Run coding tickets through a planning agent, builders and a review panel

Drag built Triglav, a local orchestrator that runs the Claude Code and Codex CLIs as a small consultancy and pointed it at Drag's own repositories. A long-running Staff session turns a ticket or prompt into a goal and acceptance criteria, which Drag approves, then starts up to two Senior subagents per job, each in its own git worktree opening a pull request. Seniors post open questions to a shared job board, where Staff and short-lived researcher runs answer with a stance before Staff records a binding decision. Every commit goes through a read-only review panel, with a Codex model as the main reviewer and a cheaper Claude model as second opinion, then an optional Copilot review, then a CI check Staff runs itself. Drag approves plans, answers escalations and merges.

In the first ten days it ran 41 jobs and finished 34. The review panel completed 96 rounds, passing 54 and sending 42 back for changes. In the first six days the agents wrote about 3.6 million output tokens but read about 3.6 billion from prompt cache, because a waking Staff re-reads 300k to 700k tokens of context. Once a session that hit a usage limit resumed itself while Drag also resumed it, leaving two Staff agents on one job.

All of it runs on your own Claude or Codex plan and Staff starts agents without asking, so watch usage. Drag limited wake-ups and compacted Staff context at about 280k tokens to cut re-reading. Merge and deploy rights are only an instruction to the agents, so protect branches on GitHub. The numbers are one person's use on Linux.