# SDD Explore: Multi-Session Terminal UX ## Executive Summary The codebase has a well-built persistent terminal foundation from the `persistent-terminal-sessions` change. `TerminalManager` currently tracks exactly one `TerminalSession` per `instance_id` in an in-memory dict. `TerminalSession` already supports WebSocket attach/detach, circular output buffer replay, idle timeout, and process lifecycle management. Implementing multi-session terminal support is a **moderate-complexity, medium-risk** change. The core backend refactor is straightforward: change the session tracking key from `instance_id` to `(instance_id, session_id)` and update the WebSocket endpoint to accept a `session_id`. The frontend work is more involved: designing a tabbed session UI that works on both desktop and mobile, handling session creation/switching/closing, and integrating with the existing `MobileTerminalWrapper`. No database schema change is **strictly required** for an MVP—sessions can remain purely in-memory with the same idle-timeout cleanup. However, adding a `terminal_sessions` table would provide cross-API-restart persistence, session auditability, and a foundation for future features like session history or named sessions. ## Current Architecture (as explored) ### Backend - **`TerminalManager`** (`apps/api/src/services/terminal_manager.py`): - `self._sessions: dict[str, TerminalSession]` keyed by `instance_id` string. - `get_or_create_session(instance_id, container_id, startup_command)` — returns the single existing session or creates a new one. - `attach_websocket(session, websocket)` — detaches any *existing* WebSocket connections on that session (closes them with code 4000) before attaching the new one. This enforces single-active-client per session. - `reset_session(instance_id, container_id, ...)` — kills the existing session and creates a new one. - Idle check loop every 60s; sessions with no WebSockets attached for 30 minutes are cleaned up. - **`TerminalSession`** (`apps/api/src/services/terminal_session.py`): - Already has a `session_id: str` field (UUID) but it is not used as a lookup key. - Manages one `docker exec` PTY process per session. - Circular buffer (10KB) for output replay. - Tracks `self._websockets: set[Any]` for attached connections. - **`api/terminal.py`** (`apps/api/src/api/terminal.py`): - WebSocket endpoint: `/ws/tool-instances/{instance_id}/terminal` - Authenticates user, verifies instance ownership/running state, then calls `terminal_manager.get_or_create_session()`. - Supports JSON control messages: `resize`, `reset`. - POST endpoint: `/projects/{project_id}/repositories/{repo_id}/instances/{instance_id}/terminal/reset` — resets the single session. - **Database**: - No `terminal_sessions` table exists. Terminal sessions are purely in-memory. - `ToolInstance` model (`apps/api/src/models/tool_instance.py`) has no terminal-related fields. ### Frontend - **`TerminalComponent`** (`apps/web/src/components/terminal.tsx`): - Single xterm.js terminal per component. - One WebSocket connection to `/ws/tool-instances/{instance_id}/terminal`. - Handles reconnect with exponential backoff (max 3 attempts). - Font size persisted globally in `localStorage` under key `terminal-font-size`. - Copy/paste buttons on mobile only. - Status indicator: connecting, connected, disconnected, error, resetting. - **`TerminalPage`** (`apps/web/src/pages/terminal.tsx`): - Desktop: renders one `TerminalComponent` inside a page shell. - Mobile: renders `MobileTerminalWrapper` which composes `MobileTerminalHeader`, `TerminalComponent`, `SpecialKeysStrip`, and `SpecialKeysPanel`. - **`MobileTerminalWrapper`** (`apps/web/src/components/mobile-terminal-wrapper.tsx`): - Already handles auto-hide header, virtual keyboard height, special keys, and mobile viewport detection. - Manages terminal ref callbacks (`sendData`, `connectionStatus`, `focusInput`, `changeFontSize`). ### Prior Art - **`persistent-terminal-sessions`** (fully implemented): - Sessions survive WebSocket disconnections. - Buffer replay on reconnect. - Idle timeout cleanup. - Reset functionality. - **`mobile-terminal-ux`** (mostly implemented): - Mobile fullscreen terminal with collapsible chrome. - Special keys toolbar. - Dynamic viewport handling for virtual keyboard. ## Architecture Options for Multi-Session ### Option A: In-Memory Multi-Session (MVP) - Change `TerminalManager._sessions` to `dict[tuple[str, str], TerminalSession]` keyed by `(instance_id, session_id)`. - Add `create_session(instance_id, container_id, ...)