Summary

The --tui v2 Ink TUI is missing ~23 of v1's 48 slash commands locally, has zero @file: / @diff / @staged / @git:N / @url: context references, no quick_commands shortcuts, no Skills Hub / Tools Config / Cron / Rollback / Snapshot / Plugins / Insights / Agents / Browser / Platforms overlays (even though backend RPCs exist for most), no MCP status on boot, no Ctrl+Z suspend, and degrades-or-breaks for any v1 command that relies on curses (e.g. /skills browse). A handful of items are also unreachable due to a registry collision on /model. Tagging @OutThisLife since you're the primary TUI v2 contributor.

This issue is a feature-parity audit — not a bug report. v2 has legitimately better UX in several places (virtual history, mouse support, Ctrl+G editor handoff, subagent accordion, sticky prompt tracker, queue-edit mode, face/verb ticker) and we don't want to lose any of that. The goal is a punch-list of what a v1 user would notice going missing when they type --tui, plus a few things v2 does that v1 users may not realize.

Methodology: audited cli.py + hermes_cli/ (v1) and ui-tui/src/ + tui_gateway/server.py (v2) end-to-end. Cross-referenced 56 backend RPCs against frontend call sites.

---

1. Slash commands — local registry gap

v1 has 48 canonical CommandDef entries in hermes_cli/commands.py. v2 registers 25 locally in ui-tui/src/app/slash/. Everything else falls through to slash.exec (captures Rich stdout as plain text) or command.dispatch (typed payloads).

Missing from v2 local registry (either degraded via slash.exec, or genuinely non-functional):

22 server RPCs are orphaned — wired on the Python side but never called by the frontend. That's your largest latent UI backlog: skills.manage, rollback.*, cron.manage, plugins.list, agents.list, tools.list/.show, toolsets.list, insights.get, browser.manage, session.title, session.history, session.save, voice.tts, process.stop, cli.exec, command.resolve, paste.collapse, config.show.

2. Curses commands break under slash.exec

tui_gateway/slash_worker.py runs a persistent HermesCLI with sys.stdout / sys.stderr redirected to a StringIO and the Rich console reconstructed over the same buffer. That works for text output. It doesn't work for:

• /skills browse and /skills inspect — use curses which needs a real TTY. Will hang or crash silently.
• Any v1 path that calls prompt_toolkit.application.run_in_terminal() — no terminal to hand off to inside the worker.
• /tools interactive picker (invoked via hermes tools externally by v1, not the slash command) — same curses constraint.

The clean fix is native Ink overlays for skills hub + tools config. The orphaned skills.manage and tools.configure RPCs already give you the data endpoints.

3. @ context references — completely missing

v1 expands these inline at send time before the message hits the model (see cli.py:7849):

• @file:path[:lines] — inline a file slice
• @folder:path — inline a recursive listing
• @diff / @staged — git diff snapshots
• @git:N — last N commits
• @url:... — fetched URL content

v2 has zero references to any of these tokens anywhere under ui-tui/src/. For Claude Code / Cursor users this is an expected baseline. Two implementation paths:

• Client-side expansion in useSubmission before prompt.submit (preserves v1 semantics, works offline).
• Server-side via a new prompt.expand_refs RPC called before prompt.submit.

4. quick_commands — missing

v1 reads config.quick_commands (user-defined exec shell shortcuts that bypass the LLM entirely) and registers them as completable slash commands. v2 has zero hits on quick_commands. Useful for users who've set up /build, /deploy, etc.

5. Keybindings

The paste-to-file behavior is a real feature difference, not just a reformat — v1 users often grep the paste files later. v2's in-memory snippet vanishes on exit.

6. Attachments

• v1 /image <path>, /paste, Alt+V, Ctrl+V, bracketed-paste-auto-image — five entry points. v2 has /image, /paste, Alt+V, bracketed-paste — Ctrl+V missing.
• v1 attaches are staged in _attached_images with persistent [📎 Image #1] badges above input until sent. v2 uses a one-shot image.attach with "attached image: X · Yk tokens" activity line — no persistent badge showing pending attachments.
• v1 Ctrl+C when idle clears staged attachments. v2 has no staging to clear.

