Autoplan
gstack's /autoplan now runs end to end under a parent agent
resume marks an unfinished attempt EXECUTION_UNKNOWN instead of dispatching it again.gstack's /autoplan now runs end to end without Claude Code. A parent agent calls next, dispatches the reviewer and submits each phase. After a crash, resume flags an unfinished attempt instead of running it again.
Outside voice
The outside voice can be Codex, an API, or another model's subagent
codex directly for now.The outside reviewer is now pluggable: the Codex CLI, an API, or a subagent on another model family. Missing Codex no longer counts as degraded when another runner reviewed. Skill text still calls Codex directly for now.
Checklists
Plan reviews load a short checklist instead of the whole methodology
Unattended and light runs now give each plan review a checklist under 35 lines. They print the skipped sections, load the full one when scope is unclear, and interactive reviews stay the same.
Doctor
gstack-doctor --check confirms an install with one grep
The old Capy check passed machines with nothing installed. Now gstack-doctor --check prints one row per component, a fix for every failure, and a single trailer line with a real exit code.
Browser
Chromium installs the first time a browser skill runs
gstack-capy-install sets up a Capy machine without Chromium.Planning machines spent almost 90 seconds installing a browser they never opened. Now Chromium installs the first time a browser skill runs, and the new Capy installer sets a machine up without it.
Limits
Spend caps, runtimes and spec reviews report what they know
--spend-cap holds across a crash; subagent spend is unknown, not zerodocs/unattended.mdthe parent loopThe spend cap is now kept in a shared ledger that survives a crash. The doctor now fails the Bun version on the default Capy image. And spec review stops at its cap without a verdict. The unattended guide covers the parent loop.
