Mothership

The problem

Coding agents are fast and forgetful. They skip requirements, claim things are “done” without proof, and leave no trail you can audit a month later. I wanted agents that work like a disciplined delivery team.

The flight plan

Every ticket flies the same six-stage route, and each stage hands off through a gate:

GROUND → CLARIFY → CANONICAL → TEMPLATE → SIMPLIFY → VERIFY → Closed → (deploy)
analyst  analyst   planner     builder    builder    verifier          deployer
                    ↑ fixer can be called from any stage ↑

• harness, the orchestrator, routes every stage and never writes code itself.
• analyst grounds the request in the real codebase, then clarifies until requirements pass a 10-point freeze gate.
• planner turns frozen requirements into a canonical plan with tasks and estimates.
• builder implements against templates, then a dedicated pass simplifies the result.
• verifier won’t close a ticket until every acceptance criterion has cited evidence.
• fixer finds the root cause and applies the smallest patch, at any stage.
• deployer always asks before publishing anything.

The flight deck

The Delivery Console is a zero-build web app (stdlib Python plus vanilla JS). It’s the only thing allowed to write ticket state.

• Kanban with a WIP limit, plus investigations, timesheets and analytics.
• Live agent chats. It streams Claude Code sessions as stream-json, so you can steer, queue or interrupt mid-turn and pick the model.
• Approval gate. A PreToolUse hook holds risky tool calls for Allow once / Allow for this chat / Deny. If nobody answers, the call is denied.
• Voice. Replies can be read aloud, and it says out loud when it needs you, e.g. “Permission needed for Bash.”
• Vault view. An Obsidian knowledge vault with a wikilink force-graph of every decision.
• Desktop. A Tauri shell wraps all of it.

Key decisions

• Files over databases. Markdown for prose, TOML for structure. Humans and agents read the same files.
• Traceability is a gate, not a report. Requirement → task → code → evidence. A gap blocks the ticket.
• One harness, two IDEs. The same agents and skills run in Claude Code and Cursor, with no duplicate definitions.

What’s next

Mothership is the v2 of the harness line. Its lessons feed Helmlock, which keeps the ideas and keeps the kernel small.