kanban-worker

To prevent failed handoffs in a multi-agent workflow, structure kanban_complete summary and metadata payloads with high-signal context for downstream parsers. This ensures downstream agents receive clear decisions, changed files, and measured results, enabling reliable task lifecycle transitions.

Tenant isolation for multi-agent worker lifecycles works by enforcing safe memory prefixing using the $HERMES_TENANT environment variable. This mechanism prevents cross-tenant context leaks during task processing, ensuring agents operate strictly within their designated tenant boundaries.

The best way to handle workspace state across scratch and Git worktrees is to adapt worker behavior based on the directory type. This prevents stale or mis-scoped edits during workflow orchestration by ensuring agents interact correctly with persistent directories and temporary worktrees.

To diagnose retry outcomes from previous kanban worker runs, interpret the results from kanban_show. This allows workers to adjust their operational behavior by understanding previous block reasons, heartbeat quality, and specific failure contexts before attempting the task again.

You should use precise kanban_block reasons during workflow orchestration when a worker hits a genuine operational wall. Providing specific block context instead of failing silently improves retry diagnostics and informs the dispatcher exactly why the task cannot proceed.