Review the design,
not the diff.
Agent writes a design doc agent
You tell the agent what to build. It gathers context, researches approaches, and writes a design document with architecture, data model, and trade-offs.
You review and approve you
Read a 2-page design doc instead of reviewing a 500-line PR. Approve or request changes. Architectural mistakes are caught before they become code.
Agent writes an implementation plan agent
Concrete steps grounded in the actual codebase. Each step is independently verifiable. You review, reorder, cut. The agent creates tasks from the approved plan.
Agent executes, step by step agent
One task at a time. Each produces a typed deliverable — code change, document, configuration. Every step is tracked, auditable, resumable across sessions.
Approval gates are human-only
The MCP tools don't expose approve or request-changes. Agents can propose designs and plans, but only a human in the dashboard can advance past a review gate.
Every decision is traceable
Why did we choose this approach? Trace from the task to the plan step, to the design doc, to the comment thread where the trade-off was discussed. The audit trail is automatic.
Knowledge compounds across issues
Every issue is searchable. The tenth issue planned through taskum is better than the first because the agent can reference the previous nine.
$ claude mcp add taskum -- taskum serve
$ opencode mcp add taskum -- taskum serve
$ codex mcp add taskum -- taskum serve
The planning layer that persists.
Every design, every plan, every decision — searchable, auditable, shareable.