OpenYabby

Second patch release. Adds OS-level tracking for Bash(run_in_background) jobs and three guards that prevent the lead from spinning in delivery loops.

Added

• OS-level tracking of Bash run_in_background jobs. New bg_tasks table, PreToolUse hook that captures the host PID (sh -c 'echo \$\$ > <file>; exec <cmd>'), and a watcher polling kill -0 every 30s. When a bg job ends after the parent CLI has exited, the agent receives a [BG_COMPLETED] follow-up so it can report the result to the user via WhatsApp / voice / web. Exposed via GET /api/tasks/:id/bg-tasks and GET /api/agents/:id/bg-tasks.

Fixed

• Lead looping "project done" after delivery. The plan_continuation auto-poke now skips when an active presentation has last_run_status='passed'. Previously the lead replied "done" once per 10-second cycle for as long as the project stayed active (75 cycles in 45 min observed).
• Sub-agent completions not reaching the lead past Phase 1. The orchestrator's phase='discovery' skip now also requires no approved plan — tasks.phase is never advanced from 'discovery', so leads were silently dropped after their first plan approval.
• Review pile-up on cascading sub-completions. The orchestrator skips a new orchestrator_review when one is already pending or processing — the existing review already reads all pending inbox messages in one go. Without this, 4 cascading completions could produce 6+ duplicate reviews via the delayed-check path.
• Watchdog killed live bg jobs (regression from 0.1.1). Removed the 10s FINAL_OUTPUT watchdog and the close-handler SIGKILL of the process group. Bg jobs now survive CLI exit; the watcher above detects completion via OS PID.

Full Changelog: https://github.com/OpenYabby/OpenYabby/compare/v0.1.1...v0.1.2