Struct ExecuteCommandTool 

pub struct ExecuteCommandTool;

execute_command — spawn a shell, run a command, capture output.

Honors three escape hatches:

• ExecContext::token (the main event): cancellation from the reducer aborts the child. This is the Ctrl+C fix.
• timeout argument: model-specified per-call cap (capped at COMMAND_MAX_TIMEOUT_SECS). Default COMMAND_TIMEOUT_SECS.
• Dangerous-command blocklist: refuses obvious destructive patterns before spawning.

Trait Implementations

impl ToolExecutor for ExecuteCommandTool

fn name(&self) -> &'static str

Canonical name the model uses to call this tool. Matches schema().name exactly.

fn schema(&self) -> ToolDefinition

JSON-schema description the model sees in the outgoing request. Adapters translate this into provider-native shape (Anthropic’s type: "custom", Gemini’s function_declarations, OpenAI’s flat tools, Ollama’s function calling). The same ToolDefinition feeds all four.

fn execute<'life0, 'async_trait>( &'life0 self, args: Value, ctx: ExecContext, ) -> Pin<Box<dyn Future<Output = ToolOutcome> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait,

Run the tool. The returned ToolOutcome is passed verbatim into Msg::ToolFinished — there’s no error-to-outcome conversion happening outside this function.

fn is_internal(&self) -> bool

True for tools that exist for internal dispatch only and should NOT be advertised to the model (e.g. the MCP proxy router, which fronts every mcp__server__tool call — the individual MCP tools are advertised separately from state.mcp.servers). Default false.