v0.63.1

Fixed

• The desktop agent can start long commands in the background, read them and stop them, inside run_shell's fences. A bench stage driven through the MCP bridge was cut at 120 s. The existing run_shell(background=true), job_status and job_cancel now close the gaps:
  • Limits: at most 3 concurrent jobs (CHIMERA_JOBS_MAX_RUNNING) and 6 hours per job (CHIMERA_JOBS_MAX_RUNTIME). An overrun job has its whole process tree killed.
  • Exit recording: the exit code and end time are recorded the moment the process ends.
  • Bounded reads: output is read by head and tail.
  • Isolation: each workspace sees and stops only its own jobs.
  • Orphans: an orphaned job reads lost, never running.
  • App exit: closing the app ends its jobs. On Windows they run in a kill-on-close Job Object.
  • POSIX sandbox: bwrap and Seatbelt now wrap the job, instead of refusing it.
  • Same fences as the foreground: the same approval card, the same workspace jail, the same scrubbed environment and the same taint marking.
  • Timeout hint: a timeout now points to background=true.
  • Where to see it: a "Background jobs" panel on the Code screen and a new MCP bridge area, desktop_shell_jobs (list, read, stop). (#658)