` that always creates a new session. - Keep `get_or_create_session()` for backward compatibility (returns the "default" or only session). - Add `get_sessions_for_instance(instance_id) -> list[TerminalSession]`. - Add `close_session(instance_id, session_id)` to kill a specific session. - **Tradeoffs**: Simplest, no DB migration, survives existing patterns. Loses sessions on API restart. ### Option B: Database-Backed Session Metadata - Create `terminal_sessions` table: ```sql id UUID PRIMARY KEY, instance_id UUID FK(tool_instances.id, ondelete=CASCADE), session_name VARCHAR(255), status VARCHAR(50), -- active, idle, closed created_at TIMESTAMPTZ, last_activity_at TIMESTAMPTZ, closed_at TIMESTAMPTZ ``` - `TerminalManager` still keeps `TerminalSession` objects in memory, but creates/updates DB rows on lifecycle events. - **Tradeoffs**: Enables cross-restart persistence, session history, named sessions, and auditability. Adds migration and async DB overhead to hot paths. ### Option C: Hybrid (Recommended) - In-memory active sessions for performance. - DB table for metadata, created on session start, updated on activity/close. - On API restart, sessions are gone (no process resurrection), but metadata remains for history. - **Tradeoffs**: Best of both worlds. Slightly more complex than Option A but much simpler than full persistence. ### Decision Matrix | Criterion | Option A | Option B | Option C | |-----------|----------|----------|----------| | Implementation complexity | Low | Medium | Medium | | DB migration required | No | Yes | Yes | | Cross-restart persistence | No | Yes (full) | Metadata only | | Resource auditability | No | Yes | Yes | | Performance | Best | Good (cacheable) | Best | | Recommended for MVP | **Yes** | No | **Preferred** | ## WebSocket Protocol Options ### Option 1: URL Path Segment (Recommended) ``` /ws/tool-instances/{instance_id}/terminal/{session_id} ``` - Clean, RESTful, easy to route in FastAPI. - Default session can use a reserved ID like `default` or keep `/terminal` as an alias. - **Tradeoff**: Breaks existing hardcoded URLs; needs backward-compatibility route. ### Option 2: Query Parameter ``` /ws/tool-instances/{instance_id}/terminal?session_id=... ``` - Easier to add without changing route structure. - Less idiomatic for WebSocket APIs. - **Tradeoff**: Query params in WebSocket URLs can be inconsistently supported by proxies. ### Option 3: First-Message JSON Payload - Client connects to `/terminal`, then sends `{"type": "attach", "session_id": "..."}`. - Server must hold the connection in limbo until the attach message arrives. - **Tradeoff**: More complex state machine; harder to reject invalid sessions early. **Recommendation**: Option 1 with a backward-compatible fallback: - `/ws/tool-instances/{instance_id}/terminal` → attaches to the "default" session (existing behavior). - `/ws/tool-instances/{instance_id}/terminal/{session_id}` → attaches to the specified session. ## Frontend UX Design Options ### Session Presentation: Tabs vs Panes | Feature | Tabs | Panes (Split) | |---------|------|---------------| | Desktop UX | Good | Excellent (tmux-like) | | Mobile UX | Good | Poor (too cramped) | | Implementation | Medium | High | | Accessibility | Good | Complex | | Recommendation | **Preferred** | Future enhancement | **Decision**: Start with tabs. A split-pane layout can be added later as an advanced feature without breaking the tab model. ### Tab Bar Design - Position: Above the terminal container on desktop; integrated into `MobileTerminalHeader` on mobile. - Contents: - Session name (auto-named "Session 1", "Session 2", or custom). - Status dot (connecting, connected, error). - Close button (×) on hover/active. - New tab button (+). - Overflow: Horizontal scroll on mobile; wrap or scroll on desktop. ### Fullscreen Mode - **Behavior**: Toggle hides all page chrome (header, sidebar, tab bar can optionally be shown as a minimal overlay). - **Trigger**: `Ctrl+Shift+F` or UI button. - **Mobile**: Should integrate with existing mobile fullscreen behavior (already hides AppShell). Fullscreen on mobile could mean hiding the special-keys strip too, with a gesture to reveal. - **Exit**: `Esc` or UI button. ### Keyboard Shortcuts | Shortcut | Action | Notes | |----------|--------|-------| | `Ctrl+Shift+N` | New session | May conflict with browser "New window" on some platforms. Consider `Ctrl+Shift+T` if not used for "Reopen tab". | | `Ctrl+Shift+W` | Close current session | Conflicts with browser "Close window". May need `Ctrl+Shift+D` or accept override with `preventDefault()`. | | `Ctrl+Shift+F` | Toggle fullscreen | Safe, no major browser conflict. | | `Ctrl+Shift+T` | Toggle tab bar visibility | Conflicts with "Reopen closed tab" in browsers. Consider `Ctrl+Shift+B` or `Ctrl+Shift+~`. | **Recommendation**: Use `preventDefault()` aggressively and show a shortcuts help modal (e.g., `Ctrl+Shift+/` or `?