Struct CustomTaskHandler 

pub struct CustomTaskHandler { /* private fields */ }

TaskHandler that injects a custom prompt into the agent loop.

When a TaskKind::Custom task is due, CustomTaskHandler reads the "task" field from the task’s JSON config, sanitises it with crate::sanitize_task_prompt_checked, and sends the resulting string on the provided mpsc::Sender. The agent loop receives the prompt and processes it as a new user message.

Sending is best-effort: if the channel is full or closed, the error is logged at warn level and Ok(()) is returned so the scheduler continues running.

Injection pattern detection: if the prompt contains a known injection marker, SchedulerError::PromptInjectionBlocked is returned and no message is sent.

Examples

use tokio::sync::mpsc;
use zeph_scheduler::CustomTaskHandler;

let (tx, mut rx) = mpsc::channel(8);
let handler = CustomTaskHandler::new(tx);

use zeph_scheduler::TaskHandler;
handler
    .execute(&serde_json::json!({"task": "Generate a daily report"}))
    .await
    .expect("handler should not fail");

let prompt = rx.recv().await.unwrap();
assert_eq!(prompt, "Generate a daily report");

Implementations

impl CustomTaskHandler

pub fn new(tx: Sender<String>) -> Self

Create a new handler that sends prompts on tx.

task_name is included in SchedulerError::PromptInjectionBlocked when an injection pattern is detected, enabling structured log correlation.

pub fn with_task_name(tx: Sender<String>, task_name: impl Into<String>) -> Self

Create a new handler with an explicit task name for diagnostics.

Trait Implementations

impl TaskHandler for CustomTaskHandler

fn injects_agent_prompt(&self) -> bool

Always true: every execution sends the sanitised prompt on tx for the agent loop to process as a new user message, so RTW-A Mechanism 4 must be able to suppress it.

fn execute( &self, config: &Value, ) -> Pin<Box<dyn Future<Output = Result<(), SchedulerError>> + Send + '_>>

Execute the task with the provided configuration.

fn reads_external_content(&self) -> bool

Returns true when execute reads content that originates outside the scheduler’s own trusted config or in-process state — e.g. a network/HTTP response, an MCP tool result, a file loaded from a plugin-marketplace skill directory, or stored memory/graph facts that may themselves have been populated from such an external source.