diff --git a/workspace-server/internal/handlers/registry.go b/workspace-server/internal/handlers/registry.go
index 047b8ea1d..2856e5ab0 100644
--- a/workspace-server/internal/handlers/registry.go
+++ b/workspace-server/internal/handlers/registry.go
@@ -869,6 +869,29 @@ func (h *RegistryHandler) evaluateStatus(c *gin.Context, payload models.Heartbea
 		})
 	}
 
+	// Auto-recovery from failed: the provision-timeout sweeper
+	// (registry/provisiontimeout.go) flips a workspace to 'failed' when it sits
+	// in 'provisioning' past DefaultProvisioningTimeout (10m for claude-code).
+	// But a slow cold-boot (EC2 image pull + LLM preflight) can still finish and
+	// start heartbeating AFTER the flip — agent_card is written unconditionally
+	// on register, so the box is genuinely serving while its status is stuck
+	// 'failed'. A live heartbeat is authoritative: recover to online. Without
+	// this, a healthy-but-slow workspace (e.g. a model that preflights slower
+	// than the 10m budget) shows 'failed' forever despite working — the
+	// mechanism behind the intermittent multi-provider e2e "boot failures". The
+	// `AND status = 'failed'` guard keeps the flip conditional (won't override
+	// 'removed').
+	if currentStatus == "failed" {
+		if _, err := db.DB.ExecContext(ctx, `UPDATE workspaces SET status = $1, updated_at = now() WHERE id = $2 AND status = 'failed'`, models.StatusOnline, payload.WorkspaceID); err != nil {
+			log.Printf("Heartbeat: failed to recover %s from failed: %v", payload.WorkspaceID, err)
+		} else {
+			log.Printf("Heartbeat: transitioned %s from failed to online (late heartbeat after provision-timeout)", payload.WorkspaceID)
+		}
+		h.broadcaster.RecordAndBroadcast(ctx, string(events.EventWorkspaceOnline), payload.WorkspaceID, map[string]interface{}{
+			"recovered_from": currentStatus,
+		})
+	}
+
 	// #1870 Phase 1: drain one queued A2A request if the target reports
 	// spare capacity. The heartbeat's active_tasks field reflects what the
 	// workspace runtime is ACTUALLY running right now, independent of
diff --git a/workspace-server/internal/handlers/registry_test.go b/workspace-server/internal/handlers/registry_test.go
index a8407b956..51640051a 100644
--- a/workspace-server/internal/handlers/registry_test.go
+++ b/workspace-server/internal/handlers/registry_test.go
@@ -193,6 +193,54 @@ func TestHeartbeatHandler_ProvisioningToOnline(t *testing.T) {
 	}
 }
 
+// ==================== Heartbeat — failed → online recovery (#616 follow-up) ====================
+
+// A workspace flipped to 'failed' by the provision-timeout sweeper must recover
+// to 'online' once it starts heartbeating: a live heartbeat proves the agent
+// booted (just slowly, past the 10m budget), so the timeout flip was premature.
+func TestHeartbeatHandler_FailedToOnline(t *testing.T) {
+	mock := setupTestDB(t)
+	setupTestRedis(t)
+	broadcaster := newTestBroadcaster()
+	handler := NewRegistryHandler(broadcaster)
+
+	mock.ExpectQuery("SELECT COALESCE\\(current_task").
+		WithArgs("ws-failed").
+		WillReturnRows(sqlmock.NewRows([]string{"current_task"}).AddRow(""))
+
+	mock.ExpectExec("UPDATE workspaces SET").
+		WithArgs("ws-failed", 0.0, "", 1, 3000, "").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	// evaluateStatus SELECT — currently failed (provision-timeout sweeper flip)
+	mock.ExpectQuery("SELECT status FROM workspaces WHERE id =").
+		WithArgs("ws-failed").
+		WillReturnRows(sqlmock.NewRows([]string{"status"}).AddRow("failed"))
+
+	// the new failed → online recovery transition
+	mock.ExpectExec("UPDATE workspaces SET status =").
+		WithArgs(models.StatusOnline, "ws-failed").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	mock.ExpectExec("INSERT INTO structure_events").
+		WillReturnResult(sqlmock.NewResult(0, 1))
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	body := `{"workspace_id":"ws-failed","error_rate":0.0,"sample_error":"","active_tasks":1,"uptime_seconds":3000}`
+	c.Request = httptest.NewRequest("POST", "/registry/heartbeat", bytes.NewBufferString(body))
+	c.Request.Header.Set("Content-Type", "application/json")
+
+	handler.Heartbeat(c)
+
+	if w.Code != http.StatusOK {
+		t.Errorf("expected status 200, got %d: %s", w.Code, w.Body.String())
+	}
+	if err := mock.ExpectationsWereMet(); err != nil {
+		t.Errorf("unmet sqlmock expectations: %v", err)
+	}
+}
+
 // ==================== Heartbeat — awaiting_agent → online recovery ====================
 // External workspaces flip to 'awaiting_agent' via healthsweep when their
 // heartbeat goes stale. When the operator's poller comes back, heartbeat