start_batch_task

Start durable per-item work; returns immediately and continues after this turn.

Use for independent subjects needing repeated multi-step work, regardless of domain: e.g. researching companies, products, or topics. Each item is one subject and outcome, not a tool call. Current executors require table-row or workflow targets; arbitrary chat/file outputs and external actions are not supported. Create tables only when requested or agreed. Submit research before doing it yourself; do not merely batch the final writes. Submit independent items once; use exact patches for mechanical edits and read-only research for enrichment. The activity panel reports progress and results after this turn. If admission is disabled, continue with normal tools within their limits and report unfinished work accurately; never claim a job exists without a successful submission.

items: [{key: unique string, executor: workflow_prompt_patch|table_update|assistant_enrichment, target_id: workflow or row ID, options: {…}}]. workflow_prompt_patch options: {node_id: “run”, replacements: [{old: exact text, new: replacement, count: expected matches}]}. table_update: {table_id, values: {column_key: value}} writes already-known values only. assistant_enrichment: {table_id, output_keys: [column_key], required_tools?: [tool_name]} uses read-only research then validates and writes only the named fields. Declare required_tools when a specific locally available read tool is essential; missing tools fail before research. In chat, omit assistant_id; the server uses this chat’s assistant. Outside chat, assistant_id is required. Read the actual table column keys and target IDs before submission; never guess them. Reuse request_key to recover a lost response. No workflow execution or arbitrary writes. Status is durable: do not poll in a loop or ask the user to say continue. User-facing descriptions should name the requested work, not batching or dispatch. Report outcomes and actionable blockers; keep routine progress in the thinking panel.

Parameters#

Generated by generate_mcp_docs from the server tool. Add to this page below this block; the block is rewritten on every run.

Research items submit structured results independently of their chat prose. Formatting correction and transient failures are handled automatically within the job budget; completed research and writes are not repeated. Genuine questions appear in the originating chat. Reply there to resume the affected items.

assistant_enrichment.options.required_tools optionally names up to ten essential local read tools. If any is unavailable after visibility and permission filtering, the item fails before research with a capability explanation. Native MCP tools are not currently available to durable research workers. Only declare tools whose exact names have been discovered; never invent names.