feat: channel-first workflow communication via Relaycast

[khaliqgant]

Summary

• Workflows now create a dedicated Relaycast channel from the YAML swarm.channel config and route all lifecycle events through it
• Upgrades @relaycast/sdk from 0.2.5 to 0.3.0 for channel management APIs
• Adds createChannel, joinChannel, inviteToChannel methods to RelaycastApi
• All spawned agents are invited to the workflow channel automatically
• Task assignments, step lifecycle events (started/completed/failed/retrying/skipped), and run completion/failure are posted to the channel
• Channel posts are fire-and-forget so they never block workflow execution

Test plan

• Run existing broker-sdk tests to verify no regressions
• Run a workflow with swarm.channel set and verify the channel is created in Relaycast
• Verify all agents appear in the channel member list
• Confirm task assignments and step lifecycle events appear as channel messages
• Verify workflow completes successfully even if Relaycast is temporarily unavailable

🤖 Generated with Claude Code

---

[devin-ai-integration]

Devin Review found 1 potential issue.

View 6 additional findings in Devin Review.

[devin-ai-integration]

🔴 Channel creation/join calls are blocking and can crash the workflow when Relaycast is unavailable

The createChannel and joinChannel calls in both execute and resume are not wrapped in try-catch, so if the Relaycast API is unavailable or returns an unexpected error, the entire workflow fails.

Root Cause and Impact

In execute at packages/broker-sdk/src/workflows/runner.ts:498-499:

await this.relaycastApi.createChannel(channel, workflow.description);
await this.relaycastApi.joinChannel(channel);

And in resume at packages/broker-sdk/src/workflows/runner.ts:636-637:

await this.relaycastApi.createChannel(resumeChannel);
await this.relaycastApi.joinChannel(resumeChannel);

These calls are inside the try block but have no individual error handling. If Relaycast is temporarily unavailable (network error, auth failure, etc.), the exception propagates and the workflow is marked as failed.

This contradicts the PR's stated design goal ("Channel posts are fire-and-forget so they never block workflow execution") and the test plan ("Verify workflow completes successfully even if Relaycast is temporarily unavailable"). The postToChannel helper at line 1094 correctly wraps its call in try-catch, but the channel setup calls do not follow the same pattern.

Impact: Any Relaycast outage or connectivity issue will prevent all workflows from running, even though channel communication is supposed to be non-critical observability infrastructure.

Prompt for agents

In packages/broker-sdk/src/workflows/runner.ts, wrap the channel setup calls in try-catch blocks so they are non-blocking, matching the fire-and-forget pattern used by postToChannel(). There are two locations to fix:

1. In the execute() method around lines 497-499:
   await this.relaycastApi.createChannel(channel, workflow.description);
   await this.relaycastApi.joinChannel(channel);

2. In the resume() method around lines 635-637:
   await this.relaycastApi.createChannel(resumeChannel);
   await this.relaycastApi.joinChannel(resumeChannel);

Both should be wrapped in try-catch with empty catch blocks (or logging), similar to how postToChannel() handles errors at line 1094-1101. For example:

try {
  await this.relaycastApi.createChannel(channel, workflow.description);
  await this.relaycastApi.joinChannel(channel);
} catch {
  // Non-critical — don't break workflow execution
}

This ensures workflows complete successfully even if Relaycast is temporarily unavailable.

---

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