Fix slow shutdown on SIGINT/SIGTERM

[brendanlong]

Summary

• Add SIGINT/SIGTERM signal handlers for graceful shutdown that stop active Claude queries and disconnect Prisma
• Fix setInterval in rate-limiter keeping the Node.js event loop alive by adding .unref()
• Add stopAllSessions() to claude-runner for bulk session cleanup during shutdown
• A second Ctrl+C force-exits immediately

Root Causes

The process was hanging on Ctrl+C because:

1. No signal handlers — there were no SIGINT/SIGTERM handlers, so Node.js relied on the event loop draining naturally
2. Rate-limiter setInterval — the hourly cleanup timer (line 216 of rate-limiter.ts) was never .unref()'d, so it kept the event loop alive indefinitely
3. Active Claude queries — running SDK queries weren't interrupted during shutdown
4. Prisma connections — the database connection pool wasn't closed

Test plan

• Unit tests pass (pnpm test:run — 361 tests)
• Integration tests pass (pnpm test:integration — 143 tests)
• Manual test: pnpm start, then Ctrl+C — should exit promptly
• Manual test: start a Claude query, then Ctrl+C — should stop query and exit
• Manual test: second Ctrl+C should force-exit immediately

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces graceful shutdown handling by registering SIGINT and SIGTERM listeners to stop active Claude sessions and disconnect Prisma. It also applies .unref() to the rate limiter's cleanup interval to prevent it from hanging the process. A review comment highlights that the registration logic should be idempotent to prevent multiple listeners from being attached during Next.js development.

[gemini-code-assist]

In Next.js, the register function in instrumentation.ts can be executed multiple times during development (e.g., due to Hot Module Replacement). This would result in multiple SIGINT/SIGTERM listeners being attached to the process, leading to redundant cleanup operations and potential race conditions during shutdown. It is recommended to ensure the signal handlers are only registered once by using a global flag, ensuring that initialization logic is idempotent and protected against concurrent or repeated execution.

    if (!(globalThis as any)._shutdownHandlerRegistered) {
      registerShutdownHandler();
      (globalThis as any)._shutdownHandlerRegistered = true;
    }

References

1. To prevent race conditions in methods with asynchronous initialization, set a processing/locked flag synchronously at the start of the method, before any await calls, to ensure concurrent invocations are rejected immediately.