Interactive mode: make it behave like a real terminal

[ANonABento]

Interactive mode already renders the real claude/codex TUI via raw-PTY xterm passthrough (arrows/Ctrl/Tab/Esc/resize already work). This closes the gaps that made it feel un-terminal-like. All green: clippy + cargo test --workspace (497+19), vitest (411), tsc, lint.

A — Single input surface

Dropped the line-injecting chat box from the interactive panel. The terminal is now the sole input, so Tab-complete, slash menus, @file mentions, y/n prompts, and multi-line paste work natively — the chat box submitted whole lines + a forced Enter, which broke all of those. A thin hint points the user at the terminal.

B — Terminal fidelity

• tmux set -g mouse on + history-limit 50000 at server start → scroll/click inside TUIs and deep scrollback (was ~2000 lines).
• Debounced the terminal-view resize cascade so a drag-resize issues one fit instead of a burst of resize_pty calls that flicker the TUI.

C — Completion is advisory, not authoritative

An interactive session is human-in-the-loop, so the watcher no longer:

• auto-advances the pipeline the instant the agent prints its done-sentinel, or
• marks a session FAILED at the 2h timeout.

Instead it keeps the TUI alive, drops the task to idle, and emits agent:<task>:interactive_done. The panel shows an "Advance column" affordance; you advance when satisfied via the new agent_advance command (same mark_complete the headless path runs). The /exit-on-success teardown is removed so the session persists for continued chat.

Notes / follow-ups

• The advisory signal is currently event-based (surfaces while the panel is attached). A persisted flag (so a closed-panel task shows a "ready to advance" badge on the card) is a sensible follow-up.
• Not in this PR (from the same audit): adaptive readiness (no hard-kill on slow cold start) and codex --append-system-prompt/resume parity.

🤖 Generated with Claude Code