dev-team 14 replaces its context ceiling with Claude Code's autocompact

Version 14 of the dev-team plugin stops blocking skill loads when a session's context fills up. Instead it sets Claude Code's own autocompact threshold and lets the harness compact the conversation. That is a breaking change, and the fix is one command, dev-team setup, run in each repo.

Until now, a hook measured how full each session was. Past 40 percent of the window, capped at 350,000 tokens, it blocked skill loads until you wrote a handoff. That hook is gone. Claude Code's own autocompact takes over, at 40 percent by default, in subagents as well as the main session.

Setup merges a single entry into the repo's settings file and leaves your other keys alone. A flag picks a different percentage, or skips it entirely. The harness can only lower its threshold, so anything above its default does nothing.

A generic compaction summary can lose your place in a plan. A new hook runs right after compaction and puts back the build phase, step, plan file and unchecked items. Build also keeps a record between steps. Whether the hook fires after a real compaction is still being checked.

Nine review agents now carry a short authoring checklist. A new script picks the ones whose scope matches the files a build step will touch, and hands the engineer only those checklists. In return, the engineer stops preloading the full code review skill, about 25,000 tokens.

Ship used to have the model walk a long decision tree to tell whether a feature was shipped, in flight, or new. Now a script checks the pull requests and issues, returns one of seven verdicts, and asks you when the check itself fails. When build's reviews already passed every changed file, ship skips its second review.

The agent readiness scanner now scores how cheaply an agent can work in your repo. It wants a root CLAUDE.md of about 200 lines or fewer, rules layered into nested files, and one command that runs lint and tests. Overall scores shift, so compare only scans from the same scorecard version.

The old guard covered every repo automatically. Now a repo that skips setup keeps the harness default, which comes much later, and gets a one-line reminder. The 350,000 token cap is gone, so on a one million token window, 40 percent means 400,000 tokens. The harness checks between turns, so a long tool loop may overshoot. And handoff is now manual only.

To move to version 14, update the plugin with its upgrade command, then run dev-team setup in each repo. Until you do, every new session prints this reminder. If you relied on the old hard stop, set a lower percentage. That's dev-team 14, replacing its context ceiling with Claude Code's autocompact.