TUI ↔ Web feature parity

Comparison of the Rust terminal client (clients/tui/) against the Quasar/Vue web app (apps/web/), plus a prioritised roadmap to close the gaps. Snapshot: 2026-06-15.

Update 2026-06-15 (commit b1c1bc2): P1 partially shipped — TUI now has a security-mode picker (F7 / /security) and the /main, /rename, /summarize, /proofread commands. /edit was dropped (awkward in a TUI); the context/budget chip is deferred (needs model-context-window plumbing the TUI doesn't track yet).

Legend: ✅ present · ⚠️ partial · ❌ missing

Chat core

Feature	Web	TUI
Send + streaming (text / thinking / tool calls / usage)	✅	✅
Load history, switch / create / rename chats	✅	✅
Pinned main chat (auto-open)	✅	✅ (Ctrl+M / F6)
Mid-run follow-up queue	✅	✅
Cancel run	✅	✅ (Ctrl+X)
Regenerate last turn	✅	✅ (/retry)
/btw side question (ephemeral)	✅	✅
Edit a user message	✅	❌
Rewind / fork (branch from a message)	✅	❌
Attachments (max 6)	✅	✅
Input history (↑/↓)	⚠️	✅

Model & run controls

Feature	Web	TUI
Model picker (grouped / searchable)	✅	✅ (F2)
Reasoning effort (off/low/med/high)	✅	✅ (/think)
Built-in tools on/off	✅	✅ (Ctrl+T)
Security mode (autonomous/approve_each/judge)	✅	✅ (F7 / /security)
Collaboration mode (plan/execute/pair)	✅	❌
Data classification (confidential)	✅	❌
Memory-access policy (scoped)	✅	❌
Per-run integrations / tool selection	✅	❌ (built-ins only, global)

Slash commands

• Web: /new /rename /btw /goal /main /summarize /proofread + coding (/terminal /review /explain /fix /tests /init)
• TUI: /new /btw /retry /model /think /tools /attach /agents /inbox /logout /help
• Custom commands (from GET /commands): ✅ both
• Missing in TUI: /goal, coding prompts (/main /rename /summarize /proofread now present)

Coding mode

Feature	Web	TUI
Shell in workspace	✅ (terminal panel)	✅ (!cmd via PTY WS)
Monaco editor + file tree	✅	❌
Workspace snapshot / revert	✅	❌
Diff rendering for file tools	✅ (cards)	✅ (unified diff in stream)

Agents & HITL

Feature	Web	TUI
Sub-agent panel (live status / usage)	✅	✅ (F5)
Sub-agent transcript	✅	✅
Tool approval (allow / remember / reject)	✅	✅
ask_user / agent-question cards (multi, custom)	✅	✅
Background-resume auto-attach	✅	✅

Inbox

Feature	Web	TUI
List + detail + reply + suggest	✅	✅ (F3)
Status filter, resolve / dismiss	✅	✅
Star / snooze / bulk / reply-all / forward	✅	❌
Live refresh	✅	✅

Other pages (web) — all ❌ in TUI

Notes, Files, Knowledge/Entities, Contacts, Calendar, Agenda/Commitments, Logbook, Scenes, Skills, Tasks, Dashboards, Workflows. TUI surfaces only Chat + Sessions + Inbox + Agents.

Cross-cutting

Feature	Web	TUI
Auth: Keycloak OIDC	✅ (Auth-Code+PKCE)	✅ (device grant)
Realtime: SSE stream + control WS	✅	✅
Adopt accent colour (ui.accent)	✅ (source)	✅ (reads it)
Settings UI (profile/appearance/behavior/devices/MCP tokens/usage/budget)	✅	❌ (CLI flags / config only)
Voice (STT / TTS)	✅	❌
Context / budget gauges	✅	❌
Org switching	✅	❌ (decoded, not surfaced)

Summary

The TUI has a solid chat foundation at genuine parity: streaming, model/reasoning, follow-ups, /btw, sub-agents, HITL (approvals + question cards), inbox basics, coding shell with diffs; shares auth/realtime/accent. The gaps are run-governance, voice, the coding editor, message lifecycle (edit/rewind/fork), inbox extras, the standalone pages, /goal, and a settings UI.

---

Roadmap (prioritised)

Ordered by value-to-effort for a terminal user. Each item notes the backend it already has (most are pure TUI work — the API exists).

P1 — cheap, high-impact (run controls + lifecycle)

1. Security mode picker — DONE (b1c1bc2): F7 / /security, passed as security_mode, shown as a 🔒 composer chip.
2. Edit a user message — DROPPED: awkward in a TUI (no in-place message editing UX); /retry covers regeneration.
3. Missing global slash commands — DONE (b1c1bc2): /main [msg], /rename <t>, /summarize, /proofread.
4. Context / budget chip — DEFERRED: needs the model context window + cumulative history tokens, which the TUI doesn't track (the chat_run WS frame carries no context tokens). Plumb context_tokens/threshold_tokens onto the run/WS event first, then add the chip.

P2 — moderate (governance + tools)

1. Per-run integrations / tool selection — let the user toggle integrations/devices for a run (integration_entries/devices in the run body). Needs a picker popup like the model one.
2. Memory-access + collaboration + classification pickers — same pattern as security mode; group them under one "run settings" popup to avoid F-key sprawl.
3. /goal autonomous-goal mode — start a goal loop in a session chat (chat.startGoal equivalent endpoint); render goal turns. Higher value for unattended terminal use.

P3 — larger surfaces

1. Inbox extras — star/snooze/bulk/reply-all/forward (endpoints exist from the upstream inbox work). Incremental key bindings on the existing inbox view.
2. Notes page — list/create/edit notes (/notes API). Small, self-contained, useful in a terminal.
3. Settings view — at least appearance (accent/language) + MCP token mint, read/write via /me/preferences and /me/api-tokens.

P4 — heavy / maybe out of scope for a TUI

1. Coding editor — a Monaco-equivalent in a terminal is a big lift; !cmd + diffs may be enough. Consider a read-only file viewer + $EDITOR handoff instead of a full editor.
2. Voice (STT/TTS) — needs audio capture/playback; low priority for a terminal client.
3. Workspace snapshot/revert, Dashboards, Knowledge/Contacts/Calendar/Workflows — port only if a concrete terminal use-case emerges.

Suggested first slice: P1 (1–4) — small, all backend-ready, and they remove the most friction for a terminal-first user.