Graceful opencode session hibernation (Session #1)

[JRedeker]

Summary

Implement graceful hibernation for idle OpenCode sessions. Instead of leaving expensive LLM sessions running, hibernate (SIGSTOP or snapshot) and restore on reattach.

Why

Running OpenCode sessions consume LLM API quota even when idle. Hibernation frees resources while preserving session state.

Acceptance Criteria

• Idle timeout triggers hibernation (configurable)
• Session state preserved across hibernate/restore
• No data loss on restore
• Status bar shows hibernated sessions distinctly
• oca attach transparently restores

Dependencies

Blocked by Session #0 (Pattern B topology).

Proposal

docs/proposals/2026-05-03-session-and-resource-architecture.md § Change #1

WSJF Estimate

Value: 7 | TC: 5 | RROE: 7 | Effort: 8 | WSJF: 1.3

[JRedeker]

Deferred to v1.1 — conditional on RAM still hurting (2026-05-11)

Research findings during v1 scope refinement:

1. Operator workload target revised: 15 simultaneous active agents (not 40 with hibernation rotating).
2. Current opencode RAM footprint: 8 procs × ~750 MB = 6 GB (modest).
3. Real RAM bottleneck found: Advance Temporal worker singleton broken \u2014 each opencode session spawns 2 workers, not 1 per project. Result: ~2.2 GB wasted across 4 projects. Filed Temporal worker singleton broken: every opencode session spawns 2 workers, not 1 per project (~2 GB waste) Advance#117 to fix upstream.
4. Hibernation would save ~3 GB at 50% idle of 8 active opencode procs. Worker-leak fix saves ~2.2 GB without OCA work.

Conclusion: with worker leak fixed at the Advance layer + Pattern B cutover, RAM pressure should resolve for the 15-active-agent target. Hibernation becomes optional.

Revisit criteria for v1.1:

• After v1.0 cutover, operator on Pattern B with worker leak fix shipped (Temporal worker singleton broken: every opencode session spawns 2 workers, not 1 per project (~2 GB waste) Advance#117)
• If 15-agent workload still stresses RAM, OR if workload target grows past 20 agents, reopen and ship as v1.1.
• UX details (idle threshold, opt-out keybind, \ud83d\udca4 marker) will be designed against real Pattern B usage data instead of paper specs at that point.

Closing in OCA Project board but keeping issue open as v1.1 candidate with the conditional reopen criteria above.