0.2.17

번역 준비 중

이 쪽은 아직 한국어로 옮기지 않았습니다. 아래는 영어 원문입니다.

Behavior changes

• with harness.queue.directSay on (default off), direct harness say/ask calls go through the seat queue (seat cap, order, same-cell refusal) and print a one-line reason; --no-queue is the emergency bypass.
• nested elanous launches are denied for Claude children through a strict MCP config, and nesting stops at a configurable depth cap.
• the hq-fence cron wrapper is generated by elanous hq fence-wrapper install, runs from the installed build instead of a stray worktree, and alerts when the fence fails.
• fix: the hq-fence cron wrapper now runs bun by the absolute path resolved at install time (cron has no PATH for bun), and writes a log line if even that bun is missing.
• every run-supervisor stop reason now maps to a named next step (harvest, decompose, self-review, human gate), recorded with the run.
• every landing path, including host-side unattended merges, now checks the operational landing freeze at one pre-merge point, and the release run's own version bump can pass an explicit freeze override.
• failed Pod runs now report the child's real stop reason (node, outcome, supervisor verdict) instead of a stray heal-loop line.
• authoring clarification questions are now self-answered in remote Pods for goal-file launches and supervisor rework passes, not only for the first say pass.
• a Pod launch that waits too long in admission or for a pool slot now logs launch-stalled with the measured reason, and a pool whose members can never hold the Job's memory limit fails fast (pod-pool-unfit) instead of waiting forever.
• Pod authoring no longer passes --author-grade to the in-Pod elanous (it rejected the flag); lite authoring is local-only for now, Pod authoring uses full until 0.2.18 reconnects it.

Plus 30 internal changes.