7. /model registry collision (bug)

ui-tui/src/app/slash/commands/session.ts registers name: 'model' (picker overlay handler). ui-tui/src/app/slash/commands/setup.ts registers name: 'model' (external hermes model spawn via withInkSuspended).

Registry construction is a Map so last-registered wins. setup.ts is registered after session.ts, so the picker path in session.ts is dead code — /model always goes through the external setup wizard with Ink suspended. The ModelPicker overlay component exists and is wired but only reachable from non-slash code paths.

Probably intended: bare /model → picker overlay, /model [spec] → switch directly, /provider → setup wizard. Currently all of these route through the setup path.

8. Display / rendering

Parity on most things but a few gaps:

• Banner: v1 shows Available Tools grouped by toolset, Available Skills grouped by category, per-MCP-server connection status (transport + tool count or failed). v2 shows tools + skills but no MCP server boot status. MCP errors surface only via stderr events into the activity trail, which is easy to miss.
• Profile row (non-default profile): v1 highlights it in the banner. v2 doesn't surface it anywhere.
• Response Box: v1 wraps assistant output in a skin-styled border ╭─⚕ Hermes─────────╮ … ╰────╯. v2 renders messages inline with no box delimiter. Subjective, but several users identify Hermes by this chrome.
• Tips-of-the-day: v1 rotates through ~100 hand-written tips in hermes_cli/tips.py (slash commands, keybindings, @ refs, CLI flags). v2 has fortunes.ts — short zen-style strings, different tone, doesn't teach features.
• Markdown: v2's markdown.tsx covers CommonMark + GFM tables + task lists, but per-language syntax highlighting is missing (only diff gets colored). v1 gets full Rich syntax highlighting on non-streaming renders. Code-heavy transcripts look better in v1.
• Inline diffs: v1 renders write_file/patch/skill_manage diffs with skin-adapted colors, 6-file / 80-line caps. v2 wires inline_diff through tool.complete — parity, good.
• Reasoning box: v1 has a dedicated dim bordered box for streaming reasoning, deferred final content until it closes. v2's Thinking component covers this — parity.
• Status bar: v2 is actually richer — face/verb ticker, voice label, bg task count, ctx bar color thresholds, duration. v1 has 3 adaptive widths but less content. Keep this.

9. Modals / overlays inventory

10. Configuration surface

v1 CLI reads ~35 config sections at startup from DEFAULT_CONFIG. v2's useConfigSync.ts polls config.get {key:"full"} every 5s and reacts to:

• display.bell_on_complete, display.tui_compact, display.tui_statusbar, display.details_mode, display.thinking_mode

Things v1 honors at load that v2's config sync does not read (or reads indirectly via config.set):

• display.streaming — stream assistant tokens vs batch
• display.show_reasoning — reveal reasoning deltas
• display.show_cost — surface $ in status bar
• display.inline_diffs — render diffs after write tools
• display.tool_progress / tool_preview_length / tool_progress_overrides
• display.personality — global personality default
• approvals.timeout / clarify.timeout — modal countdown defaults
• voice.auto_tts / voice.silence_* — continuous voice mode tuning
• quick_commands (see §4)
• personalities (user-defined, not just the 14 built-ins)
• skills.external_dirs

Several of these affect rendering choices the TUI makes. display.streaming in particular is semantically important: v1 renders assistant tokens inside the response box live; v2 seems to always accumulate until message.complete — confirm?

