dispatcher-core

What problem does it solve?

It eliminates brittle, one-off automation by providing a reliable way to orchestrate multi-step workflows that call Linear, Gmail, Calendar, Margot, Composio, or any registered MCP tool and persist progress across crashes and restarts.

Core Features & Use Cases

• Ordered cross-tool flows: Executes a declared sequence of steps where each step’s inputs can be derived from earlier step outputs.
• Resumable, journaled execution: Persists flow state to .harness/dispatcher_state.json and appends an immutable audit trail to .harness/swarm.jsonl.
• Step-level error handling: Supports abort, log-and-continue, and retry behavior per step to match real workflow tolerance.
• Kill-switch aware pausing: Observes TAO_SWARM_ENABLED mid-flow and pauses without losing progress.
• Safety guardrails: Enforces HITL routing for gated sending steps (e.g., telegram-draft-for-review), applies per-flow tool allowlists, prevents unbounded recursion, and guarantees immutable auditing.

Example use case: Create a Linear issue, run deep research via Margot, then prepare a Telegram draft for HITL review—all while guaranteeing resumability and controlled failure behavior.

Quick Start

Ask the Dispatcher to execute a flow that chains an MCP tool to create a ticket, calls a research skill, and then generates a Telegram draft-for-review, using a max_runtime_minutes limit and kill_switch_aware execution.