Agent Tools

term2 exposes a controlled set of tools to the agent, gated by the active profile and sandbox policies.

Filesystem Tools

• read_file: Reads file contents with support for line-number ranges (start_line and end_line, inclusive and 1-based) and size bounds.
• apply_patch: Applies unified patches to files with preview and self-healing syntax recovery.
• search_replace: Performs precise contiguous block replacements within a file.
• create_file: Creates a new file in the workspace (or overwrites an existing file if explicitly approved).
• glob: Finds files and directories using glob patterns (leveraging fd when installed).
• grep: Performs regex or literal text searches across files (leveraging ripgrep when installed).
• read_code_outline: Generates high-level symbol outlines (classes, functions, interfaces) for rapid file comprehension.
• code_context_search: Searches code symbols and definitions across the codebase.

Shell Execution

• shell: Executes shell commands within the active sandbox policy. Commands that read or write outside approved boundaries prompt for explicit confirmation.
• Background Shell Jobs: Long-running commands can run detached in the background:
  • get_shell_job: Check running state, exit codes, and output byte counts.
  • monitor_shell_job: Attach an output watch that delivers matching lines as notifications.
  • cancel_shell_job: Terminate background jobs gracefully or forcefully.
  • cancel_shell_monitor: Stop an output watch.

Sandboxed JavaScript Execution (run_code)

• run_code: Executes JavaScript in a sandboxed V8 context directly inside the Node.js process.
• Nested Tool Invocation: Inside a run_code script, the agent can call other tools programmatically via await tools.<toolName>(args).
• Tool Approvals in Code: Nested read-only tool calls run automatically, while actions requiring approval prompt the user before continuing.
• Script Isolation: Scripts execute in an isolated environment without access to process globals or host credentials.
• Nested Shell Commands: Scripts can invoke tools.shell with the same sandbox constraints, approvals, timeouts, and job tracking as top-level shell calls.
• JSON Inputs: Pass structured data separately from the script source with the inputs parameter; values are available through the script’s inputs binding.
• Result Contracts: Every scriptable tool declares its scripted return contract. Compact entries include return-shape guidance, and tools.describe exposes the full contract before invocation.
• Nested Trace: Tool calls made by a script are shown in the conversation and remain subject to the active approval policy.

User & Advisory Tools

• ask_user: Asks clarifying questions to the user with an interactive prompt and selectable choices.
• ask_mentor: Consults the higher-tier mentor model for architectural guidance or code review.
• activate_skill: Loads specialized instruction sets and workflows for domain-specific tasks.
• list_models: Discovers available models and registered providers with pagination, credential readiness, and local session usage stats.

Web Research

• web_search: Search the web through the configured Tavily, Exa, Firecrawl, or SearXNG provider.
• web_fetch: Fetch and extract content from a URL.