Files

2.4 KiB

name, description, version, author, license, platforms, metadata
name description version author license platforms metadata
kanban-collaboration Umbrella for Hermes Kanban collaboration: orchestrator decomposition, worker lifecycle pitfalls, task graph routing, and board hygiene. 1.0.0 Hermes Agent MIT
linux
macos
windows
hermes
tags related_skills
kanban
orchestration
workers
task-graph
multi-agent
kanban-orchestrator
kanban-worker
hermes-agent

Kanban Collaboration

Use this umbrella when operating Hermes' durable Kanban board for multi-agent or multi-profile work. It covers both orchestrator behavior (decompose and route work) and worker behavior (execute an assigned card safely and report completion/blockers).

Orchestrator responsibilities

  • Decide whether the board is warranted; do not route trivial single-agent tasks through Kanban.
  • Understand the goal and constraints before creating cards.
  • Sketch a task graph with dependencies, not a flat TODO dump.
  • Create small, independently verifiable cards with clear acceptance criteria.
  • Assign to profiles based on user-configured roles/capabilities; do not assume a fixed roster.
  • Resist doing worker tasks yourself. The orchestrator's job is decomposition, routing, monitoring, and integration.

Worker responsibilities

  • Treat the assigned card and board context as the source of truth.
  • Work in the assigned workspace/tenant; do not bleed state across tenants.
  • If creating follow-up cards, capture returned IDs before linking/claiming them.
  • Complete only when acceptance criteria are met and evidence is attached.
  • Block quickly with actionable reasons when credentials, permissions, missing context, or upstream dependencies prevent progress.

Task graph hygiene

  • Link dependencies explicitly.
  • Keep cards narrow enough to verify but broad enough not to create micro-task noise.
  • Use comments for discoveries that future workers need.
  • Reclaim or unblock stale tasks only after checking recent activity.
  • Summaries should include files changed, commands run, artifacts produced, and remaining risks.

Good block reasons

A block reason should allow the orchestrator/user to act immediately: missing API key name, inaccessible repository/path, ambiguous product decision, failing upstream card, or unavailable service. Avoid vague 'could not proceed' messages.

Verification

The board is durable; verify by reading cards/comments/links back after creating, updating, completing, blocking, or linking tasks.