🔧 fix: autonomous agent scheduling, provider fallback, and duration bugs

- Background agent manager: use lastCycleToolCalls instead of lifetime
  totalToolCalls for adaptive scheduling — agents now properly accelerate
  when active and slow down when idle on a per-cycle basis
- Autonomy engine + background agent runner: replace silent hardcoded
  fallback to openai/gpt-4o-mini with clear error when no provider
  is configured, preventing confusing auth errors
- Subagent manager: use startedAt (when available) instead of spawnedAt
  for durationMs calculation on cancel/error, excluding queue wait time
- Fix edge-mqtt-client tests: add vi.useRealTimers() in afterEach to
  prevent fake timer leak across tests (3 pre-existing failures fixed)

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

Author: ersinkoc

packages/gateway/src/autonomy/engine.ts

@@ -471,8 +471,13 @@ export class AutonomyEngine implements IPulseService {
       const { getOrCreateChatAgent } = await import('../routes/agent-service.js');
       const { resolveForProcess } = await import('../services/model-routing.js');
       const resolved = await resolveForProcess('pulse');
-      const provider = resolved.provider ?? 'openai';
-      const model = resolved.model ?? 'gpt-4o-mini';
+      if (!resolved.provider || !resolved.model) {
+        throw new Error(
+          'No AI provider configured for autonomous pulse. Set a default provider in Settings → AI Models.'
+        );
+      }
+      const provider = resolved.provider;
+      const model = resolved.model;
       const fallback =
         resolved.fallbackProvider && resolved.fallbackModel
           ? { provider: resolved.fallbackProvider, model: resolved.fallbackModel }

packages/gateway/src/services/background-agent-manager.ts

@@ -60,6 +60,7 @@ interface ManagedAgent {
   cyclesThisHour: number;
   hourWindow: number; // hour timestamp for rate limiting
   persistTimer: ReturnType<typeof setInterval> | null;
+  lastCycleToolCalls: number; // tool calls in the most recent cycle (for scheduling)
 }
 
 // ============================================================================
@@ -205,6 +206,7 @@ export class BackgroundAgentManager {
       cyclesThisHour: 0,
       hourWindow: this.getCurrentHour(),
       persistTimer: null,
+      lastCycleToolCalls: 0,
     };
 
     this.agents.set(config.id, managed);
@@ -427,8 +429,8 @@ export class BackgroundAgentManager {
       if (lastResult) {
         // Backoff on error
         delay = CONTINUOUS_MAX_DELAY_MS;
-      } else if (managed.session.totalToolCalls === 0) {
-        // No activity yet — use idle delay
+      } else if (managed.lastCycleToolCalls === 0) {
+        // Last cycle was idle (no tool calls) — slow down
         delay = CONTINUOUS_IDLE_DELAY_MS;
       } else {
         // Active — fast delay
@@ -571,6 +573,7 @@ export class BackgroundAgentManager {
     session.lastCycleAt = new Date();
     session.lastCycleDurationMs = result.durationMs;
     session.lastCycleError = result.error ?? null;
+    managed.lastCycleToolCalls = result.toolCalls.length;
 
     // Clear inbox after cycle (messages were included in the cycle prompt)
     session.inbox = [];

packages/gateway/src/services/background-agent-runner.ts

@@ -88,8 +88,15 @@ export class BackgroundAgentRunner {
         model = this.config.model;
       } else {
         const resolved = await resolveForProcess('pulse');
-        provider = this.config.provider ?? resolved.provider ?? 'openai';
-        model = this.config.model ?? resolved.model ?? 'gpt-4o-mini';
+        const resolvedProvider = this.config.provider ?? resolved.provider;
+        const resolvedModel = this.config.model ?? resolved.model;
+        if (!resolvedProvider || !resolvedModel) {
+          throw new Error(
+            'No AI provider configured. Set provider/model on the agent or configure a default in Settings.'
+          );
+        }
+        provider = resolvedProvider;
+        model = resolvedModel;
       }
 
       // 2. Create agent with full tool access

packages/gateway/src/services/edge-mqtt-client.test.ts

@@ -130,6 +130,7 @@ describe('EdgeMqttClient', () => {
   });
 
   afterEach(() => {
+    vi.useRealTimers();
     delete process.env.MQTT_BROKER_URL;
   });
 

packages/gateway/src/services/subagent-manager.ts

@@ -197,7 +197,8 @@ export class SubagentManager {
     managed.session.state = 'cancelled';
     managed.session.completedAt = new Date();
     managed.session.durationMs =
-      managed.session.completedAt.getTime() - managed.session.spawnedAt.getTime();
+      managed.session.completedAt.getTime() -
+      (managed.session.startedAt ?? managed.session.spawnedAt).getTime();
 
     log.info(`Cancelled subagent "${managed.session.name}" [${subagentId}]`);
 
@@ -348,7 +349,8 @@ export class SubagentManager {
         managed.session.error = getErrorMessage(err);
         managed.session.completedAt = new Date();
         managed.session.durationMs =
-          managed.session.completedAt.getTime() - managed.session.spawnedAt.getTime();
+          managed.session.completedAt.getTime() -
+          (managed.session.startedAt ?? managed.session.spawnedAt).getTime();
 
         log.error('Subagent execution error', {
           subagentId: id,