fix: workflow local mode — don't connect to Relaycast when disabled

When AGENT_RELAY_WORKFLOW_DISABLE_RELAYCAST=1 is set, the broker was
still receiving --channels which triggered Relaycast registration. Now:

- Skip --channels arg when channel list is empty (client.ts)
- Pass empty channels to broker when Relaycast is disabled (runner.ts)
- Validate config in execute() to catch DAG cycles before run starts
- Initialize abortController early so abort() works during setup
- Mark pending/running steps as failed on cancellation
- Fix continue strategy: treat failed steps as done for run completion
- Fix verification test: avoid anti-injection false positive

All 28 tests across workflow-dag, workflow-lifecycle, workflow-runner,
workflow-verification, and workflow-patterns now pass.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: khaliqgant

packages/sdk/src/client.ts

@@ -486,8 +486,7 @@ export class AgentRelayClient {
       'init',
       '--name',
       this.options.brokerName,
-      '--channels',
-      this.options.channels.join(','),
+      ...(this.options.channels.length > 0 ? ['--channels', this.options.channels.join(',')] : []),
       ...this.options.binaryArgs,
     ];
 

packages/sdk/src/workflows/runner.ts

@@ -1165,8 +1165,15 @@ export class WorkflowRunner {
     workflowName?: string,
     vars?: VariableContext
   ): Promise<WorkflowRunRow> {
+    // Set up abort controller early so callers can abort() even during setup
+    this.abortController = new AbortController();
+    this.paused = false;
+
     const resolved = vars ? this.resolveVariables(config, vars) : config;
 
+    // Validate config (catches cycles, missing deps, invalid steps, etc.)
+    this.validateConfig(resolved);
+
     // Resolve and validate named paths from the top-level `paths` config
     const pathResult = this.resolvePathDefinitions(resolved.paths, this.cwd);
     if (pathResult.errors.length > 0) {
@@ -1250,6 +1257,10 @@ export class WorkflowRunner {
 
   /** Resume a previously paused or partially completed run. */
   async resume(runId: string, vars?: VariableContext): Promise<WorkflowRunRow> {
+    // Set up abort controller early so callers can abort() even during setup
+    this.abortController = new AbortController();
+    this.paused = false;
+
     const run = await this.db.getRun(runId);
     if (!run) {
       throw new Error(`Run "${runId}" not found`);
@@ -1312,9 +1323,7 @@ export class WorkflowRunner {
     const { run, workflow, config, stepStates, isResume } = input;
     const runId = run.id;
 
-    // Start execution
-    this.abortController = new AbortController();
-    this.paused = false;
+    // Start execution (abortController already set by execute()/resume())
     this.currentConfig = config;
     this.currentRunId = runId;
     this.runStartTime = Date.now();
@@ -1386,7 +1395,7 @@ export class WorkflowRunner {
         this.relay = new AgentRelay({
           ...this.relayOptions,
           brokerName,
-          channels: [channel],
+          channels: relaycastDisabled ? [] : [channel],
           env: this.getRelayEnv(),
           // Workflows spawn agents across multiple waves; each spawn requires a PTY +
           // Relaycast registration. 60s is too tight when the broker is saturated with
@@ -1578,8 +1587,15 @@ export class WorkflowRunner {
       this.log(`Executing ${workflow.steps.length} steps (pattern: ${config.swarm.pattern})`);
       await this.executeSteps(workflow, stepStates, agentMap, config.errorHandling, runId);
 
+      const errorStrategy =
+        config.errorHandling?.strategy ?? workflow.onError ?? 'fail-fast';
+      const continueOnError =
+        errorStrategy === 'continue' || errorStrategy === 'skip';
       const allCompleted = [...stepStates.values()].every(
-        (s) => s.row.status === 'completed' || s.row.status === 'skipped'
+        (s) =>
+          s.row.status === 'completed' ||
+          s.row.status === 'skipped' ||
+          (continueOnError && s.row.status === 'failed')
       );
 
       if (allCompleted) {
@@ -1624,6 +1640,19 @@ export class WorkflowRunner {
       await this.updateRunStatus(runId, status, errorMsg);
 
       if (status === 'cancelled') {
+        // Mark any pending or in-progress steps as failed due to cancellation
+        for (const [stepName, state] of stepStates) {
+          if (state.row.status === 'pending' || state.row.status === 'running') {
+            state.row.status = 'failed';
+            state.row.error = 'Cancelled';
+            await this.db.updateStep(state.row.id, {
+              status: 'failed',
+              error: 'Cancelled',
+              updatedAt: new Date().toISOString(),
+            });
+            this.emit({ type: 'step:failed', runId, stepName, error: 'Cancelled' });
+          }
+        }
         this.emit({ type: 'run:cancelled', runId });
         this.postToChannel(`Workflow **${workflow.name}** cancelled`);
         await this.trajectory.abandon('Cancelled by user');

tests/integration/broker/workflow-lifecycle.test.ts

@@ -149,12 +149,23 @@ test('workflow-lifecycle: abort cancels a running workflow', { timeout: 120_000
       { cwd }
     );
 
+    // Wait for the step to actually start running before aborting
     let currentRunner = harness.getCurrentRunner();
     for (let i = 0; i < 20 && !currentRunner; i += 1) {
       await sleep(250);
       currentRunner = harness.getCurrentRunner();
     }
     assert.ok(currentRunner, 'Expected workflow runner to be available while running');
+
+    // Wait for step:started event before aborting so the step is actually in-flight
+    await new Promise<void>((resolve) => {
+      const unsub = currentRunner!.on((event) => {
+        if (event.type === 'step:started') {
+          unsub();
+          resolve();
+        }
+      });
+    });
     currentRunner.abort();
 
     const result = await runPromise;

tests/integration/broker/workflow-verification.test.ts

@@ -39,7 +39,7 @@ function makeConfig(overrides?: Partial<RelayYamlConfig>): RelayYamlConfig {
           {
             name: 'step-verify-output',
             agent: 'worker',
-            task: 'Return DONE',
+            task: 'Do one thing',
             verification: { type: 'output_contains', value: 'DONE' },
           },
         ],