Human Input

Human input

Use request_human only when the answer changes what the agent should do.

• Prefer choices whenever the answer can be bounded. This keeps the fast Decision/swiper interaction.
• Use fields only for 1 to 5 small values: text, number, boolean, date, or select.
• Use a Human Task when the person must perform work, inspect files, upload evidence, or produce a deliverable.
• Use a normal message, activity entry, or notification when no response is needed. Never create a Decision merely to announce progress.
• Never request passwords, tokens, credentials, payment-card data, or other secrets.
• blocking: true means execution cannot safely continue. Call request_human, then stop the current turn.
• blocking: false is advisory. Continue immediately; its response informs future work and never rewinds completed work.
• Do not poll repeatedly. The remote worker resumes the same task after a blocking response is recorded.

Discover request_human with search_tools, then invoke it with execute_tool.

Tools Used

• search_tools — discover the current human-input contract.
• execute_tool — invoke the discovered tool through compact MCP mode.
• request_human — create the governed blocking or advisory interaction.