Struct VerifyCommandTool 

pub struct VerifyCommandTool { /* private fields */ }

The verify tool: runs the harness-configured verification command(s) in the workspace and reports a structured, gate-able PASS/FAIL. Its two reasons to exist over raw bash: (1) the command is configured by the harness (the agent never guesses what “passing” means), and (2) the result is a machine-greppable VERIFY_SENTINEL the goal judge / a deterministic gate can key on. Commands run sequentially and fail-fast (stop at the first non-zero exit), so the final sentinel is the overall verdict.

Trust model: the command is operator-configured, not agent-supplied, so it runs with ambient authority — like a CI command — and is deliberately NOT sandboxed the way the agent-facing bash tool is (it does not apply BashTool::with_sandbox’s Landlock / EnvPolicy / CorePathPolicy). It intentionally does not reuse BashTool for that reason: the agent’s bash sandbox and the operator’s verify authority are kept separate, and the small spawn/timeout plumbing here is the price of that separation.

Implementations

impl VerifyCommandTool

pub fn new(workspace: impl Into<PathBuf>, commands: Vec<String>) -> Self

Create a verify tool that runs commands (in order, fail-fast) with the workspace as the working directory.

pub fn timeout(self, timeout: Duration) -> Self

Override the per-command timeout (default 300s).

pub fn max_tail(self, max_tail: usize) -> Self

Override the per-command output tail cap (default 4000 bytes).

Trait Implementations

impl Tool for VerifyCommandTool

fn definition(&self) -> ToolDefinition

Return the tool’s definition (name, description, input schema) for the LLM.

fn execute( &self, _ctx: &ExecutionContext, _input: Value, ) -> Pin<Box<dyn Future<Output = Result<ToolOutput, Error>> + Send + '_>>

Execute the tool with the LLM-supplied input JSON.

fn redact_for_history(&self, output: &str) -> String

Optional override: produce a compact version of the tool output suitable for inclusion in conversation history sent to subsequent LLM turns. The default implementation returns the input verbatim.