overlay, fuzzy search, navigation)
• Read-only commands: skill:list, mcp:list, memory:stats, view:cost, view:tools, view:config, view:autonomy

Phase 2: Mutating Commands

• skill:reload, skill:trust <name> <level>
• mcp:connect <id>, mcp:disconnect <id>
• memory:search <query>, memory:embed-missing

Phase 3: Runtime Config

• Arc<RwLock<>> for mutable config fields
• config:summarize-output, config:autonomy, config:model
• Toast/notification widget for feedback

Design doc

.local/plan/tui-command-menu.md

Key findings from research

Subsystem	Runtime API exists	Priority
MCP servers	Full (add/remove/list)	High
Skills (reload, trust)	Partial (reload yes, trust via SQLite)	High
Memory (search, stats)	Yes (recall, counts)	Medium
Tools (list, permissions)	Read-only	Medium
Cost tracking	In MetricsSnapshot but not rendered	Low
Provider switching	No runtime API	Deferred