Coven kills claude sessions before they get to record the final message

[Crazytieguy]

There was a recent change (related to interaction with background bash commands I think) where we switched from closing stdin for claude to killing it in some cases. I think this was probably the right call, but there's a side effect: we kill the claude session before it has a chance to save the last message to the session file (so a resumed session might not have it). Please investigate and propose possible fixes

[Crazytieguy]

@claude

[claude]

View run

Completed work on #32

• Read issue
• Investigate the session killing behavior
• Identify relevant code paths (stdin close vs kill)
• Implement graceful shutdown method
• Replace kill() calls with shutdown() in ralph, worker, fork, reload
• Run tests (all 149 pass)
• Push and create PR

PR: #34

Changes: Added SessionRunner::shutdown() which closes stdin, waits up to 2s for claude to save session state, then falls back to SIGKILL. Replaced runner.kill() with runner.shutdown() in ralph, worker, fork, and reload paths. Kept immediate kill() for Ctrl+C interrupts.