Act IX: Armor for Production Servers — Universal Command Guard & v0.8.6 Release

Entry 47. Armor for Production Servers: Universal Command Guard, Taming Bash History, and Xterm Buffer X-Ray

1. The Trigger: When a Protection Toggle Clicks into the Void

During real-world dogfooding of Shellit, I opened the edit modal for one of our working servers assigned to the default environment (defaultEnv), scrolled down to the settings, and enabled the toggle: “Production Command Guard”.

The intent seemed crystal clear: I explicitly instructed the client to guard this host against destructive mistakes. I saved the host, initiated the SSH session, the terminal connected… and nothing changed. No alerting border, no warning banner, no command interception. I typed a destructive command — and it was dispatched directly to the remote server without resistance.

The immediate feedback to the AI was unambiguous:

“audit the prod guard implementation"
"I see zero visual difference, in addition to the bugs you discovered"
"record all discovered bugs! then we will discuss the plan"
"we need to explain this in the toggle subtitle. Also, explain to me what selecting items from this environment dropdown even affects?”

2. The Blind Spots: A Disconnected Switch and a Cascade of Vulnerabilities

Digging beneath the surface revealed that this was far more than an isolated visual oversight. It was a chain of 6 interconnected architectural and ergonomic vulnerabilities:

1. Disconnected Environment Logic (BUG-037): The logic inside ProdGuardBorder and the keystroke interceptor in terminal_screen.dart strictly checked host.environment == HostEnvironment.production. For hosts configured as defaultEnv, staging, or development, the protection switch was completely dead code — saved to the encrypted database, but completely ignored by the active terminal screen.
2. Fatal History Bypass (BUG-038, P0 Blocker): The terminal’s character accumulator _commandLineBuffer gathered printable keystrokes while discarding ANSI escape sequences. When a user pressed Up Arrow to cycle back through Bash history to a previous command such as rm -rf /var/cache/*, _commandLineBuffer remained completely empty! Pressing Enter dispatched the dangerous command straight to the SSH channel without triggering DangerousCommandChecker or the confirmation modal.
3. Regex Holes & Utility Blind Spots (BUG-039): The pattern matcher resided in the presentation layer terminal_ui instead of the domain core and tested for \brm\s+. Commands using absolute paths (/bin/rm), reordered flags (rm -f -r), or critical destructive tools (wipefs, truncate -s 0, systemctl poweroff/reboot, chmod -R 777) slipped right through.
4. Disruptions Inside nano and vim (BUG-040): In full-screen TUI editors, pressing Enter continued to be parsed as shell commands. If a config line contained a matching substring, the confirmation dialog popped up, and pressing Cancel sent a Ctrl+C (\x03) interrupt that broke the editor session.
5. Divergent MCP AI Gateway Logic (BUG-041): The MCP server gateway for external AI agents used an ad-hoc substring check lower.contains('rm -rf'), completely out of sync with terminal safety checks.
6. Zero Hardcoded Strings Violation & Lack of Tab Context (BUG-042): Warning banners were hardcoded in English, and open terminal tabs lacked visual indicators of their target environment or active guard status.

3. Architectural Solutions & Implementation

Resolving these issues demanded a unified overhaul across multiple layers:

1. Universal Command Guard & Color Distinction

We expanded “Prod Guard” into a comprehensive Universal Command Guard:

• On servers declared as Production (HostEnvironment.production), destructive command protection is always active by default, framed with a vibrant red safety border (2.5px, #FF3333) and a top banner [⚠️ PRODUCTION ENVIRONMENT].
• On servers in any other environment (defaultEnv, staging, development) where the user explicitly turned on the Command Guard toggle, the terminal renders a distinctive amber warning border (2.0px, #FFB020) and top banner [🛡️ COMMAND GUARD ACTIVE].
• In HostFormDialog, the toggle is renamed to “Command Guard (Destructive Protection)” and features a dynamic subtitle explaining that Production hosts are always guarded automatically, while non-production hosts can be guarded on demand.

2. Xterm Screen Buffer X-Ray & Shell Prompt Stripping

Rather than attempting complex ANSI cursor position tracking inside _commandLineBuffer, upon pressing Enter the terminal takes a direct snapshot of the active screen line:

final screenLine = _terminal.buffer.lines[_terminal.buffer.cursorY].toString();
final candidate = DangerousCommandChecker.cleanPromptAndExtractCommand(screenLine);

The new static helper cleanPromptAndExtractCommand intelligently strips arbitrary shell prompts (user@host:~$ , root@srv:~# , (venv) [user@host /var]$ ), isolating the candidate command. If the candidate contains destructive actions, input is intercepted and the confirmation modal is triggered.

3. Respecting Alternate Screen Buffers (TUI Apps)

Full-screen terminal applications (vim, nano, htop, less) transition xterm into an alternate screen buffer (\x1b[?1049h). Using xterm.dart’s native property _terminal.isUsingAltBuffer, we placed an early exit at the head of the input handler:

if (_terminal.isUsingAltBuffer) return;

Inside TUI editors, the command interceptor steps aside, allowing frictionless line breaks and text editing.

4. Domain Core Migration & MCP Gateway Alignment

DangerousCommandChecker was relocated to packages/core_foundation/lib/src/security/dangerous_command_checker.dart. Patterns now support binary paths (?:[^\s]*\/)?, arbitrary flag order, and a comprehensive roster of destructive system commands (wipefs, truncate, dd, mkfs, systemctl, init 0, chmod -R). The identical validator is wired into the MCP Server Gateway.

5. Tab Badges & Complete Localization

Terminal tab headers in top_bar_tabs.dart now display color-coded environment chips (PROD red, STAGE amber, DEV blue) alongside an active [🛡️ GUARD] badge. All UI strings are extracted into context.tr('prod_guard.*') with full Russian and English translations.

4. The Outcome: Release v0.8.6

• All 6 issues (BUG-037 through BUG-042) are marked as RESOLVED.
• 48 domain tests and 9 regression UI tests established.
• 100% green test pass across all modules:
  • core_foundation: 27/27 passed.
  • terminal_ui: 126/126 passed.
  • desktop_plugin_sdk: 62/62 passed.
  • apps/shellit: 61/61 passed.
• Static analysis with flutter analyze completed with 0 issues found.
• Version officially bumped to v0.8.6.

Server protection in Shellit is now intuitive, visually unmistakable, resilient against history bypasses, and unobtrusive during daily configuration workflows.