v1.23.24

• agents fork <id> --device <box> works again, and so does every other command that declares its own --device. The fleet router rejected --device on any top-level name missing from its hand-kept list, so agents fork, agents cloud run, agents projects status and agents projects pull all failed with "does not support --device". The router now steps aside when the resolved command declares --device itself (the --device that addHostOption adds stays the router's, and --devices/--hosts route exactly as before), and a test walks the whole command tree so a moved or new command cannot lose --device silently. agents config list --device <name> now runs its own handler (that device's config read here) instead of being forwarded to the device. The fork sibling starts in the source session's folder (--cwd) when it lands on the machine that holds the source session, and agents ps preview exits 1 when the hop to the owning device fails, so agents fork reports the peer's error instead of "could not read the source". Source: cli/src/lib/hosts/routing-flag.ts, cli/src/lib/hosts/passthrough.ts, cli/src/commands/fork.ts, cli/src/commands/ps.ts.

• --device auto skips boxes with a nearly full disk and honours preferred everywhere. A device with under 20 GiB free on / now reads disk low in agents devices list and is excluded from automatic placement (reason disk-low), since an agent that cannot create its worktree does nothing. Memory pressure is judged by free GiB rather than percent used, so a 120 GiB box at 48% no longer reads busy and loses every pick to a small idle box. agents devices pick (and so scripts/test.sh) and agents teams placement now apply the auto-launch.preferred boost that agents run --device auto already did. agents devices list --json adds autoSkip, the reason auto placement would pass a device over (off, personal, desktop, not-a-worker, offline, unreachable, loaded, disk-low), or null. Source: cli/src/lib/devices/health.ts, cli/src/lib/devices/pool.ts, cli/src/lib/teams/scheduler.ts.