11. Startup / boot

• v1 banner shows per-MCP connect status + profile row + plugin discovery summary. v2 shows logo + tool/skill counts + "N commits behind" if applicable. MCP status is silent unless it errors.
• v1 background update-check via prefetch_update_check. v2 surfaces the result in the SessionPanel — slightly better UX, good.
• v1 rotates 100+ tips. v2 has fortunes (shorter, zen-style). Consider importing hermes_cli/tips.py for at least the "feature discovery" subset.
• v1 resume-display: hermes -c replays history with display.resume_display=full|compact tuning. v2 has HERMES_TUI_RESUME env which resumes by id but the replay fidelity vs display.resume_display is worth verifying.
• v1 activated-skills line shown at startup when skills are preloaded via -s name1,name2. v2: not surfaced.

12. Info / introspection

• /usage: both have it, v2's panel covers model + tokens + cost + context % + compressions — parity.
• /help: v1 has categorized Rich-bordered tables per category. v2 has a Panel with category sections — functional, slightly less structured.
• /commands (paginated browser): v1 gateway-only. v2 no equivalent.

13. Error handling

• v1 surfaces rate limits via format_rate_limit_display on /usage. v2 shows tokens/cost but no rate-limit awareness in session.usage output.
• v1 distinguishes tool failure suffixes: [exit N], [error], [full] (memory). v2's tool tree marks errors but the distinction between exit-code vs exception vs context-overflow may be lost.
• v1 "double Ctrl+C within 2s" force-exit on idle. v2 has a single Ctrl+C chain that eventually exits — check timing parity.

14. What v2 does better than v1 (don't regress these)

So this doesn't read as pure v1-worship:

• Virtual-windowed history (v1 can't do this under prompt_toolkit at all).
• Mouse support — wheel, click-to-position in composer, scrollbar drag, chevron toggle.
• Ctrl+G → $EDITOR with alt-screen handoff.
• Sticky prompt tracker (↳ <previous user message> chip when scrolled back).
• Queue-edit mode (↑/↓ to navigate queued messages, Ctrl+K to dispatch head).
• Subagent accordion tree with Shift+click expand-all.
• Face + verb ticker every 2.5s in the status line.
• GoodVibes heart on "thanks" / "good bot" — small but charming.
• Live skin hot-swap via skin.changed without restart.
• Config mtime polling → auto MCP reload (v1 also has this now, but TUI's behavior is cleaner).
• Bracketed-paste tokenization [[paste:label]] for large pastes — reads better than v1's file-path placeholder inline.
• Typed command.dispatch payloads (alias / skill / exec / plugin) — cleaner than v1's subprocess stdout scraping.

15. Proposed prioritization

Tier 1 — user-visible daily friction:

1. Native Skills Hub overlay (/skills is the most-hit config surface in v1).
2. Native Tools Config overlay (/tools picker, not just enable/disable text).
3. @file: / @diff / @staged / @git:N / @url: context refs.
4. Fix /model registry collision so bare /model opens the picker.
5. MCP status surfaced on boot and on /status / banner.
6. quick_commands honored.

Tier 2 — feature gaps that break workflows: 7. /rollback overlay (checkpoint browser + diff + restore). 8. /cron manager overlay. 9. /platforms + /profile + /status + /history rendered natively. 10. Per-language syntax highlighting in markdown code blocks. 11. Ctrl+Z suspend + Ctrl+V fallback paste. 12. Paste-to-file behavior for 5+ line pastes (v1 ~/.hermes/pastes/ convention).

Tier 3 — polish: 13. Response Box chrome around assistant messages (skin-styled). 14. Import hermes_cli/tips.py rotation for boot tips. 15. /plugins, /agents, /insights, /browser, /snapshot, /debug, /fast, /gquota, /update native surfaces. 16. Resume-display fidelity parity with display.resume_display. 17. Profile row in banner. 18. Rate-limit display in /usage.

---

/cc @OutThisLife — raw materials for the audit are in cli.py, hermes_cli/commands.py (v1 side) and ui-tui/src/app/slash/registry.ts + tui_gateway/server.py @method decorators (v2 side). Happy to split this into sub-issues per overlay if that's easier to work against.