Tools & review

Tools are how the model acts. Every tool call goes through the same pipeline: review (compute what to show, decide whether to auto-run), then execute, then the result is sent back as a message.

Built-in tools

Tool	Effect	Review
`read_files`	Read one or more files	auto
`search_lores`	Regex search over lore libraries	auto
`exec_shell`	Run a shell command	manual
`diff`	Apply a unified diff to a file	manual
`replace`	Apply exact search/replace patches	manual
`agent`	Launch a sub-agent	manual

Auto-execution is not a hardcoded list: a tool auto-runs when its RPC is declared idempotency_level = NO_SIDE_EFFECTS. The same rule applies to remote tool engines, so a GetIssue runs silently while UpdateIssue waits for you.

What you see

Each tool renders its own request:

exec_shell shows the command as a sh code block;
diff and replace show a unified diff computed at review time — the patch has already been applied to a copy and validated, so what you approve is exactly what lands;
engine RPCs show Service/Method and the request as JSON.

Verdicts

When a call awaits review the status line says so and the input box turns into a reason field:

Key	Verdict
`alt`+`y`	Accept the call under review
`alt`+`shift`+`y`	Accept every pending call
`alt`+`shift`+`r`	Reject; the input text is sent back as the reason
`alt`+`shift`+`a`	Always accept this tool for the rest of the session

A rejected call returns an error result carrying your reason, so the model can adjust rather than retry blindly. Cancelling a turn (ctrl+c) marks every unresolved call as never run.

Choosing tools

Tools are the union of three sources, deduplicated:

--tool name on the command line (repeatable);
@tool(...) directives of the role (and included roles);
chat.default_tools from the configuration.

Inside a chat alt+shift+t opens a fuzzy picker to toggle any tool — built-in or engine — for the current chat.