`). ### Session Naming - **Auto-name**: "Session 1", "Session 2", etc. based on creation order. - **Custom name**: Editable by double-clicking the tab. Persisted in DB if Option B/C, or in-memory only for Option A. - **Default session**: The first session created for an instance can be unnamed or named "Default". ### Reset/Kill Semantics Current behavior: "Reset Terminal" kills the single session and starts fresh. With multi-session: - **Close Session** (× on tab): Kills the `docker exec` process and removes the session. - **New Session** (+ on tab bar): Creates a new session and switches to it. - **Reset Session** (in menu): Same as current reset but scoped to the active session. - **Reset All** (optional, in menu): Kill all sessions for the instance and recreate a default one. ### Font Size Persistence - Currently global (`localStorage` key `terminal-font-size`). - With multi-session, users may want different font sizes per session (e.g., larger for presentations, smaller for logs). - **Options**: 1. Keep global (simplest, no change). 2. Per-session font size (stored in session state or DB). 3. Per-instance font size. - **Recommendation**: Keep global for MVP. Per-session font size is a nice-to-have that adds complexity. ### Status Per Session - Each tab shows a status dot. - Possible statuses: `connecting` (pulsing), `connected` (green), `disconnected` (yellow), `error` (red), `closed` (gray). - The terminal component already tracks these statuses; they just need to be surfaced at the tab level. ## Database Schema Recommendation (Option C) ```python class TerminalSessionModel(UUIDPrimaryKeyMixin, TimestampMixin, Base): __tablename__ = "terminal_sessions" instance_id: Mapped[uuid.UUID] = mapped_column( UUID(), ForeignKey("tool_instances.id", ondelete="CASCADE"), nullable=False ) name: Mapped[str | None] = mapped_column(String(255), nullable=True) status: Mapped[str] = mapped_column( String(50), nullable=False, default="active" ) # Not storing process PID here — that's runtime-only in TerminalManager created_at: Mapped[datetime] = mapped_column( DateTime(timezone=True), nullable=False, default=datetime.utcnow ) last_activity_at: Mapped[datetime | None] = mapped_column( DateTime(timezone=True), nullable=True ) closed_at: Mapped[datetime | None] = mapped_column( DateTime(timezone=True), nullable=True ) ``` **Migration**: New alembic revision adding `terminal_sessions` table. ## Open Questions Needing User/Product Decisions 1. **Max sessions per instance?** Suggest 5 for MVP to prevent resource exhaustion. 2. **Should we persist sessions across API restarts?** Option A = no; Option C = metadata only. Product call. 3. **Tab vs Pane UI?** Strongly recommend tabs for MVP. Panes as future work. 4. **Keyboard shortcuts — override browser defaults?** `Ctrl+Shift+W` closes browser window. We can `preventDefault()` but should warn users. 5. **Should the existing `/terminal` endpoint remain as a default-session alias?** Yes for backward compatibility, but confirm. 6. **Session idle timeout per session or global per instance?** Currently per session. Keep per session. 7. **Should font size be global, per-instance, or per-session?** Recommend global for MVP. 8. **Copy/paste on desktop — any gaps?** Current desktop relies on native xterm.js copy/paste (`Ctrl+C`/`Ctrl+V` with selection). This is standard and sufficient. Mobile already has buttons. ## Risks and Feasibility Assessment | Risk | Likelihood | Impact | Mitigation | |------|------------|--------|------------| | Resource exhaustion from too many docker exec processes | Medium | High | Enforce max sessions per instance (5). Idle timeout already exists. | | Mobile UX degradation from tab bar clutter | Medium | Medium | Integrate tabs into existing `MobileTerminalHeader` auto-hide. Limit visible tabs, overflow scroll. | | Backward compat breakage from URL change | Low | Medium | Keep `/terminal` as default-session alias. | | Concurrent WebSocket policy bugs | Medium | High | Ensure "close existing" only applies within same `(instance_id, session_id)`, not across sessions. | | Scope creep (panes, detachable windows) | High | Medium | Explicitly exclude split panes and detachable windows from MVP. | ## Feasibility: Green/Yellow/Red **Yellow-Green**. The backend changes are well-scoped and build on solid existing infrastructure. The frontend tab UI is the largest unknown, especially mobile integration, but the existing `MobileTerminalWrapper` provides a good foundation. No external dependencies needed. ## Recommended Next Step **Proceed to `design` phase** after resolving these scoping decisions: 1. Choose Option A or C for session storage (recommend Option C). 2. Confirm max sessions limit (recommend 5). 3. Confirm tab-only UI for MVP (no panes). 4. Confirm backward-compatible WebSocket URL strategy. Then write `design.md` with concrete decisions and `tasks.md` with implementation steps.