Tools and Output Limits

Tools and Output Limits

Built-in tools

Tool	What it does	Default permission
read	read a file (text, or an image the model can see)	allow
write	create or overwrite a file	ask
edit	replace text in a file	ask
bash	run a shell command	ask
glob	find files by pattern (gitignore-aware; ripgrep when available)	allow
grep	search file contents (gitignore-aware; ripgrep when available)	allow
todo	keep the agent's task list	allow
skill	list skills and load one (Skills)	allow
task	spawn a subagent (Agents)	allow

MCP servers add more tools, named mcp__<server>__<tool> (MCP Servers). See Permissions and Trust to change the defaults.

bash runs every command in a new session with no controlling terminal, so it can never read from or take over the TUI's terminal. Cancelling a run (or a command timing out) kills the command's whole process group, background children included.

Output limits

Every tool result the model sees is capped at 50 KB (cut at a UTF-8 boundary, with [output truncated at 50 KB] appended). Some tools bound their own output more tightly:

• read stops before 50 KB and says which offset to continue from.
• grep lines are cut to 500 characters, and the search stops once 100 matches are found.
• bash keeps the last 30 KB of output and saves the full output to a private per-run temp directory (mode 0700, removed when jig exits).