Fix broker PID persistence and orphaned process detection

[khaliqgant]

Summary

• Write correct broker PID: Changed PID file to store the Rust broker process PID (relay.brokerPid) instead of the Node.js CLI PID (deps.pid). When the CLI exits but the persist-mode broker survives, the PID file now correctly identifies the running broker, preventing false "stale PID" detection on next agent-relay up.
• Fix orphaned broker detection: killOrphanedBrokerProcesses now greps by broker name (--name relay) instead of only by project root path. The installed binary (~/.agent-relay/bin/agent-relay-broker) doesn't contain the project root in its command line, so the old grep never matched orphaned brokers. Falls back to path-based grep for dev builds.
• Use fixed-string grep: Changed grep -e to grep -F for broker name matching to prevent regex metacharacter issues in project names containing dots or special chars.

Test plan

• All 22 existing core.test.ts tests pass
• Manual test: agent-relay up, kill the Node.js CLI (not the broker), run agent-relay up again — should detect running broker correctly
• Manual test: agent-relay down --force followed by agent-relay up — should start cleanly
• Verify PID file contains the Rust broker PID (not the Node.js PID) after startup

🤖 Generated with Claude Code

---

[devin-ai-integration]

Devin Review found 1 potential issue.

View 3 additional findings in Devin Review.

[devin-ai-integration]

🔴 Writing broker child PID instead of parent PID causes agent-relay down to orphan the parent CLI process and dashboard

The change from deps.pid to relay.brokerPid ?? deps.pid writes the broker child process PID to the PID file. When agent-relay down reads this PID and sends SIGTERM (src/cli/lib/broker-lifecycle.ts:1103), it kills the broker child directly. However, the parent agent-relay up process is still blocked on holdOpen() (src/cli/lib/broker-lifecycle.ts:997), which is a promise that never resolves (src/cli/commands/core.ts:339). The parent's SIGTERM/SIGINT signal handlers (src/cli/lib/broker-lifecycle.ts:976-995) are never triggered, so shutdownOnce() is never called, meaning the dashboard process (dashboardProcess) is never killed (src/cli/lib/broker-lifecycle.ts:701-706) and the parent process itself is never exited. This leaves both the parent CLI process and the dashboard as orphans after agent-relay down completes.

Affected scenario: background mode

In --background mode, the parent is a detached process (src/cli/lib/broker-lifecycle.ts:717-729). After agent-relay down kills only the broker grandchild, the detached parent becomes an invisible orphan that the user cannot easily discover or stop.

Suggested change

	writeBrokerPid(brokerPidPath, relay.brokerPid ?? deps.pid, deps);
	writeBrokerPid(brokerPidPath, deps.pid, deps);

---

Was this helpful? React with 👍 or 👎 to provide feedback.