Persistent Terminal MCP has visible tool definitions with reasonable schemas and descriptions. However, the evaluation reveals significant gaps in definition quality: (1) Inconsistent description depth, some tools like create_terminal have rich parameter docs while others like open_terminal_ui lack meaningful context. (2) Parameters lack validation constraints, no enum definitions for the 'mode' parameter in read_terminal despite listing valid options ('full', 'head', 'tail', 'head-tail'); LLMs will guess. (3) Missing output schema documentation, no description of what each tool returns; agents cannot plan downstream steps. (4) Error handling guidance absent, tools describe what they do but not how to recover from failures. (5) Security concern: no visible input sanitization for shell execution (create_terminal accepts arbitrary 'shell' and 'env' parameters). (6) Composition issue: two near-identical tools (create_terminal and create_terminal_basic) introduce cognitive load for LLM selection without clear differentiation. The schema for read_terminal has typed parameters but the descriptions do not explain the semantic difference between 'head', 'tail', and 'head-tail' modes clearly enough to guide tool use. Overall, definitions are functional but below production baseline, they lack the LLM-optimized descriptions and error recovery guidance expected of A-grade tools.

• Add formal JSON Schema output definitions to every tool. Example: read_terminal should document 'returns { output: string, lineCount: number, timestamp: number }'. Agents need this to plan multi-step calls.
• Convert 'mode' parameter in read_terminal to an enum constraint in schema: enum: ['full', 'head', 'tail', 'head-tail']. Remove the enum list from description; let the schema enforce it.
• Consolidate create_terminal and create_terminal_basic into a single tool with optional parameters. Document: 'Omit shell/env/cols/rows for defaults.' This simplifies LLM selection.
• Add validation constraints to numeric parameters: 'cols: 10-500 (terminal width)', 'rows: 3-200 (terminal height)', 'timeout: 100-300000 (milliseconds)', 'port: 1024-65535'. Include these in parameter descriptions.
• Document error recovery for all tools. E.g., write_terminal: 'If terminalId invalid, call list_terminals() to confirm ID. If terminal is dead, create a new one with create_terminal().'
• Add confirmation pattern to kill_terminal: 'This is destructive. Return signal=SIGTERM and require agent to confirm with send=true before executing SIGKILL.'
• Add input sanitization: validate shell parameter is an absolute path (regex: ^/[a-zA-Z0-9/_.-]+$); validate env keys are alphanumeric+underscore; reject shell wildcards and pipes.
• Implement pagination for list_terminals. Add parameters 'limit' (1-100, default 20) and 'offset' (default 0). Return { terminals: [...], total: N, hasMore: boolean }.
• Clarify open_terminal_ui purpose in description: 'Opens a web-based terminal interface. Use this for interactive exploration of sessions; programmatic access use other tools.' Document port range and autoOpen side effects.

Inferred effective spec: 2025-06-18+.

• Relies on Sampling (deprecated) - integrate directly with the LLM provider API